Uif!Obuvsf!pg! !uif!Xpse STUDIES IN HONOR OF PAUL KIPARSKY
EDITED BY
Kristin Hanson and Sharon Inkelas
The Nature of the Word
Current Studies in Linguistics Samuel Jay Keyser, general editor A complete list of books published in the Current Studies in Linguistics series appears at the back of this book.
The Nature of the Word Studies in Honor of Paul Kiparsky
edited by Kristin Hanson and Sharon Inkelas
The MIT Press Cambridge, Massachusetts London, England
6 2009 Massachusetts Institute of Technology All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. For information about special quantity discounts, please e-mail
[email protected] This book was set in Times New Roman and Syntax on 3B2 by Asco Typesetters, Hong Kong. Printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data The nature of the word : studies in honor of Paul Kiparsky / edited by Kristin Hanson and Sharon Inkelas. p. cm. — (Current studies in linguistics) Includes bibliographical references and index. ISBN 978-0-262-08379-9 (hardcover : alk. paper) — ISBN 978-0-262-58280-3 (pbk. : alk. paper) 1. Grammar, Comparative and general. 2. Lexicology. 3. Word (Linguistics) 4. Kiparsky, Paul. I. Kiparsky, Paul. II. Hanson, Kristin. III. Inkelas, Sharon. P201.N375 2009 415—dc22 2007041980 10
9 8 7 6 5
4 3 2 1
To see a World in a Grain of Sand And a Heaven in a Wild Flower Hold Infinity in the palm of your hand And Eternity in an hour —William Blake, ‘‘Auguries of Innocence,’’ 1–4
Paul Kiparsky. Photo by Cleo Condoravdi.
Contents
Preface xi Biography of Paul Kiparsky
xv
I
Metrics
1
1
On ‘‘The Phonological Basis of Sound Change’’ (after Hopkins)
3
Andrew Garrett 2
On Stress and Meter and on English Iambics in Particular
5
Morris Halle 3
Nonlexical Word Stress in the English Iambic Pentameter: A Study of John Donne 21
Kristin Hanson 4
The Prosodic Word as a Unit in Poetic Meter
63
Draga Zec 5
The Word in Tiberian Hebrew
95
Bezalel Elan Dresher 6
Faithfulness and Componentiality in Metrics
113
Bruce Hayes II
Phonology and Morphology
7
The Phonology of Perceptibility Effects: The P-Map and Its Consequences for Constraint Organization 151
Donca Steriade
149
viii
8
Contents
Vowel Duration, Syllable Quantity, and Stress in Dutch
181
Carlos Gussenhoven 9
Sympathy Meets Argentinian Spanish
199
Ellen M. Kaisse 10
Vowel Length, Cyclicity, and Output-Output Correspondence
215
Cemil Orhan Orgun 11
Level Ordering in Nuuchahnulth
225
John Stonham 12
Inside Access: The Prosodic Role of Internal Morphological Constituency
241
Patricia A. Shaw 13
Morphosyntactic Correspondence in Bantu Reduplication
273
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda 14
Patterns of Reduplication in Yoruba
311
Douglas Pulleyblank 15
Multiple Tenses in the Malayalam Verb
359
Tara Mohanan and K. P. Mohanan On Pa¯nini 2.4.81 (a¯mah) ˙ ˙
387
III
The Lexicon and Change
395
17
Lexical Irregularity and the Typology of Contrast
16
S. D. Joshi and J. A. F. Roodbergen
397
Rene´ Kager 18
Derived Environment Effects in Colloquial Helsinki Finnish
433
Arto Anttila 19
A Historical Perspective on Nonderived Environment Blocking: The Case of Korean Palatalization 461
Young-mee Yu Cho 20
Lexical Storage and Phonological Change
Geert Booij
487
Contents
21
ix
The Dental Preterites in the History of English
507
Aditi Lahiri 22
Analogical Morphophonology
527
Andrew Garrett and Juliette Blevins IV
Syntax and Semantics
547
23
Second-Position Clitics in Tagalog
549
Stephen R. Anderson 24
‘‘Elsewhere’’ in Gender Resolution
567
Stephen Wechsler 25
The Force of Lexical Case: German and Icelandic Compared
Dieter Wunderlich 26
Constraints on Source/Goal Co-occurrence in Carrier
William J. Poser 27
Punctual Until as a Scalar NPI
631
Cleo Condoravdi 28
The Existential Tense in Hungarian
Christopher Pin˜o´n V
Poetics
675
29
The Art of Fusion
677
Haj Ross References 697 Contributors 743 Language Index 745 Subject Index 747
655
621
587
Preface
This book honors Paul Kiparsky, whose contributions as a scholar and teacher have transformed virtually every subfield of contemporary linguistics. In taking the form of a collection of studies of the word, it reflects the distinctive focus of his own attention, and the consequent shape of his influence. The word has traditionally been recognized as a fundamental entity of language. Ordinary speakers often identify knowing a language with knowing its words, children attend to and acquire words before any other structural units in a language, and when one language influences another it is most commonly its words that are borrowed. From a more scholarly perspective too, the word has for centuries occupied a central place as the repository of basic phonological patterns, morphological structures, syntactic capabilities, semantic content, correspondences allowing the reconstruction of historical relationships among languages, and poetic possibilities. Kiparsky’s work belongs to this tradition, but in a way unique to him. As T. S. Eliot writes of poets in ‘‘Tradition and the Individual Talent,’’ If the only form of tradition, of handing down, consisted in following the ways of the immediate generation before us in a blind or timid adherence to its successes, ‘‘tradition’’ should positively be discouraged. We have seen many such simple currents soon lost in the sand; and novelty is better than repetition. Tradition is a matter of much wider significance. It cannot be inherited, and if you want it you must obtain it by great labour. It involves, in the first place, the historical sense . . . ; and the historical sense involves a perception, not only of the pastness of the past, but of its presence; the historical sense compels a man to write not merely with his own generation in his bones, but with a feeling that the whole of the literature of Europe from Homer and within it the whole literature of his own country has a simultaneous existence and composes a simultaneous order. This historical sense, which is a sense of the timeless as well as of the temporal and of the timeless and of the temporal together, is what makes a writer traditional. And it is at the same time what makes a writer most acutely conscious of his place in time, of his contemporaneity.
Kiparsky’s ‘‘place in time’’ is in the first group of PhD students graduated in 1965 from the new program in linguistics at MIT, a program directed by Morris Halle and committed to developing Noam Chomsky’s revolutionary definition of language
xii
Preface
as a generative system, and the hypothesis that crucial aspects of its formal structure are universal because they are innately determined. Kiparsky’s ‘‘historical sense’’, it seems to us, has compelled him to express within that revolutionary paradigm ‘‘a feeling’’ that the whole of language and within it the whole of each language ‘‘has a simultaneous existence and composes a simultaneous order.’’ It is labor to obtain Kiparsky’s tradition within the revolution that has defined his particular intellectual influence that is represented here. At the outset of the early generative period, theoretical interest in productive rules in both syntax and phonology drew attention away from words, associated with the lexicon and memorized forms; but within only a few years, representing systematicity within the lexicon itself and relating the lexicon to other subsystems of grammar emerged as essential to generative theory. Di‰culties in defining the word vexatious to everyone from Scrabble players to dictionary editors to poets turned out to reflect real complexities in the structure of grammar that had to be taken into account: Is a compound word like blackbird one word or two? Why is the ‘‘g’’ of longer pronounced like that of a simple word like finger rather than that of a complex word like longing? Why do poets not treat the stress of a preposition like between quite the same way that they treat that of a verb like believe? How can we know that a word that expresses such a dazzling set of complex syntactic and semantic relationships as the Finnish adverb voimisteluttelemasta, ‘from having on and o¤ caused to do gymnastics’, really is just one word, and not the nine that English needs to convey the same meaning? Kiparsky’s serious attention to questions like these has shown the word to be as central to grammar as tradition had always suggested. His work on formalizing its role, from his theory of Lexical Morphology and Phonology to his theories of morphosyntax and of the role of analogy in sound change, has been fundamental to the development of lexicalist approaches to every subsystem of grammar within the generative tradition. The range of Kiparsky’s influence derives not only from his rigorous focus on such an intricate and central entity as the word, but also from another intellectual commitment, equally represented in this book’s title. Chomsky’s definition of language famously cast linguistics as a science, and its object of study as a phenomenon of nature. In Kiparsky this idea must have fallen on especially fertile soil, as it were; for, as fine a gardener as he is a linguist, Kiparsky approaches words rather as a botanist approaches plants, fascinated equally by their beauty, by their structure, and by their evolution, and aware that only by considering them from these multiple perspectives can one begin to understand their true nature. The contributions to this volume represent these multiple perspectives on the word that Kiparsky has explored and inspired, and are organized around the several complementary senses of the word nature that they seem to reflect. Part I, ‘‘Metrics,’’ recognizes the nature of the word not only as a source of beauty, but also as something
Preface
xiii
that is not artifice, a recognition whose profound significance for literary theory Kiparsky’s own work has drawn out. Part II, ‘‘Phonology and Morphology,’’ explores the nature of the word as a matter of its formal composition, precisely the sense in which morphology refers equally to language and to organisms. Part III, ‘‘The Lexicon and Change,’’ addresses the nature of the word simultaneously in the sense of having an idiosyncratic character, and in the sense of having a genetic inheritance, in something of the same way that nineteenth-century historical linguists’ conceptions of the ‘‘genius’’ of an individual language combined historical accidents with formal necessity. Part IV, ‘‘Syntax and Semantics,’’ captures the nature of the word in the sense of its having certain aptitudes, the way an organism’s nature determines what it can do in the world. Finally, part V, ‘‘Poetics,’’ acknowledges that the nature of the word, like nature in general, is not, finally, objectively describable, but only inferrable, partially and imperfectly, from observations that will always be di¤erent under di¤erent conditions. Poetry takes this as a central fact, and for Kiparsky, with a humbleness that paradoxically makes his work especially powerful, scientific inquiry into language is, in this regard, not fundamentally di¤erent from poetry. Representation of this range of perspectives on language has partially determined the contributors to this book, but of course much more remains to be said about the inevitably di‰cult matter of their selection. The scope of Kiparsky’s influence vastly exceeds what any single book could contain, and we have chosen to deal with this by simply acknowledging the limitations of our own perspectives, and focusing on those scholars whose close intellectual associations with Kiparsky we were especially aware of in our own work as students entering the graduate linguistics program at Stanford in 1984, the same year that he came there from MIT. These include Kiparsky’s own teachers, his previous students, our fellow students, and colleagues whose relationship to him, with the remarkable indi¤erence to institutional limitations so characteristic of him, derived purely from shared intellectual curiosities. The book largely excludes the many students and colleagues who have been important to him since that time, because we were aware when we began this project that that set would have grown and would continue to grow in ways we were simply not in a position to keep up with. It also excludes many whose own expertise overlapped insu‰ciently with our own at that time for Kiparsky’s relationship to them to have been salient to us. And it undoubtedly excludes still others simply through errors for which we can only hope we will be forgiven. In this raggedness, at least, it represents something of Kiparsky’s own helpful, honest awareness of how knowledge gets advanced: a bit here, a bit there, with gaps that hopefully can somehow, sometime be filled in. Finally, it must be added that Kiparsky’s influence does not, of course, derive from intellectual accomplishment alone. The same sense that any single word or flower, if studied closely enough in both its specialness and its generalness, can yield secrets of the universe, Kiparsky brings to his interactions with students, colleagues, and
xiv
Preface
friends—finding, cultivating, and delighting in each one’s individual talent and potential relationship to tradition, not just within linguistics but within human experience more broadly. Shared exploration of language is for Kiparsky thus inseparable from shared enjoyment of life, and hence from friendship; which is, of course, in the end the deepest motivation for this book. In this spirit we are therefore especially grateful for the patience, cooperation, expertise, and kindness of the many people who have helped bring the book into being. Dikran Karagueuzian helped us conceive, plan, and set in motion the entire project. Ann Banfield, Jim Blevins, Ed Flemming, Andreas Kathol, Paul Kay, and especially Gary Holland made helpful suggestions about individual papers, as did in fact virtually all the contributors, reading each others’ papers as well as writing their own. Jeremy Ecke helped transform twenty-nine idiosyncratic papers into a single, coherent manuscript. At The MIT Press, Tom Stone enthusiastically took on the project, and Sandra Minkkinen e‰ciently and elegantly saw it through to completion. Everyone working with her contributed stellar editorial and production work much appreciated by all the authors, who remained heroically patient during a long wait to see their papers appear. We were especially grateful for the involvement of Anne Mark—in the words of the book’s own honoree, ‘‘a legend.’’ Renowned for miracles of a di¤erent kind, Samuel Jay Keyser made a breeze blow again at a time when everything seemed becalmed. And at the Library of the Royal Botanic Gardens, Kew, with all the discretion of a royal retainer, Samantha Cox beautifully answered our peculiar request for botanical drawings of Finnish orchids.
Biography of Paul Kiparsky
Born in Helsinki on January 28, 1941, Rene´ Paul Victor Kiparsky is a scion of a long line of scholars, not least his parents Dagmar (ne´e Jaatinen) and Valentin Kiparsky. His great-grandfather Valentin Ludwig Kiparsky wrote a dissertation on plastic surgery in 1864, at a time of academic renaissance at Tartu University and the emergence of Estonian nationalism. Paul’s grandfather, Rene´ Valentinovich Kiparsky (1867–1938), was a distinguished gynecologist in St. Petersburg who developed several innovative surgeries and wrote a standard text with Dmitri Ott, head of the research institute that now bears his name (Ott 1914).1 Paul’s father, Valentin Julius Alexander Kiparsky (1904–1983), is best known among linguists for his book on Russian accent and his three-volume historical grammar of Russian (V. Kiparsky 1962, 1963–1975). Equally important, in a long career devoted to Russian and the Baltic linguistic area, are his many detailed studies of linguistic contact and di¤usion, including a dissertation on prehistoric Germanic-Slavic contact and monographs on topics ranging from the loan vocabulary of Baltic German to the history of terms for the walrus (V. Kiparsky 1934, 1936, 1952). Valentin Kiparsky is indeed still cited for his early statement, in 1938, that a language’s receptiveness to borrowing depends as much on social factors as it does on facts about linguistic structure. It is relevant to add that when he joined the University of Helsinki faculty in the same year, he introduced modern phonology (Prague School structuralism) to Finland. Paul Kiparsky’s career continues the family tradition. After studying at Alabama College and the Universities of Helsinki and Minnesota, Kiparsky received his PhD in 1965 from MIT. For two decades he taught at MIT, and since 1984 he has taught at Stanford University, where he is Bass Professor in the School of Humanities and Sciences. Among other honors he has served as Collitz Professor of the Linguistic Society of America (1973), been awarded an honorary doctorates by Go¨teborg University (1985) and the University of Konstanz (2008), and received the Alexander von Humboldt Prize (1993). He has directed over thirty dissertations in fields as diverse as poetics, language change, semantic and syntactic theory, and every area of phonology and morphology.
xvi
Biography of Paul Kiparsky
Kiparsky’s work in linguistics has been wide-ranging and foundational. He has redefined no fewer than four fields of linguistics (so far) and made major contributions to several others, notably semantics and Paninian studies. His dissertation and a series of papers mainly devoted to Greek phonology (Kiparsky 1965, 1967a, 1967c, 1967d), together with several other articles published in his second book (Kiparsky 1982a) and elsewhere, established him as the central figure in a new field of generative historical linguistics—the analysis of sound change and analogy in the grammatical terms of generative linguistics. More recent papers such as his survey of phonological change and several radical reanalyses of morphological changes in Gothic and Latin (Kiparsky 1995, 1998, 2000a, 2000b) show that he continues to be as much a leader in this field as ever. His work throughout the 1960s on phonological and morphological change led Kiparsky to assess the relationship between phonology and morphology more broadly, and the result of his and others’ research in this area was the field of lexical phonology as it arose in the 1970s and 1980s. This field remains at the center of linguistic theory, and the questions it poses are urgent for linguists of all theoretical persuasions; still among its foundational texts are a series of papers by Kiparsky (1982b, 1982c, 1982d, 1984, 1985). At the same time, starting with a classic treatment of Indo-European modality (Kiparsky 1968), Kiparsky has set a new standard in the field of historical syntax. He continues this line of research unabated, in the last decade through a dazzling set of studies tracing the evolution of Indo-European word order structures in Germanic and English and in the history of Greek (Kiparsky 1994, 1996c, 1997c; Condoravdi and Kiparsky 2001). Finally, Kiparsky created and still defines the modern linguistically informed analysis of poetics and poetic meter. His earliest work in this vein was a study of Finnish meter, to which he has recently returned (Kiparsky 1967d; Hanson and Kiparsky 1996); other major contributions appear in classic volumes on oral poetry and on the linguistics of meter (Kiparsky 1976, 1989). Kiparsky’s students and friends will agree that in accomplishment and breadth of interest he is the truest successor to his teacher Roman Jakobson, whose paraphrase of Terence equally defines Paul Kiparsky’s own career: Linguista est; linguistici nihil a se alienum putat. Andrew Garrett University of California, Berkeley Note 1. See Ol’shanets’kyi 1950, 103–104. I am grateful to Boris Maslov for archival research in St. Petersburg and to Lisa Conathan, Cleo Condoravdi, and Leslie Kurke for discussion and other information.
The Nature of the Word
I
Metrics
1
On ‘‘The Phonological Basis of Sound Change’’ (after Hopkins)
Andrew Garrett
Laulelevi virsia¨nsa¨, laulelevi, taitelevi —Kalevala 3.5–6 I taught last April spring’s companion, Ling 230’s thriller, thickly theory-thought Chapter, in its writing On di¤usion level-ordered short-a tensing claims, and gliding By there, how it does for all the facts an answer bring In analogy! then o¤, at once it’s a-swing Onto sound change brought forth in perception: structural tiding Preserved in such shifts. My heart in hiding Stirs for a word,—the achieve of, the mastery of the thing! Brute beauty and thinking react, the pattern of sound’s here Lexical! and it’s grammar guiding speech when, if alien Forms seem optimal, more simplified, they will appear! No wonder of it: she´er plo´d may scan a zillion Lines of Finnish song, but in that mere Paul trawls alone, and finds gold in prillion. [Falling paeonic rhythm, sprung and outriding]
2
On Stress and Meter and on English Iambics in Particular
Morris Halle
2.1
Introduction
An important di¤erence between poetry and prose is that all poetry is composed of lines, whereas prose has sentences, clauses, and phrases but no lines. In metrical poetry, the length of the lines is governed by a set of principles and rules in a manner detailed below; in free verse—as far as is known—no such principles or rules determine line length, yet poems in free verse are invariably made up of lines. The line must thus serve essential poetic purposes in all kinds of verse, even though in free verse the principles that govern lineation—and motivate its being printed without justification at the right margin—remain to be discovered (see Smith 1968; Steele 1990 for some discussion). In what follows we shall be concerned with metrical verse exclusively. A central proposition of this chapter is that every known type of meter is based on an iterative footing rule of the kind encountered in the assignment of word stress in many languages. As in Kiparsky (1977, 190), it is assumed here that feet are computed ‘‘by a kind of paraphonology that modifies the phonological derivations of the language and produces as output one or (more commonly) several alternative representations that may di¤er from phonetic representations.’’ In Kiparsky 1977 the metrically relevant properties of the line were characterized by means of the prosodic theory of Liberman 1975. The main innovation of the present study is that it characterizes the metrically relevant properties of the line—that is, it scans the lines—with the help of the prosodic theory of Idsardi 1992. My aim is to show that this innovation provides important insights into metrical phenomena of all kinds. 2.2
On Word Stress
An important advance in the understanding of the prosody of words was Liberman’s (1975) study of the English intonational system. Liberman’s key insight was that unlike [back], or [round], or [continuant], stress is not a phonetic feature. Instead, he
6
Morris Halle
argued, stress is a reflex of the tendency of languages to group syllables—more precisely, stressable phonemes—into feet. Stress, on this view, is the phonetic reflex of the foot structure of words. This implies that the computation of stress involves two separate steps. As illustrated below, we first compute the foot structure of the word and use it to assign stress—that is, high tone or some other phonetic mark—to phonemes that occupy special (head) positions in the feet so constructed. There have been a number of attempts to implement Liberman’s insight formally, for example, Liberman and Prince 1977, Hayes 1981, Prince 1983, Halle and Vergnaud 1987, and Idsardi 1992 (see also Halle and Idsardi 1995). It is the latter formalization that is adopted below. An obvious fact about stress is that not all phonemes are capable of bearing stress. In most languages all and only vowels are stressable; other phonemes are unstressable. This, however, is not true in all languages. For example, in Indonesian the schwa vowel may never bear stress (see, e.g., Halle and Idsardi 1994 and references therein). On the other hand, in Lithuanian, not only vowels but also glides, nasals, and liquids that are part of the syllable nucleus are stressable (Halle and Vergnaud 1987). A minimal requirement for an adequate stress notation is therefore that it include a means for indicating which phonemes in a sequence can bear stress. This requirement is implemented here formally by projecting stressable phonemes on a separate autosegmental plane, and it is sequences of the projected— that is, stressable—phonemes represented by asterisks in (1) and elsewhere that are grouped into feet. Since on the account adopted here stress is a reflex of foot structure, we need a formal device for grouping the stressable phonemes into feet. As illustrated in (1), the grouping of stressable phonemes into feet is accomplished by foot boundary markers or junctures, represented here by ordinary parentheses. A left parenthesis (foot boundary) foots the stressable elements on its right, whereas a right parenthesis foots those on its left; elements that are neither to the left of a right parenthesis nor to the right of a left parenthesis are unfooted. Thus, in the first and third example in (1b) the last asterisk (¼stressable phoneme) is not footed, but in the examples in (1a) and (1c) all asterisks are footed. The crucial di¤erence between this formalization and earlier ones (such as those of Liberman 1975; Hayes 1981; or of Halle and Vergnaud 1987) is that in the earlier notations, feet, like syntactic constituents, had two ends and therefore were delimited by a pair of matched parentheses, whereas in the present notation a single, unmatched parenthesis defines a foot. (1) a. Maranungku (Hayes 1981, 51) (* * (* * (* *( (* * (* * (* w e´ l e p e` n e m a` n t a l a´ n g k a r a` t e t `ı ‘kind of duck’ ‘prawn’
On Stress and Meter and on English Iambics in Particular
b. Pintupi (Hayes 1995, 62) )* *) * )* *) * *) j m a´ L a w a` n a t u´ T a y a ‘many’ ‘through from behind’ c. Weri (Hayes 1981, 53) )* *)* *) *) * *) * *) u l u` a m ´ı t a` k u n e` t e p a´ l ‘mist’ ‘times’
7
)* *) * *) * p u´ L i N k a` l a t j u ‘we (sat) on the hill’
It has been remarked above that feet are defined by parentheses, but to this point nothing has been said about where these parentheses come from. It is assumed here that speakers of a language have the ability to break a sequence of stressable phonemes into subsequences or feet. We picture this formally (literally) as the ability to insert parentheses into strings of asterisks, as illustrated in (1). The feet in the examples in (1b,c) are defined by right parentheses, those in (1a), by left parentheses. The parentheses are inserted by rules of Iterative Footing, that is, by rules that work their way across the stressable phonemes (asterisks) of a word starting at one edge and ending at the opposite edge. The iterative rules illustrated in (1) insert parentheses next to every other asterisk that they encounter, that is, they skip over two unfooted asterisks in each step. In (1a,b) the insertion begins at the left edge and proceeds rightward; in (1c) the insertion proceeds in the reverse direction, from right to left. In addition to binary footing rules like those illustrated in (1) there are also ternary footing rules, where three asterisks are skipped in each iteration. Stress systems with ternary feet, though considerably rarer than those with binary feet, are encountered, for example, in Cayuvava and Alutiq. (See Idsardi 1992 for some discussion.)1 The present theory provides only two kinds of iterative feet—binary and ternary—feet of greater length, such as quaternary, are specifically excluded as being beyond the innate capabilities of normal speakers. The machinery available for stress assignment—and for footing lines of metrical verse—is thus limited to counting by pairs and/or by triplets. Since the notation makes available both left and right parentheses, the type of parenthesis to be inserted must be specified in each case. As seen by comparing the rightmost examples in (1a) and (1b), the choice of di¤erent parentheses may result in scanning the same number of asterisks (syllables, stressable phonemes) into di¤erent numbers of feet. Thus, the three left parentheses in the five-asterisk sequence in (1a) generate three feet, whereas the three right parentheses in (1b) generate only two feet. Because a foot may be composed of more than one element, foot structure alone cannot define the placement of stress. To this end we have recourse to the fact that every foot has exactly one head and that stress is placed on the head of
8
Morris Halle
a foot. It is, moreover, the case that the head of a foot is almost either its left-most or its rightmost element, and the choice between the two is determined by a special, language-particular Head rule. In Maranungku and Pintupi (shown in (1a,b)) feet are left-headed, whereas in Weri (1c) feet are right-headed. Iterative Footing rules are not the only source of parentheses: parentheses may also be inserted by Accent rules. The Accent rules supply parentheses to specific syllables (more precisely, to asterisks projected from specific syllables). For example, many languages stress syllables with heavy rimes, or syllables that have special morphological characteristics. In the theory adopted here this is done by an Accent rule that inserts a parenthesis next to (asterisks projecting from) such syllables. Parenthesis insertion of this type is the main stress mechanism in languages such as Koya, Selkup, Khalkha Mongolian, and Komi. (For some discussion, see Idsardi 1992 and, from a di¤erent theoretical perspective, Hayes 1995:296¤.) In the Indo-European languages with free stress, such as Russian, Serbo-Croatian, Sanskrit, and Lithuanian, parentheses are supplied to some morphemes in their underlying representation. (For additional discussion, see Halle 1997 and Kiparsky and Halle 1977.) Of particular interest for present purposes are languages that make use both of Accent rules and of Iterative Footing. In such languages, Iterative Footing respects the partial footing imposed by the Accent rules. This is illustrated by the Finnish examples in (2), where right parentheses are inserted by Accent rules and left parentheses are inserted by Iterative Footing. (2) Finnish (Hanson and Kiparsky )* *) * )* *) (* o´ m e n a o´ m e n a` t ‘apple’ ‘the apples’ )* *) * (* l u´ m o s i v a` t ‘they enchanted’
1996, 301) )* *) (* *) l u´ m o t t u` i h i n ‘enchanted’
)* *) (* *) (* k o´ s k e m a` t t o m a` n ‘untouched’
)* *) * (* *) * (* *) v o´ i m i s t e l u` t t e l e m a` s t a ‘from having on and o¤ caused to do gym’
)* *) * (* *) o´ p p e t t a m a` s t a ‘from teaching’ )* *) * o´ m e n a t ‘the apples’
)* *) * *) l u´ m o s `ı v a t ‘they enchanted’
)* *) (* *) * k o´ s k e m a` t t o m a n ‘untouched’ In Finnish, according to Hanson and Kiparsky (1996, 301), ‘‘[s]econdary stress is subject to phonological and morphological conditioning (with variability in some contexts).’’ The formal reflex of this conditioning here is the Accent rule (3a), which inserts a left parenthesis before the syllables in question. (See the underlined paren-
On Stress and Meter and on English Iambics in Particular
9
theses in (2)). Binary feet are then constructed by a rule of Iterative Footing (3b), which inserts right parentheses left to right at binary intervals. Stresses are assigned by positing that feet in Finnish are left-headed, that is, trochaic, and that stress (high tone) is assigned to the heads. As noted, Iterative Footing respects the rudimentary foot structure assigned to the word by the Accent rule. Formally, this is implemented by ordering the Accent rule before Iterative Footing. It is to be noted that Iterative Footing applies only to sequences of unfooted syllables; it does not foot a pair of asterisks separated by a (previously inserted) foot boundary. Thus, in o´ppettama`sta and voimistelutttelemasta in (2), Iterative Footing inserts a right parenthesis after the fifth—and not after the fourth—syllable. It is the latter fact that is reflected in the rather complex part (italicized below) of Hanson and Kiparsky’s formulation of the distribution of secondary stresses: ‘‘Secondary stress falls on every second syllable after the initial one, skipping an additional light syllable if the syllable after that is heavy’’ (p. 301). There are two exceptions to the preceding. (i) According to Hanson and Kiparsky, ‘‘[f ]inal syllables not preceded by stressed syllables are optionally stressed.’’ In the formalism adopted here, this means that in word-final position the Finnish Accent rule (3a) is optional. Such alternative stressings without a left parenthesis before the word-final syllable are exemplified in the last line of (2). (ii) The second, that is, peninitial, syllable of a word is never stressed, even if heavy. This fact is taken into account in the formulation of the Accent rule in (3a).2 The basic footing rules for Finnish stress are given in (3). (3) a. Accent rule In postpeninitial position, insert left parentheses before heavy syllables and before specially marked syllables. (This rule applies optionally before wordfinal syllables.) b. Iterative Footing rule Insert right parentheses iteratively from left to right starting at the left edge of the string and skipping over two consecutive syllables at each step. Heads: left Since in Finnish the main stress of the word is distinguished from its secondary stresses, additional machinery is needed that will allow us to distinguish among stressed syllables—that is, among heads of feet. In the Idsardi notation each head of a foot is projected onto the next higher line in the metrical grid. This property of the notation also provides a straightforward way of distinguishing among heads of feet. We need only posit that the sequence of heads of feet projected onto line 1 is footed by an Accent rule that inserts a left parenthesis at the left edge of line 1 and that line 1 feet are left-headed. This is illustrated in (4), where the left parenthesis inserted by the Accent rule on line 1 is represented by a square bracket rather than a parenthesis.
10
Morris Halle
(4) * [* )* *) * o´ m e n a
* [* * )* *) (* o´ m e n a` t
* [* * )* *) * (* l u´ m o s i v a` t
* [* * * )* *) (* *) (* k o´ s k e m a` t t o m a` n
* [* * )* *) * (* *) o´ p p e t t a m a` s t a 2.3
* [* * )* *) (* *) l u´ m o t t u` i h i n
line 2 line 1 line 0
line 2 line 1 line 0
* [* * * )* *) * (* *) * (* *) v o´ i m i s t e l u` t t e l e m a` s t a
line 2 line 1 line 0
The Strict Meters of English Verse
The formal machinery sketched in the preceding section accounts also for the di¤erent types of metrical verse encountered in the languages of the world. It is not the case that the stress rules of a language are the same as those that account for its metrical verse. For example, English word stress is trochaic (see Halle 1998 for some discussion), but English metrical poetry is, of course, not restricted to trochaic meters. Though not identical, the rules governing word stress and the rules of meter are of the same kind. Underlying both English word stress and the main meters of English is a rule of Iterative Footing (cf. (3b), (9a), and (15a)). Iterative Footing is supplemented by an Accent rule both in the assignment of stress to Finnish words discussed (cf. sec. 2) and in the computation of the loose meters of English discussed in section 2.4 (cf. (3a) and (15b)). When examined in detail, the rule of Iterative Footing involves the setting of the four binary parameters in (5). (5) a. b. c. d.
The type of parenthesis inserted: left or right The direction of insertion: (left to right) L>R or (right to left) R>L The number consecutive asterisks skipped after each insertion: two or three Head placement: left- or rightmost
There are thus, in principle, 2 4 ¼ 16 distinct ways of footing a string of asterisks. Not all of these 16 footings appear to be used equally in metrical verse. In English poetry from Chaucer to the end of the nineteenth century, the overwhelming majority of lines are composed in one of the four meters (feet) in (6). In these meters, which are referred to here as the strict meters of English, head placement (5d) determines both the type of parenthesis inserted (5a) and the direction of insertion (5b):
On Stress and Meter and on English Iambics in Particular
11
when left-headed feet are constructed, left parentheses are inserted from left to right; when right-headed feet are constructed right parentheses are inserted from right to left. (6) a. Trochees Insert (, left to right, b. Dactyls Insert (, left to right, c. Iambs Insert ), right to left, d. Anapests Insert ), right to left,
heads: left, binary heads: left, ternary heads: right, binary heads: right, ternary
Examples of these four meters are given in (7). (7) a. Trochees (6a) Life is real! Life is earnest! (* * (** (* *(* *( And the grave is not its goal: (* * (* * (* * (* (Longfellow, ‘‘A Psalm of Life’’) b. Dactyls (6b) Half a league, half a league, (* * * (* * *( Half a league onward, (* * * (* * (Tennyson, ‘‘The Charge of the Light Brigade’’) c. Iambs (6c) Let me not to the marriage of true minds )* *) * *) * *) * *) * *) Admit impediments. Love is not love )* *) * *) * *) * *) * *) (Shakespeare, Sonnet 116) d. Anapests (6d) ‘Twas moonset at starting; but, while we drew near * *) * * *) * * *) * * *) L o k e r e n, t h e c o c k s c r e w, a n d t w i l i g h t d a w n e d c l e a r. *) * * *) * * *) * * *) (Browning, ‘‘How They Brought the Good News from Ghent to Aix’’) In (7) like in (4) the consecutive syllables of each line are represented by asterisks underneath the words, and parentheses are inserted into this sequence of asterisks by
12
Morris Halle
the rule of Iterative Footing given in bold type at the right. For example, in (7d) right parentheses are inserted from right to left skipping three syllables after each insertion. Since (7d) is an instance of a strict meter, head placement determines both the direction of insertion and the kind of parenthesis inserted. As shown in (7d), the result of this procedure is to construct four anapests on both lines. Footings other than those in (7) can be imposed on the lines. For example, in (8) I have footed one of the iambic lines in (7c) into anapests by means of rule (6d) and one of the anapestic lines in (7d) into trochees by rule (6a). (8) a. By rule (6d) Admit impediments. Love is not love *) * * *) * * *) * * *) b. By rule (6a) ‘Twas moonset at starting; but, while we drew near (* * (* * (* * (* * (* * (* The availability of the alternative footings in (8) raises the question as to how these patently wrong footings are eliminated in favor of the correct ones in (7). The obvious answer is that in the English verses under discussion here, the choice of the correct footings is determined by the position of the stressed syllables in the line. While traditional approaches have treated all stressed syllables as equal in determining the well-formedness of a line, it is assumed here—following Halle and Keyser 1999— that only the main stresses of polysyllabic words determine whether a particular footing is a valid instantiation of a given meter. We call such stressed syllables Maxima and require that Maxima be placed only in head positions of feet. This requirement is violated by the footings in (8), where the respective Maxima of Admit in (8a) and of moonset in (8b) occur in nonhead positions of the feet, and it is these violations that rule out the footing in (8). By contrast, none of the footings in (7) violates the requirement just proposed. The fact that stressed monosyllables play no role in determining the well-formedness of lines is illustrated by the clause Love is not love in (7c). As shown there, the first word Love is in a nonhead position, even though this word has greater stress than either of the two syllables adjoining it. Being in a monosyllabic word, however, this stress is not a Maximum and does not violate the requirement above. We have now reviewed all the main ingredients involved in the scansion of the lines in (7). The rules and conditions that define the strict meters of English are stated in (9). These consist of a definition of Maximum, a rule of Iterative Footing (9a), and two conditions (9b). (9) Definition A Maximum is the syllable bearing main stress in a polysyllabic word. (To be modified below cf. (15).)
On Stress and Meter and on English Iambics in Particular
13
a. Iterative Footing rule Select one of the four parameter settings in (6). b. Conditions i. Maxima can be placed only in head positions of feet. ii. Incomplete feet are allowed to terminate the iterative footing procedure. (Incomplete feet may therefore appear at the beginning of iambic and anapestic lines and at the end of trochaic and dactylic lines.) Condition (9bii) accounts for the variation in line length found in the examples in (7). For example, in (7a) the first line is eight syllables long, whereas the second line has only seven syllables. The di¤erence in line length is a consequence of the manner in which foot boundaries are inserted by the rule of Iterative Footing. As stated in (6a), trochees are generated by a rule of Iterative Footing that inserts left parentheses L(eft)>R(ight), skipping two consecutive asterisks after each insertion. This rule will obviously foot a sequence of eight syllables into four feet. However, since a foot may consist of a single asterisk, this rule will also foot a sequence of seven syllables into four feet. Since in (7a) Iterative Footing proceeds L>R, the incomplete foot appears at the end of the line. Where Iterative Footing applies R(ight)>L(eft) the incomplete foot appears at the beginning of the line. We see an example of this in (7d), where the first line begins with a disyllabic foot and the second line with a monosyllabic foot. Since in anapests Iterative Footing skips three syllables after each insertion, it may generate incomplete feet that are either monosyllabic or disyllabic. 2.3.1
Elision
As illustrated in the even-numbered lines of (10), extra syllables may appear in some lines. (10) a. O, n o, i t i s a n e v e r f i x e` d m a r k, )* *) * *) * *) * *) * *) That looks on tempests and is never shaken; )* *) * *) * *) * *) * *) (Shakespeare, Sonnet 116) b. All the buds and bells of May, (* * (* * (* * (* From dewy sward or thorny spray; (* * (* * (* * (* (Keats, ‘‘Fancy’’) Traditionally the second line in (10a) is said to end with a ‘‘feminine’’ rime, and the extra syllable is not counted for purposes of determining the meter. This same procedure is followed here and such metrically ‘‘omitted’’ syllables are signalled by
14
Morris Halle
representing them not with an asterisk but with a bullet (), as shown in (10). In (10a) the ‘‘omitted’’ syllable is located at the end of the line, that is, at the starting point of Iterative Footing (cf. (6c)). The symmetrical situation obtains in the second line in (10b), where the Iterative Footing rule applies L>R. In this case, the metrically omitted syllable begins the line. The status of the omitted syllable resembles that of incomplete feet (see (9bii)). The two cases di¤er in that incomplete feet are admitted as the last iteration of a rule of Iterative Footing, whereas the metrically omitted syllables under discussion here occur at the beginning of Iterative Footing. Metrically omitted syllables occur also line medially, but their occurrence there is more highly marked and therefore used only sparingly. A few examples are given in (11). (11) a. H e s t a r e d a t t h e P a c i f i c—a n d a l l h i s m e n * *) * *) )* *) * *) * *) (Keats, ‘‘On First Looking into Chapman’s Homer’’) b. This fortification, gentlemen, shall we see it? )* *) * *)* *) * *) * *) (Shakespeare, Othello III, 2, 5) c. But as I past, I worshipt: if those you seek. )* *) * *) * *) * *) * *) (Milton, Comus, 302) d. Yet dearly I love you and would be love`d faine, )* *) * *) * *) * *) * *) (Donne, Holy Sonnet XIV) Such metrically omitted syllables are especially common in Donne’s poetry, where they occur most frequently in hiatus, that is, word finally before a vowel-initial word.3 Like Donne’s line (11d), the line from Shakespeare (11b) includes two lineinternal omitted syllables. (This and two additional lines are cited by Hanson and Kiparsky 1996, 298.) Both types of omission have traditionally been referred to as ‘‘elision,’’ which Merriam-Webster’s unabridged dictionary (1961) defines as a ‘‘deliberate syllablereducing suppression. . .of a. . .vowel in poetry for the sake of the meter.’’ The present theory treats all such vowels by not projecting them onto line 0 of the metrical grid. Special note is to be taken of the fact that metrical omission says nothing about the pronunciation of these syllables. While the meter and the pronunciation of a line are related, neither determines the other completely. 2.3.2
Ancipitia
A well-known peculiarity of English iambic verse is illustrated in (12), where every line begins with a stressed syllable of a polysyllabic word.
On Stress and Meter and on English Iambics in Particular
15
(12) a. Mighty and dreadful for thou art not so: )* *)* *) * *) * *) * *) (Donne, Holy Sonnet X) b. Kissing with golden face the meadows green )* *) * *)* *) * *) * *) (Shakespeare, Sonnet 33) c. Mindless of its just honours; with this key )* *) * *) * *) * *) * *) Shakespeare unlock’d his heart; the melody )* *) * *) * *) * *)* *) (Wordsworth, Sonnet) d. S i l e n t u p o n a p e a k i n D a r i e n. )* *) * *) * *) * *) **) (Keats, ‘‘On First Looking into Chapman’s Homer’’) Treating the line-initial stressed syllable in (12) as a Maximum results in all examples in a violation of condition (9bi). Since lines of this kind are readily found in the works of the best poets, it is necessary to modify the definition of the Maximum so as to exclude line-initial stresses. The restated definition is given in (13). (13) A Maximum is a syllable bearing main stress in a polysyllabic word provided that it is preceded and followed by a syllable with less stress. The e¤ect of (13) is to limit further the contexts where the stressed syllable of a polysyllabic word is counted as a Maximum. Since line-initial stresses are not preceded by a syllable without stress, they are also not Maxima, according to (13). The lines in (12), therefore, do not violate condition (9biii). The insight captured by this redefinition is that of anceps in classical metrics, which, according to Marouzeau 1943, 30 is ‘‘a syllable counted freely as short or as long (Lat. anceps ¼ ‘‘ambiguous’’)’’. For the role of ancipitia in Latin verse, see Embick and Halle in prep. The special treatment of line-initial stress is a peculiarity of English verse, which is not shared by the syllabotonic verse of Russian and German. This has been noted for Russian by Zˇirmunskij 1925/Eng. translation 1966, p. 53, and Bjorklund (1989, 156) for German. The definition of Maximum in Russian and German verse must therefore di¤er from that in English. Whether this di¤erence is connected with the di¤erence in phonetic stress pattern of the words in these two languages and how this connection is to be expressed formally is at present an open question. (For some discussion, see Hanson and Kiparsky 1996.) In view of the redefinition of Maximum in (13), we expect to find examples where the stressed syllable in a polysyllabic word is not counted as a Maximum in versemedial position. Such examples are given in (14).
16
Morris Halle
(14) a. A maid of grace and complete majesty (Shakespeare, Love’s Labour’s Lost I, 1, 137) b. His means of death, his obscure funeral (Shakespeare, Hamlet IV, 5, 213) c. Sir, in good sooth, in sincere verity. (Shakespeare, King Lear II, 2, 111) d. The divine property of her first being (Milton, Comus, 469) e. In profuse strains of unpremeditated art. (Shelley, ‘‘To a Skylark,’’ 6) The examples in (14), which have been copied from Bridges 1921, are frequently cited in discussions of English meter. Bridges attributed these to ‘‘recession of accent. . .not now heard’’ (p. 67). As the main (sole?) evidence for the accent recession is the scansion of the lines, Bridges’s proposal is not compelling, especially since accent movement has not been adduced by Bridges or anybody else as an explanation for the similar line-initial examples in (12); for example, nobody has suggested that Shakespeare in (12c) is subject to accent advance. In view of this there is reason to question accent recession as an explanation for the facts in (14). This is yet another instance of the indirect relationship between metrical pattern and its instantiation in verse. The definition of Maximum in (13) accounts for both classes of exceptions: those in (12) as well as those in (14).4 2.4
Loose Meters in English
In addition to the ‘‘strict’’ iambics discussed above, English poets, as well as poets writing in German and Russian, have employed what Robert Frost has called ‘‘loose’’ iambics. In addition to Frost, whose practice has been studied in Halle and Keyser 1999, loose iambics were employed by many other English poets, for example, Blake, Tennyson, Yeats, and Auden.5 The main di¤erence between strict and loose iambics involves the relationship between the footing and placement of Maxima. In strict iambics, placement of the Maxima is determined by the footing of the line: because of condition (9bi), Maxima can be placed only in head positions of the feet—that is, in even-numbered positions in iambic lines. In loose iambics it is the other way around—that is, the Maxima determine the location of the heads of the feet. Formally, we capture this di¤erence by replacing (9) with (15). This change includes the Accent rule (15b) which places a right parenthesis after each Maximum. Once the Maxima have been marked, the rule of Iterative Footing (15c) applies and inserts left parentheses from right to left, skipping two consecutive asterisks after each insertion. Since Iterative Footing is
On Stress and Meter and on English Iambics in Particular
17
ordered after the Accent rule, Iterative Footing respects the parentheses placed by Maximum marking. In sum, in loose iambics—and in loose meters, in general— footing of the line is governed by the placement of the Maxima, whereas in strict iambics, the placement of Heads is determined by Iterative Footing and the words must be chosen so that their Maxima coincide with the heads of the feet.6 The definitions and rules for English loose iambics are given in (15). (15) a. Definition A Maximum is i. the syllable bearing main stress in a polysyllabic word ii. a stressed monosyllabic word when followed by two syllables with less stress iii. a stressed monosyllabic word at the end of a major syntactic constituent b. Accent rule Insert a right parenthesis after a Maximum. c. Iterative Footing rule Insert left parentheses from right to left starting at the right edge of the line and skipping two asterisks in each iteration. Heads: right d. Condition Unfooted syllables and nonmaximal feet are admitted in all positions in the line. It is to be noted that the definition of Maximum for loose iambs in (15a) di¤ers from that of the Maxima in the strict meters in (13). The inclusion among Maxima of stressed monosyllabic words followed by two unstressed syllables expresses formally the fact noted, for example, by Tarlinskaja (1993, 57) that two consecutive stressed syllables in the line are separated by no more than two syllables. An example of loose iambs is Blake’s ‘‘The Sick Rose’’ in (16). (16) O Rose, thou art sick. (* *) * (* *( The invisible worm *( * *) * ( * *( That flies in the night (* *) * (* *( In the howling storm * (* *) (* *( Has found out thy bed (* *) * (* *( Of crimson joy, (* *) (* *(
1 2 3 4 5 6
18
Morris Halle
And his dark secret love * (* *) * (* *( Does thy life destroy. * (* * (* *) (
7 8
In (16) the left parentheses are inserted by the Accent rule (15b), and the right parentheses are inserted by the Iterative Footing rule (15c). In lines 1, 3, and 5 of (16), monosyllabic words are treated as Maxima, for they are followed by two syllables with less stress. The lines vary in length from four to six syllables, but in terms of feet there is no variation: each line has exactly two feet. The footing of the line explains the variations in syllable number on the assumption that the poem is subject to restrictions beyond those in (15). In particular, since the poem has only masculine rimes, unfooted syllables are not authorized in line-final position, nor are unary feet admitted anywhere in the poem. This admits five- and six-syllable lines into the poem in addition to four-syllable lines. Metrically the most interesting is line 7, for that line would have three feet if the lines were scanned as having no stress Maximum. The metric irregularity is removed by treating dark as a stress Maximum. This scansion is supported by the reading of dark secret as a compound adjective, like bright yellow (screen) or fair feathered (friend). (For some discussion of the metrical treatment of such structures, see Kiparsky 1975b, section 2.2, and Kiparsky 1977, section 6.) 2.5
In Conclusion
In the preceding I have attempted to show that a single grouping mechanism underlies both the assignment of stresses to words in di¤erent languages and the computations of meters in poetry. In section 2.2 the stress systems of a number of languages (Pintupi, Maranungku, Weri, and Finnish) was examined, and it was shown that the stress contours of the words in these languages were accounted for by grouping the syllables into pairs or triplets. In sections 2.3 and 2.4 I showed that the same kind of grouping of syllables into pairs or triplets accounts also for the main meters of English poetry. Both in stressing words and in computing meters, the grouping of the syllables is formally implemented by inserting diacritic markers—parentheses—among the syllables in the sequence. A left parenthesis groups the syllables on its right, a right parenthesis groups the syllables on its left. The grouping is carried out by a version of the rule of Iterative Footing (cf. (3b) and (9a)). The grouping operation, moreover, designates as head the left-, resp. right-most syllable in each of the groups. In the strict meters, the grouping of the syllables (footing) is determined exclusively by the Iterative Footing rule. In loose meters, grouping is determined by the interac-
On Stress and Meter and on English Iambics in Particular
19
tion of Iterative Footing with an Accent rule such as (15b). We encounter a parallel di¤erence in stress systems: in Maranungku and the other languages illustrated in (1), stress distribution is determined entirely by a rule of Iterative Footing; in Finnish, by contrast, stress assignment involves the Accent rule (3a) in addition to the Iterative Footing rule (3b). An important di¤erence between stress systems and meters is that in stress systems a phonetic mark—often a special pitch—is assigned to the heads of the feet, but no phonetic mark is supplied to the heads in meters; the heads of the di¤erent feet are rather subject to special conditions, such as (9bi), requiring maxima to appear in head position exclusively. Both stress and meter require information about head placement, which can be computed only by grouping the syllables in the sequence. It is therefore the grouping of syllables into feet (and into higher order groups) that underlies the parallels between stress and meter first noted by Liberman 1975. In the seven years that have elapsed since the original version of this paper was sent to the editors, Nigel Fabb and I have studied meters in many languages including, among others, French, classical Greek, Spanish, Italian, Arabic, Vietnamese, and Sanskrit. These studies have shown that with the notable exception of the metrical poetry of the Old Testament, all meters of all languages involve grouping of syllables into pairs or triplets of the same kinds as that met with in the English examples in this chapter. For details interested readers are referred to Nigel Fabb and Morris Halle (to appear in 2008) Meter in Poetry: A New Theory, Cambridge University Press. Notes I am grateful to Sylvain Bromberger, David Embick, John Frampton, Kristin Hanson, Jay Keyser, Michael Kenstowicz, Bert Vaux, as well as to an anonymous reviewer for helpful comments. 1. Hayes 1995, chapter 8 discusses additional instances of ternary stress patterns and o¤ers an alternative account of these as special cases of binary footing. I am not altogether persuaded by Hayes’s proposals, but because of the complexity of the issues involved, a discussion of these proposals must be deferred to another occasion. 2. I am indebted to Michael Kenstowicz for drawing my attention to this fact. 3. The word love`d in (11d) is counted as disyllabic for purposes of the meter. This is another deviation from the syllabification of normal speech that is common in English poetry (see Major 1901 for some discussion). Like the omission of word-medial syllables, this type of syllable epenthesis, is employed only sparingly. We take formal account of these facts by adding appropriate stipulations to the list of conditions (9b), noting in particular their marked status. 4. The treatment discussed here is along the lines of Halle and Keyser 1971.
20
Morris Halle
5. Extensive use of loose meters is found in German poetry. Among poems in these meters are such ‘‘war-horses’’ as Goethe’s Der Erlko¨nig and Heine’s Die Grenadiere and Lorelei. In an unpublished lecture (Harvard, February, 1997), I have shown that loose meters of all kinds— trochees, dactyls, and anapests in addition to iambs—are widely employed in Russian poetry, where they are referred to by the term dol’nik. 6. A parallel di¤erence is found in stress systems. As shown in section 2.2, Finnish word stress parallels the loose meters, whereas the stress systems of Maranungku, Pintupi, and Weri parallel the strict meters.
3
Nonlexical Word Stress in the English Iambic Pentameter: A Study of John Donne
Kristin Hanson
3.1
Introduction
In their study of Chaucer that inaugurated generative metrics, Halle and Keyser (1966, 201–202) suggest that the stress borne by certain function words might be irrelevant to the fundamental question of whether a line is metrical or not. As metrical theory has changed with the vicissitudes of phonological theory, that idea has been disputed (Wimsatt 1970, 777–782; Magnuson and Ryder 1970, 791–795; 1971, 205– 206), clarified and defended (Halle and Keyser 1971b, 156–157, 163–164), explicitly extended to the modern English tradition and justified (Kiparsky 1975b, 249–250), abandoned (Kiparsky 1977, 219–221), and neglected (Hayes 1989b). My purpose here is to restore it, in a revised form, as a claim that in the modern English iambic pentameter, as in other meters (Hanson 1992; Hanson and Kiparsky 1996, 321–325), whether metrical constraints apply to lexical or postlexical prosodic structure is a significant parameter of variation across poets. My argument comes from the metrical treatment by three poets of disyllabic function words. Drawing on Liberman and Prince’s (1977) model of linguistic rhythm or prosody as fundamentally relational, Kiparsky’s (1977) study of the English iambic pentameter shows that syllables that are prosodically strong relative to other syllables in the same word (1a) are always highly constrained in that meter, while those that are strong only relative to other syllables in di¤erent words (1c) are treated more laxly: (1) a. amo´k hu´nter
b. amo´ng u´nder
c. a mo´nk hu´nt for
But what counts as a word with respect to such constraints? While function words (1b) might intuitively be expected to pattern with the content words of (1a), and are predicted by Kiparsky’s (1977) analysis to do so, I will argue here that this is correct only for some poets, such as, apparently, Shelley (Hayes 1989b). For other poets function words pattern instead with the phrases in (1c). This is subtly so for
22
Kristin Hanson
Shakespeare, but strikingly so for Donne, for whom the prosodically strong syllables of (1b,c) are not constrained at all, while those of (1a) are. In fact, the possibility of a crucial distinction between (1a) and (1b,c) alongside the more intuitive one between (1a,b) and (1c) reveals systematicity in the metrical practice of Donne, the English poet most renowned for the putative irregularity of his iambic pentameter. The theory of Lexical Phonology (Kiparsky 1982d) provides a principled explanation for both possible patternings of function words. On that theory, the phonology of a language is organized into distinct components according to its morphology. The broadest distinction is between the lexical phonology, which determines the phonological structure of words in the special sense of being the output of the wordformation component of a language’s grammar, called lexical words; and the postlexical phonology, which determines the phonological structure of syntactic configurations more broadly, including that of words in the more general sense of being syntactic terminals. In English, nouns, verbs, adjectives, and morphologically complex adverbs are lexical words, and their prosodic structure is determined by the lexical phonology. Function words—which include conjunctions, prepositions, pronouns, copular and auxiliary verbs, and adverbs of the type traditionally referred to as particles (see below)—are nonlexical words and therefore have their prosodic structure determined by the postlexical phonology (Inkelas and Zec 1993). The prosodic structure of phrases must also be determined by the postlexical phonology, by definition. Hence, a metrical constraint on prosodic strength within words will apply to (1a,b) if it takes postlexical structure into account, but to (1a) alone if it respects only lexical structure. By the same token, characteristic treatments of postlexical prosodic structure may be manifest either in (1c) alone, or jointly in (1b,c). The chapter is organized as follows. Section 3.2 summarizes its basic assumptions about the English iambic pentameter and the relevant prosodic properties of the language. Section 3.3 summarizes Kiparsky’s (1977) description of Shakespeare’s treatment of the three types of structures in (1), and proposes a reanalysis of his treatment of the disyllabic nonlexical words of (1b) based on the lexical/postlexical distinction. Section 3.4, the core of the chapter, shows that this approach permits a systematic account of Donne’s metrical practice. Finally, section 3.5 briefly contrasts Donne’s practice with Hayes’s (1989b) description of Shelley’s. 3.2 3.2.1
Metrical and Prosodic Assumptions Basic Structure of the English Iambic Pentameter
Following Halle and Keyser 1966, 1971b, a meter is assumed here to be defined by an abstract underlying structure, and a set of constraints on how language can be mapped into that structure. The phonological structure of the language itself is taken as a given, although it should be noted that it may di¤er slightly from the phonology
Nonlexical Word Stress in the English Iambic Pentameter
23
of the ordinary language in ways that will not concern us here (Kiparsky 1977, 190, 235–244). A scansion is a mapping of a line into the underlying structure that satisfies the constraints, and a line that has a scansion is metrical. The underlying structure of the English iambic pentameter consists of five metrical feet (Kiparsky 1977), each composed of a weak (W) metrical position followed by a strong (S) metrical position (Halle and Keyser 1971b). The feet are in turn grouped into cola, and the cola into a line. These constitutents too are binary with a rising rhythm, but the odd number of feet in the pentameter line poses complications. Both synchronic and diachronic metrical evidence shows that the basic underlying structure of the English iambic pentameter has an initial colon as in (2a), but an alternative (2b) in which the third foot also belongs to the intial colon is available as well (Kiparsky 1977; Piera 1981).1 (2) a.
b.
The ambiguity made available by (2) functions in scansion just like ambiguity made available by phonological options a language may allow: so long as there exists some mapping of some legitimate phonological structure of a line into some legitimate underlying metrical structure of the meter in question, the line is metrical. One constraint defining the English iambic pentameter limits the amount of linguistic material that can be mapped into any single metrical position of (2). The norm is a single syllable, apart from standard elisions and the possibility of phonologically weak extrametrical syllables (‘‘hi’’) at the ends of lines and major phrases (Kiparsky 1975b; 1977, 230–236). All three poets considered here, however, also allow two unstressed syllables in W, as in (3), or a stressed light syllable followed by an unstressed syllable in the same word together in S, as in (4): (3) a. I’ th’ name of fame and honor, which dies i’ th’ search w s w s w s hi w s w s (Shakespeare, Cymbeline 3.3.51) b. To feed on that, which to’disus’d tasts seemes tough. w s w s w sw s w s (Donne, ‘‘Elegie: His Picture’’ 20)
24
Kristin Hanson
c. Of the riven soul, sent its foul dreams to sweep w sw s w s w s w s (Shelley, ‘‘The Revolt of Islam’’ 1300) (4) a. And spends his prodigal wits in bootless rhyme w s w s w s w s w s (Shakespeare, Love’s Labour’s Lost 5.2.64) b. Then the wise politique horse would heretofore, w s w s w s w s ws (Donne, ‘‘Satyre I’’ 80) c. I looked, and lo! one stood forth eloquently, ws w s w s w s w s (Shelley, ‘‘The Revolt of Islam’’ 649) Formally this means that their iambic pentameter is best analyzed as limiting each position to a moraic trochee (Hanson and Kiparsky 1996). Although the distinction between that and the more familiar limiting of each position to a syllable will not figure directly in our analysis of the distinctions in (1) apart from constraining the scansions on which it rests, it is worth noting that the analysis actually lends further support to the argument at hand.2 The other constraint defining the English iambic pentameter, which is the one of central concern here, specifies how the rhythmic contrast between W and S is manifest. As Halle and Keyser (1966) observe, S is free to contain any kind of syllable (e.g., the unstressed its of (3c), the of (4b), or -ly of (4c)), but W is not. As mentioned above, Kiparsky (1975b, 1977) argues that across the tradition, syllables that are prosodically strong relative to other syllables within the same word are generally prohibited in W. Thus in (3) and (4), for example, the stressed syllables of the polysyllabic honor, disus’d (cf. 27b), riven, prodigal, politique, and eloquently are all in S, while stressed monosyllables can occur either in S or in W, the latter case appearing in seemes (3b), foul (3c), or wise (4b). 3.2.2
Lexical Prosodic Structure
Kiparsky’s formalization of the class of syllables prohibited in W is intended to account for not only the simple distinction between polysyllables and monosyllables illustrated above but also an entire gamut of subtle distinctions poets’ practices seem sensitive to. The correct formalization has been a matter of debate ever since Halle and Keyser’s first proposals, partly because of the complexity of the class, but also because available options depend on the theory of prosodic phonology. That theory has developed rapidly since that time, including undergoing significant changes since Kiparsky’s (1977) description of the patterns under discussion here. The most substantive of these changes has been the positing of the phonological foot (‘‘f’’), as part of a universal hierarchy of prosodic constituents into which lan-
Nonlexical Word Stress in the English Iambic Pentameter
25
guage is held to be organized, which also includes the syllable (‘‘s’’) and the phonological word (‘‘o’’). In this organization each of these constituents is assumed to consist of a group of constituents of the type below it in the hierarchy, among which exactly one member is the head of the superordinate constituent. An exception is constituents created by adjunction, in which a constituent that is not in a position to be grouped into a well-formed constituent of the superordinate type joins an adjacent one as a weak sister, forming a new constituent of the same type it joins, indicated here by the superscript ‘‘ 0 ’’. In the distinction drawn by Liberman and Prince (1977), which plays a crucial role in Kiparsky’s (1977) description, constituents that also contain a nonhead are branching; their heads are strong and their nonheads weak, indicated here with the subscripts ‘‘s ’’ and ‘‘w ’’, respectively. A terminal element dominated only by heads within a constituent is that constituent’s designated terminal element (DTE). All surface prosodic phenomena are assumed to be reflections of this structure: most notably, a stressed syllable is one that heads a foot, and a primary stressed syllable is the DTE of a phonological word. For the English words to be considered here, it will be assumed that syllables are grouped into left-headed feet, specifically moraic trochees, although ones whose weight patterns are sometimes disturbed by resolution and resyllabification. It will also be assumed that any syllable adjunction is leftward if possible, otherwise rightward. These assumptions are consistent with certain major generalizations about English prosodic phonology—for example, that vowels may be reduced only in unstressed syllables, and that expletives may be inserted within words only between feet. For various aspects of all these assumptions, see Liberman and Prince 1977; Hayes 1980, 1995; McCarthy 1982; McCarthy and Prince 1986; Selkirk 1984; Kager 1989; Hanson and Kiparsky 1996; Hanson 2001. On these assumptions, the following sample words whose metrical distribution in Shakespeare is described by Kiparsky (1977) would have the structures in (5):3 (5) Lexical stress in English
26
Kristin Hanson
The syllables Kiparsky (1977) found to be generally prohibited in W are those underlined. They have in common that they are the DTE of a branching consituent within o. This property thus formalizes what it means for a syllable to be strong relative to another syllable in the same word (Hanson and Kiparsky 1996). Note that this class of syllables is only a small subset of those that are stressed:4 wise in (5a), mainin (5d), -tile in (5e), -fies in (5g), and fan- in (5i) are all stressed, but not strong, and occur in W. Thus, the constraint that Kiparsky (1977, 194–195) takes as basic to the English iambic pentameter can be updated as in (6): (6) Basic prominence constraint of English iambic pentameter A strong syllable is prohibited in W if its weak sister is within the same o. (cf. Kiparsky’s (1977) (9b)) 3.2.3
Postlexical Prosodic Structure
The metrical treatment of words is to be compared here with that of phrases; but the theory of prosodic phonology has also developed in ways that a¤ect what structure
Nonlexical Word Stress in the English Iambic Pentameter
27
phrases can be assumed to have. Kiparsky’s (1977) description assumes that for phrases the constituent structure to which meter is sensitive is syntactic structure, but Hayes (1989b) has shown that it is more plausibly the constituent structure given by the prosodic hierarchy, derived from syntactic structure but not identical to it (Selkirk 1984; Nespor and Vogel 1986; Hayes 1989b). Various versions of that hierarchy have been proposed for English; the specific one assumed here will be that of Inkelas and Zec 1993, because it gives an explicit account of the third point of comparison at issue here, the prosodic structure of nonlexical words. On Inkelas and Zec’s model, postlexical prosodic structure is based on a syntactic representation in which the individual words (the terminal elements X 0 in (10) below) have exactly the structure given by the Lexical Phonology. Thus, lexical words have foot and phonological word structure as shown in (5), but nonlexical words do not. Both types of words are grouped together into phonological phrases (‘‘j’’) according to the algorithm in (7), giving sample phrasings such as those in (8). Phonological phrases themselves are further grouped into intonational phrases (‘‘i’’), again by pairing those most closely related in the syntax (Hayes 1989b): (7) Phonological phrasing in English (Inkelas and Zec 1993, 222) a. Branching clause From the bottom up in a syntactic tree, build an obligatory branching phonological phrase over adjacent phonological words that correspond to syntactic sisters. b. Adjunction clause Phrase a terminal element that lacks phonological word structure together with its immediate sister. c. Default clause Map any terminal element left unphrased by the other clauses into a phrase of its own. (8) a.
b.
28
Kristin Hanson
c.
It is within the context of the phonological phrase that the special status of nonlexical words emerges. Phrasal stress in English normally falls on the final stressed word of a phrase (Chomsky and Halle 1968; Liberman and Prince 1977), both at the phonological phrase level, and again at the intonational phrase level. Thus in (8b), for example, Jim bears more stress than bought within the phonological phrase they share, as well as more stress than chair within the intonational phrase. But nonlexical words do not bear phrasal stress in final position: in (8c), otherwise parallel to (8b), phrasal stress falls on fought, not for; and analogously in (9) it falls on stepped and talking, not over or about: (9) a. that chair he ste´pped over b. that chair they were ta´lking about At the same time, nonlexical words in final position do not lack stress. Of course, they may sometimes have stress for special rhetorical purposes, such as contrastive stress. But even under ordinary stress patterns, a reduced vowel is not possible for most function words in phrase-final position—in (8c), for example, a reduced vowel in for, common in other positions as in a chair for Jim, is impossible—and inability of vowels to reduce is normally attributed to their syllable being stressed.5 More strikingly, disyllabic nonlexical words as in (9) obviously have fixed stress regardless of their position within a phrase: only o´ver and abo´ut are correct, never *ove´r or *a´bout. Given that it is a general assumption of prosodic phonology that only syllables that already have stress at one level can manifest stress at a higher level—it must be talk- that manifests the word talking’s phrasal stress in (9b), not -ing—the hypothesis that only lexical prosodic structure is present in the representation to which phrasal stress is assigned is what Inkelas and Zec (1993) claim accounts for the lack of phrasal stress on nonlexical words. Specifically, they propose that nonlexical words are assigned stress only later, in conjunction with a default mapping of all syntactic terminals into phonological words (shown in all representations below). Within those postlexical phonological words, stress is assigned to disyllabic words according to the
Nonlexical Word Stress in the English Iambic Pentameter
29
principle that the final syllable will be stressed if and only if it is heavy, and to monosyllabic words if they are phrase-final, as defined by (7). These rules are summarized in (10): (10) Nonlexical stress assignment in English (Inkelas and Zec 1993, 222) a. Default phonological word mapping: X 0 ! [X 0 ]o b. Within a stressless phonological word, build an obligatorily binary foot with the right node strong if and only if it branches. c. Build a foot on a phrase-final phonological word that lacks metrical structure. On this approach, then, the fact that the stress of nonlexical words as in (1b) is phonologically distinct from that of lexical words as in (1a) is formalized as its being assigned only postlexically, like that in (1c), with the result that it is irrelevant to certain phenomena in the grammar of rhythm, such as assignment of phrasal stress. We are now in a position to see that this state of a¤airs is precisely what the metrical practice of English poets reflects. 3.3 3.3.1
Shakespeare Lexical Words
Kiparsky’s (1975b, 240–241) description of Shakespeare’s metrical practice takes the prohibition in (6) against W containing syllables that are strong within words to be almost absolute. One of only two systematic sets of exceptions he notes involves a very general phenomenon whereby certain metrical constraints are relaxed on initial positions of major domains; thus all English poets allow exceptions to (6) in the first position of a line, as in (11a), and many, including Shakespeare, also do so lineinternally for the first syllable of phonological or higher-level phrases as in (11b,c) (Kiparsky 1975b, 1977, 217; Hayes 1983, 1989b, 247–249): (11) a. Angels and ministers of grace defend us! w s w s ws w s ws hi (Hamlet 1.4.39) b. Brief let me be. Sleeping within mine orchard, w s w s w s w s w s hi (Hamlet 1.5.59)
30
Kristin Hanson
c.
Lines like those in (12), in contrast, in which W contains a syllable that is strong within a word and also phrase-internal, Kiparsky claims do not occur: (12)
Nonlexical Word Stress in the English Iambic Pentameter
31
In fact this description may be rather too severe, because exceptions have periodically been noted (Youmans 1983; Golston 1998). I will return to this point briefly in section 3.4 after having explored the consequences of Kiparsky’s description more fully. 3.3.2
Phrases
The constraint in (6) bans from W syllables that are strong relative to others within the same word, but Kiparsky (1977) observes that syllables that are strong relative to an adjacent syllable in a di¤erent word in the same phrase are also sometimes restricted. Most relevant to the issue here is the case when the adjacent syllable is a nonlexical word, as in collocations like a monk or hunt for in (1c).6 For these cases, scansions with the strong syllable in S as in of grace in (11a) are clearly the norm, with exceptions allowed only under certain conditions. Kiparsky’s (1977) characterization of these conditions forms part of an argument that metrical well-formedness is sensitive not only to syllable count and prosodic prominence but also to alignment of linguistic and metrical constituency. Specifically, he notes that strong syllables of phrases may occur in W if the beginning of the constituent within which the mismatched syllable’s strength is defined coincides with the beginning of a metrical foot, as in (13) (Kiparsky 1977, 206–209):7
32
Kristin Hanson
(13)
a.
b. If I dream not, thou art Aemilia. ws w s w s w sws (The Comedy of Errors 5.1.347) c. Thou dost love her because thou know’st I love her, w s w s ws w s ws hi (Sonnets 42.6) d. Or how haps it I seek not to advance w s w s ws w s w s (1 Henry VI 3.1.31) They may also occur in W if that entire constituent is itself subordinated to a still stronger syllable, as in (14) (Kiparsky 1977, 206–209) or even (15) (Kiparsky (1975b, 243): (14)
a.
b. If I lose thee, my loss is my love’s gain, ws w s ws w s w s (Sonnets 42.9) c. I understand thy kisses, and thou mine, ws w s w s w s w s (1 Henry IV 3.1.202)
Nonlexical Word Stress in the English Iambic Pentameter
33
d.
. . . Hang there like fruit, my soul, Till the tree die! w s w s (Cymbeline 5.5.263–4) e. Of the world’s ransom, blessed Mary’s Son; w s w s w s w s w s (Richard II 2.1.56) (15)
a.
b. I do suspect I have done some o¤ense w s w s w s w s w s (Richard III 3.7.111) c. When I come where he calls, then he is gone. w s w s w s w sw s (A Midsummer Night’s Dream 3.2.414) d. Shall we hear from you, Catesby, ere we sleep? w s w s w s ws w s (Richard III 3.1.188) Lines in which neither condition holds, as in (16), are claimed not to occur, or at least to be so rare as to require inclusion of some restriction on them as part of the meter’s description (Kiparsky 1977, 207; Youmans 1983): (16)
a.
34
Kristin Hanson
b.
Syllables that are strong relative to an immediately preceding weak syllable in a di¤erent lexical word are restricted in the same way. Thus Kiparsky (1977, 207) claims that lines like Wyatt’s in (17) also do not occur in Shakespeare: (17)
Strong syllables of phrases are fine in W, however, if the preceding syllable is a lexical word in its own right (Kiparsky 1977, 208):
Nonlexical Word Stress in the English Iambic Pentameter
35
(18)
Kiparsky’s (1977, 205–206) formal description of this pattern captures so many nice distinctions that it is risky to tamper with it, but some modification is required in light of the more elaborate prosodic structures entailed by the assumption of the phonological foot described in section 3.2.2, and the assumption described in section 3.2.3 that it is the prosodic hierarchy and not the syntax directly that defines metrically relevant constituent relationships. Many relationships that are immediate under Kiparsky’s assumptions become more distant once these new assumptions are taken into account.8 For example, Kiparsky’s description refers to the relationship of command defined as in (19), but in (17), the syllable blode does not command -ent as his formalization intends it to, only the phonological word does. Similarly in (14a), the phrase her cheek is not a weak sister of melts in the same way it would be if their prosodic relationship depended directly on the syntax, because within the prosodic hierarchy its strength relative to which intervenes. Moreover, Kiparsky’s own (1977, 209) description of the subordination that allows (14) and (15) is consciously vague: he notes that the domain within which subordination is metrically relevant is limited in a way that is di‰cult to define. Hayes (1989b) argues that such limitations refer to the prosodic hierarchy: on the specific assumptions about phrasing outlined in section 3.2.3, the relevant subordination is often within the same phonological phrase (Hayes 1989b), but always within the same intonational phrase as in (14a).9 Hayes (1983, 384) also observes that the relevant subordination tends to be to the syllable in the S immediately following the mismatched strong one as in (14), but while such adjacency plausibly plays a metrical role, it does not seem to be a requirement here, since it does not hold for the lines in (15). Rather, what these lines seem to suggest is that the relevant limitations may also refer to the metrical hierarchy shown in (2), and in particular to colon boundaries, whose signficance has been somewhat neglected (Youmans 1983). The lines in (15) have in common that a rising triple rhythm is repeated twice, concluding with a strong syllable in the strong position of one colon, the latter property being shared with those in (14). Clearly any attempt to improve on Kiparsky’s description would be a major undertaking in its own right, but on the basis of his cited examples we can hazard a guess that both limitations might be relevant, a point to which we return briefly in section 3.4.
36
Kristin Hanson
Taking these considerations into account, then, we can use Kiparsky’s (1977, 205) definition of command in (19), and update his description (p. 206) as follows: (19) P commands Q if and only if there is a node R that dominates Q and immediately dominates P. (20) Additional constraint on the English iambic pentameter A strong syllable s1 is prohibited in W if a. The immediately preceding syllable s2 is not the DTE of a lexical o (¼ Kiparsky’s (1977) (46i)) and b. The head of the constituent k1 of which s1 is the DTE commands s2 (¼ Kiparsky’s (1977) (46ii)) and c. k1 is not itself commanded by a strong constituent k2 whose DTE s3 is within the same i as s1 and within the same colon, and is itself in S (¼ Kiparsky’s (1977) (46iii)) Stated less formally, (20)’s prohibition under a conjunction of three conditions is equivalent to allowance under any of their opposites. Thus (20a) allows (18) by confining the prohibition against strong syllables in W to cases where the syllable immediately preceding the mismatched prominent one is nonlexical, unstressed, or only secondarily stressed, and so to sequences that have the kinds of stress patterns words themselves might have, a property we might somewhat fancifully call lexisimilitude. (20b) allows the lines in (13) because their mismatched strong syllables are not grouped with the syllable to their left, but instead begin new constituents aligned with the beginnings of metrical feet, a property we will call left alignment. Finally, (20c) allows the lines in (14) and (15) because their mismatched strong syllables are subordinated to correctly matched stronger syllables, a property we will call right subordination. Thus (20) in e¤ect identifies various mappings that do or do not a‰rm aspects of the underlying metrical pattern (Magnuson and Ryder 1971), and states that while those that do not are not unmetrical in and of themselves, neither can they be multiplied without limit, requiring instead some compensatory a‰rmation from ones that do. For Shakespeare, (6) constrains the distribution of syllables that are strong within words, but the less absolute set of conditions in (20) constrains syllables that are strong only within phrases. 3.3.3
Nonlexical Words
The phonological theory within which Kiparsky (1977) frames these generalizations distinguishes prosodic strength within words from that within phrases by the respective absence or presence of a word boundary between the strong syllable and its weak
Nonlexical Word Stress in the English Iambic Pentameter
37
sister, predicting that nonlexical words should pattern just like lexical words with respect to (6), never occurring with their strong syllables in W. But they do not. Rather, they constitute the second systematic class of exceptions to (6): as Kiparsky himself observes (1977, 220; 1975b, 249–250), although prepositions are most frequently positioned with their strong syllables in S as in (21), they also sometimes occur with them in W as in (22) and (23): (21) a. And mock you with me after I am gone. w s w s ws w s w s (Sonnets 71.14) b. And make me travel forth without my cloak, w s w s w s w s w s (Sonnets 34.2) (22) a. To glean the broken ears after the man w s w s w s w s w s (As You Like It 3.5.102) b. Like a white hind under the gripe’s sharp claws, w s w s w s w s w s (The Rape of Lucrece 543) c. And with declining head into his bosom, w s w s w s w s w s hi (The Taming of the Shrew Ind.1.119) (23) a. Henceforth be never numb’red among men! w s w s w s w s w s (A Midsummer Night’s Dream 3.2.67) b. Words before blows; is it so, countrymen? w sw s ws w s w s (Julius Caesar 5.1.27) c. Weigh’d between loathness and obedience, at w s w s w s ws w s (The Tempest 2.1.131) d. And I will comment upon that o¤ense; w s w s w s w s w s (Sonnets 89.2) Kiparsky’s explanation assimilates the exceptional behavior of these words to exceptional behavior requiring explanation in any case for lexical words. First, as already noted in (11), Shakespeare allows strong syllables of words in W phraseinitially. When the strong syllable of a preposition is initial in its word as in (22), because of the syntactic distribution of prepositions it is also normally initial in its phrase as in (24), and so can occur in W:
38
Kristin Hanson
(24)
In support of this position, Kiparsky (1977, 221) notes the absence of lines like the following in which a mismatched preposition is preceded by a proclitic and so, as Hayes (1989) observes, would not be initial in its phrase: (25) *Unlock this casket, if after three nights w s w s w s w s w s (construct) Second, apparent exceptions to the constraint on strong syllables of words in W arise for lexical words from the operation of the Rhythm Rule, which repairs the arrhythmicality of adjacent strong constituents as in (26a) by adjusting them to the form in (26b), subject to the constraint that the first W, if terminal, must be stressed (Kiparsky 1977, 218). The Rhythm Rule thus shifts, among other things, primary stress from a syllable at the end of a word to a secondarily stressed syllable at the beginning of a word when the word is subordinated in its phonological phrase to one beginning with an even more strongly stressed syllable, making the stress pattern of a word like thı`rte´en di¤erent in isolation from what it is in a phrase like thı´rte`en me´n (Liberman and Prince 1977; Inkelas and Zec 1993; Nespor and Vogel 1986; Hayes 1984, 1989b, and others). Shakespeare’s metrical practice reflects the operation of the Rhythm Rule in the language itself; thus words like forlorn are systematically positioned with their final syllable in S as in (27a) except when the conditions for the Rhythm Rule are met as in phrases like forlorn world in (27b), suggesting that the actual stress pattern of the word is di¤erent in the two lines (Kiparsky 1975b, 247; 1977, 218): (26) a.
b.
Nonlexical Word Stress in the English Iambic Pentameter
39
(27)
Kiparsky (1977, 220) claims that words like divine that do not have any stress on their initial syllable and so remain una¤ected by the Rhythm Rule are, in contrast to (27b), positioned with their strong syllable in W even when subordinated to a more strongly stressed syllable within their phrase as in (28): (28)
The prepositions with final stress in (23) are typically subordinated to a following stronger stress, just as forlorn is in (27b):
40
Kristin Hanson
(29)
They do not, however, have the stress on their initial syllable that forlorn does and that the operation of the Rhythm Rule requires, being in that respect instead like divine in (28). Kiparsky (1977, 220) addresses this similarity and this di¤erence by suggesting that the Rhythm Rule could nonetheless still be the explanation for the positioning of the words in (23), if it is assumed to have applied more freely to function words, suspending the usual requirement that the syllable onto which stress is shifted bear secondary stress. However, given the theoretical distinction between the prosodic structure of lexical words and that of nonlexical words since shown to be available independently in the phonology, such an otherwise empirically unsubstantiated and theoretically problematic supposition of a rhythmic adjustment is unnecessary. The metrical distribution of the strong syllables of the nonlexical words of (22) and (23) is simply like that of the strong syllables of the phrases in (13) and (14)–(15). If they are initial in their word as in (22) they are allowed in W because (20b) does not obtain; that is, they are left-aligned. And if they are final as in (23) they are allowed because (20c) does not obtain; that is, they are right-subordinated. For strong syllables to usually be in S according to (6) but occasionally in W according to the set of conditions in (20) is thus the characteristic distribution for Shakespeare of syllables that are strong in postlexical prosodic structure, a¤ecting nonlexical words and phrases alike. Considering Shakespeare alone, such a reanalysis accompanies its theoretical gain with only scanty empirical improvement. Certain additional problems that might arise in treating the Rhythm Rule as applicable to a line like (23d) that has no adjacent stressed syllables (in spite of having the same tree geometry at a higher level as the unseen good old man (Kiparsky 1977, 219)) are avoided. On the other hand, the explanation for the absence of lines like (25) evaporates; all that can be said is that since the number of prepositions with their strong syllables in W is already quite a small proportion of the total (Magnuson and Ryder 1971), it is not clear that the absence of such lines should be endowed with much significance. Nonetheless, the reanalysis does e¤ect a significant improvement in the account of the tradition as a
Nonlexical Word Stress in the English Iambic Pentameter
41
whole, because it extends naturally to explain the distribution of these words in Donne, to whom we now turn. 3.4
Donne
Donne’s metrical practice is famous for being among the laxest within the entire tradition of the English iambic pentameter. Noting that Donne allows strong syllables of words in W where no other poet in his study does, Kiparsky (1975b, 258–259) takes this reputation for liberty as confirmation of his analysis of the centrality of word stress in the rest of the tradition. I propose to turn the argument around and show that Donne’s practice is actually systematic in a way that confirms Kiparsky’s analysis of the more subtle stress relations within phrases. This accords with Stein’s (1944, 1962) defense of Donne’s systematicity, and also with the findings of Tarlinskaja and Teterina’s (1974) statistical study of Donne’s ‘‘Satyres,’’ which places the ‘‘Satyres’’ beyond what their study takes to be the threshold of metricality for English iambic pentameter with respect to the number of mismatches between metrical and prosodic prominence in them, but not with respect to the distribution of those mismatches. 3.4.1
Lexical Words
In Donne’s iambic pentameter, as in Shakespeare’s, most lexical words have their strong syllables in S, as does, for example, wander in (30a). There are also exceptions of the same kind allowed by Shakespeare at the beginnings of phonological and higher-level phrases, as in (11): (30)
a.
b. Which, as the soule quickens head, feet and heart, w s ws w s w s w s (‘‘Elegie: The Bracelet’’ 37)
42
Kristin Hanson
c. Dost woe my soule, for hers o¤ring all thine: w s w s w s w s w s (‘‘Holy Sonnet: ‘Since she whome I lovd’ ’’ 10) d. Immensity cloysterd in thy deare wombe. w s ws w s w s w s (‘‘La Corona’’ 28) But there are also quite a few exceptions of other kinds, which are strikingly similar to those described by the set of constraints in (20). Strong syllables of lexical words also occur in W under left alignment, parallel to Shakespeare’s phrases in (13) and nonlexical words in (22) allowed by (20b): (31)
a.
b. Doth with doubtfull melodious murmuring, w s w s ws w s ws (‘‘Elegie: ‘Oh, let mee not serve so’ ’’ 22) c. That would have need to be pittied by thee? w s w s w s ws w s (‘‘La Corona’’ 40) d. By sicknesse, deaths herald, and champion; w s w s ws w s ws (‘‘Holy Sonnet: ‘Oh my blacke Soule!’ ’’ 2) e. And then as worst surfets, of best meates bee, w s w s w s w s w s (‘‘The Crosse’’ 39) f. Make sinnes, else equall, in mee, more heinous? w s w s w s w s w s (‘‘Holy Sonnet: ‘If poysonous mineralls’ ’’ 6) g. Bu¤et, and sco¤e, scourge, and crucifie mee, w s w s w s wsw s (‘‘Holy Sonnet: ‘Spit in my face’ ’’ 2) h. Hee keepes, and gives to me his deaths conquest. w s w s w s w s w s (‘‘Holy Sonnet: ‘Father, part of his double interest’ ’’ 4)
Nonlexical Word Stress in the English Iambic Pentameter
43
This practice is not unprecedented in the tradition at Donne’s time; it is also found in Wyatt (as well as later in the tradition in other poets, most notably Milton) (Kiparsky 1977, 202): (32) a. There is written her faier neck rounde abowte: w s w s w s w s ws (VII: ‘‘Who so list to hounte’’ 12) b. Ffyre that purgith allthing that is vnclene w s w s w s w s w s (LXXVI: ‘‘Venemus thornes’’ 5) c. Thancked be fortune, it hath ben othrewise w s w s w s w s w s (XXXVII: ‘‘They fle from me’’ 8) But unlike Wyatt, Donne also sometimes allows strong syllables of lexical words in W where this condition is not met. The lines in (33a), (34b), (37c), and (38b) are cited by Kiparsky (1975b, 258–259) as evidence of Donne’s anomalousness with respect to the English iambic tradition, but additional lines of the same type given here show that they are not anomalous with respect to Donne’s own practice. Nor are they random violations of (6). Rather, all involve some form of right subordination. Those in (33) are directly parallel to Shakespeare’s phrases in (14) and nonlexical words in (23a,b) allowed by (20c): (33)
a.
b. I, like an usurpt towne, to’another due, w s w s w s w s w s (‘‘Holy Sonnet: ‘Batter my heart’ ’’ 5)
44
Kristin Hanson
c. O perverse sexe, where none is true but shee, w s w s w s w s w s (‘‘Twicknam garden’’ 26) d. Or is thy devout Muse retyr’d to sing w s w s w s ws w s (‘‘To Mr. R.W.: ‘Zealously my Muse’ ’’ 9) e. And begin soone, lest my griefe grieve thee too, w s w s w s w s w s (‘‘To Mr. B.B.’’ 9) f. Remembers what of old, shee esteem’d most, w s w s w s w s w s (‘‘The Lamentations of Jeremy’’ 27) g. As China, when the Sunne at Brasill dines. w s w s w s w s w s (‘‘To the Countesse of Bedford: ‘You have refin’d mee’ ’’ 18) h. Of refin’d manners, yet ceremoniall man, w sw s hi w s w s w s (‘‘Satyre I ’’ 28) Those in (34) di¤er from Kiparsky’s examples from Shakespeare in that the correctly matched syllable, while still in the same colon, is not in its DTE: (34) a. I said to all my profane mistresses, ws ws w s w s w s (‘‘Holy Sonnet: ‘What if this present’ ’’ 10) b. If faithfull soules be alike glorifi’d w s w s w sw s ws (‘‘Holy Sonnet: ‘If faithfull soules’ ’’ 1) Note that the scansions of (33) and (34) cannot be consistently explained as e¤ects of the Rhythm Rule, even though that rule certainly a¤ects stress in Donne in the same way it does in Shakespeare, giving pairs of lines like (35) parallel to (27): (35) a. No Crosse is so extreme, as to have none; w s w s w s w s w s (‘‘The Crosse’’ 14) b. But extreme sense hath made them desperate; w s w s w s w s ws (‘‘Elegie: Death’’ 8) The initial syllables of many of the misaligned words of (33) and (34) do not bear secondary stress; behold God in (33a), for example, is prosodically parallel to divine soul in (28). Moreover, there is nothing systematic about the scansion of the subordi-
Nonlexical Word Stress in the English Iambic Pentameter
45
nated words in (33) and (34); they sometimes occur in prosodically similar configurations but with their strong syllables in S, parallel to (28) (Combs and Suilens 1940):10 (36) a. So my devout fitts come and go away w s ws w s w s w s (‘‘Holy Sonnet: ‘Oh, to vex me, contraryes meet’ ’’ 12) b. As my prophane Love, and as soone forgott: w s w s w s w s w s (‘‘Holy Sonnet: ‘Oh, to vex me, contraryes meet’ ’’ 6) And in any case, the correctly matched syllable does not necessarily follow immediately, but in a few cases only more distantly, as in (37), parallel to Shakespeare’s (15) and (23d):11 (37)
a.
b. Open to’all searchers, unpriz’d, if unknowne. w s w s w s w s w s (‘‘Elegie: Change’’ 6) c. In vaine this sea shall enlarge, or enrough w s w s w s w s w s (‘‘Metempsychosis’’ 52) Thus what (33)–(34) and (37) seem to represent, especially when considered together with (31), is a replacement of (6) by an extension to lexical prosodic structure of some version of (20), in a significant but systematic innovation in the tradition.
46
Kristin Hanson
The result of the extension of (20c) in particular is a signature cadence of Donne, with a misaligned prominent syllable followed by a more prominent correct alignment often having a characteristic dramatic e¤ect of inducing, then resolving, tension. Some lines pose certain complexities in relation to this analysis. Those in (38) suggest that closer attention may be required as to how the rules for phrasing in (7) apply to coordinate structures: (38) a. Corrosivenesse, or intense cold or heat. w s w s w s w s w s (‘‘To Sr. Edward Herbert. At Julyers’’ 26) b. All demands, fees and duties; gamsters, anon w s w s w shi w s ws (‘‘Satyre V’’ 39 ) Intuitively, both are like those in (37) insofar as the entire coordinate structure is a single intonational phrase with its strongest syllable in the final S of its colon. But while (38a) also seems metrical by virtue of the more direct subordination of intense to cold, like the lines in (34), and even perhaps a potential e¤ect of the Rhythm Rule, (38b) di¤ers from the lines in (37) in lacking their regular triple rhythm, instead having a stressed syllable, fees, adjacent to the mismatched one. And those in (39) are metrical by virtue of (20) only if there is also assumed to be some special rhetorical stress on thee, an assumption that is plausible in both cases: (39) a. This kinde of beast, my thought shall except thee; w s w s w s w s w s (‘‘Elegie: The Expostulation’’ 24) b. Fear frownes? And, my Mistresse Truth, betray thee w s w s w s w s w s (‘‘Satyre IV’’ 163) But only three lines seem to fall outside the scope of (20) entirely:12 (40) a. Why should’st thou forget us eternally? w s w s *w s ws w s (‘‘The Lamentations of Jeremy’’ 385) b. Is not this excuse for mere contraries, w s w s *w s w s ws (‘‘Satyre III’’ 98) (41) How fresh our love was in the beginning; w s w s w s w s*w s (‘‘Elegie: His parting from her’’ 77)
Nonlexical Word Stress in the English Iambic Pentameter
47
(40a) comes close to being allowed by (20c); the strong syllable of eternally to which the mismatched one of forget is subordinated is within the same intonational phrase, and creates triple rhythms similar to those in (37). But it cannot be regarded as in the same colon without assuming the underlying structure to be that in (2a), counterintuitively introducing still other mismatches, since the colon boundary would then not match even a word boundary, and its strongest position would contain no stress. In (40b) the subordination of excuse is not within the same intonational phrase on the assumptions sketched in section 3.2.3. Both lines seem like self-conscious transgressions: (40a) in particular is a wrenching climax to the poem’s description of alienation from God just prior to its final plea, Restore us Lord, to thee; and that in (40b) is explicitly about being contrary. Alternatively, of course, they might represent the most extreme version possible of a constraint like (20c), allowing subordination that is limited only by the line. The line in (41), which is formally di¤erent, is also textually uncertain; preceded in its clause only by ‘‘the Spring,’’ it is syntactically problematic as well. We have up to this point avoided tampering with Kiparsky’s (1977) formalization as much as possible, casting the net in (20c) wide rather than struggling to articulate constraints he himself did not identify for Shakespeare, but at this point the issue of complexity that the range of examples from Donne raises bears mention. Intuitively, the lines in (33) that have the correctly aligned stress in the strong position of a colon seem simpler than those in (34) that do not. Those in (33) that have it following immediately also seem simpler than those in (37) that do not. (How those in (37) compare with those in (34) is not clear, suggesting perhaps di¤erent stylistic choices among metrical possibilities of equivalent complexity.) Even among those in (33), those in (33a–g) in which the correctly aligned syllable concludes a constituent at the colon boundary seem simpler than that in (33h) where it does not; similarly, a line like (16b), even if unmetrical, seems simpler than (16a). Hayes (1989b) even suggests that for constructions like those described by (20c), subordination itself might not be strictly necessary so much as just nonfinality of a mismatched syllable (though see also note 15), in which case the assumption of stress on thee in (39) might make those lines simpler, but not be necessary to their metricality. All these considerations suggest that (20c) is a conflation of conditions belonging to a larger set whose members’ identities and interrelationships ought properly to be teased apart and explored.13 As Youmans (1983, 91) notes, identifying such conditions may be more fundamental to metrical description than identifying any single inviolable rule, since they reveal the ‘‘plausible metrical relationships’’ among relaxations poets allow. Even if (20) is imperfectly formulated, however, and the di‰culty of making absolute statements is compounded by the textual, linguistic, and other metrical uncertainties involved in assessing Donne’s meter, (20) seems quite clearly to describe the trajectory along which his relaxation of (6) takes place. Donne’s practice may be
48
Kristin Hanson
extreme within the English tradition, but is not outside it: in it strong syllables of lexical words come to be governed by the same basic kinds of conditions that for Shakespeare govern those of postlexical prosodic structures. 3.4.2
Nonlexical Words
For nonlexical words, however, it is not at all clear that the same conditions obtain. Nonlexical words with initial stress cannot help but escape condition (20b), and so be parallel to the left-aligned lexical words in (31): (42)
f ss sw a. Marke, if she marke whether her woman came. w s w s w s w s w s (‘‘To Mrs. M.H.’’ 44) b. Or, as a Ship after much paine and care, w s w s w s w s w s (‘‘To Mr. Tilman after he had taken orders’’ 9) c. And shee’s unto her selfe a bitternesse. w s w s w s w s w s (‘‘The Lamentations of Jeremy’’ 16) d. So wee, If wee into our selves will turne, w s w s w s w s w s (‘‘To Mr. Rowland Woodward: ‘Like one who’in her third widdowhood’ ’’ 22) e. Why shoulds’t thou that dost not onely approve, w s w s w s w s w s (‘‘Satyre I’’ 37)
But those with final stress not only are positioned at the furthest reaches of what (20c) allows, but also violate it with frequency out of all proportion to the overall number of such words in the language. Prepositions appear with their strong syllable in W whether subordinated to an immediately following stronger stress as in (43a,b), or to a more distant one as in (43c,d), or followed only by an unstressed syllable to which they are not subordinated at all, as in (42e,f ); in the latter neither the meter nor the meaning seems to urge stress on the final pronouns in quite the same way the lines in (39) do: (43) a. Here upon earth, we’are Kings, and none but wee w s w s w s w s w s (‘‘The Anniversarie’’ 23)
Nonlexical Word Stress in the English Iambic Pentameter
49
b. Which still pursues them, without strength they go. w s w s w s w s w s (‘‘The Lamentations of Jeremy’’ 24) c. And against mee all day, his hand doth fight. w s w s w s w s w s (‘‘The Lamentations of Jeremy’’ 180) d. And without such advantage kill me then. w s w s w s w s w s (‘‘The Dampe’’ 16) e. And betweene us all sweetnesse may be had; w s w s w s w s w s (‘‘Sapho to Phil¾nis’’ 43)
f.
Conjunctions as in (44) and pronouns as in (45) also occur with their strong syllables in W whether (20c) is (distantly) met as in (44a,b) and (45a), or not, as in (44c) and (45b): (44) a. Seekes her at Rome, there, because hee doth know w s w s w s w s w s (‘‘Satyre III’’ 45) b. That wilt consort none, untill thou have knowne w s w s w s w s w s (‘‘Satyre I’’ 33) c. Is’t because thou thy self art blind, that wee w s w s w s w s w s (‘‘Elegie: His parting from her’’ 15) (45) a. Which himselfe on the Doctors did bestow; w s w s w s w s w s (‘‘La Corona’’ 46)
50
Kristin Hanson
b. Durst looke for themselves, and themselves retrive, w s w s w s w s w s (‘‘To Sr. Henry Wotton: ‘Sir, more then kisses’ ’’ 44) And most strikingly, certain adverbs occur with their strong syllable in W, again whether admitted under some interpretation of (20c) as in (46a–c), or not, as in (46d): (46) a. That wak’t from both, I againe risen may w s w s ws w sw s (‘‘La Corona’’ 83) b. A better Sun rose before thee to day, w s w s w sw s w s (‘‘Resurrection, imperfect’’ 4) c. What before pleas’d them all, takes but one sense, w sw s w s w s w s (‘‘Farewell to love’’ 18)
d.
Adverbs are commonly classified as lexical in English, but those in (46) are all of the kind traditionally referred to as particles. Syntactically, like all nonlexical words other than prepositions they do not head phrasal projections. Phonologically, in contemporary English at least, their behavior is somewhat inconsistent in its patterning with respect to the lexical/nonlexical distinction: enough in (47b) rejects phrasal stress just as the indisputably nonlexical words in (9) do, while others such as before in (48b) do not, patterning instead with the lexical adverb previously: (47) a. Make sure the suspect is su‰ciently we´ak. b. Make sure the suspect is we´ak enough. (48) a. That’s the excuse she used pre´viously. b. That’s the excuse she used befo´re.
Nonlexical Word Stress in the English Iambic Pentameter
51
But even if this suggests that the distinction between lexical and nonlexical words is gray in some cases, or its consequences not fully understood, these adverbs’ syntactic closed-class status, their relation to clearly nonlexical forms, and most important, their lack of participation in any morphologically complex structures, are consistent with the assumption that nonlexical status is the basis of their special metrical treatment; they are the exceptions that prove the rule. In sum, then, while nonlexical words may conform to (20) or even (6) more often than not, there is little evidence that (20) serves as the genuine constraint on them that it seems to be on lexical words. Even allowing for the possibility that a few lexical words as in (40) may be exceptions to (20), to treat the nonconforming nonlexical words as exceptions of the same type would leave the frequency with which exceptions to (20) involve nonlexical words as simply an unexplained coincidence. And because even among the nonlexical words conforming to (20c) so many involve the most distant placement of the corrective to the mismatch that it permits, this frequency cannot be attributed to the rhythmic configuration these words fall in most often, such as having a following stronger stress as in (43a,b). Rather, they just seem to be freer. 3.4.3
Phrases
This freedom that Donne accords nonlexical words is the same freedom that Kiparsky (1977, 206) notes Donne accords phrases. One of the most important sources of the commonly noted roughness of Donne’s iambic pentameter is that syllables that are strong within phrases fall freely in W, whether the conditions of (20) are met only in some attenuated form as in (49a) or (50a), or not at all as in (49b–e) and (50b): (49) a. Both the yeares, and the dayes deep midnight is. w s w s w s w s w s (‘‘A nocturnall upon S. Lucies day’’ 45)
b.
c. Whither, why, when, or with whom thou wouldst go. w s w s w s w s w s (‘‘Satyre I’’ 64)
52
Kristin Hanson
d. Sooner may one guesse, who shall beare away w s w s w s w s w s (‘‘Satyre I’’ 57) e. If our Soules have stain’d their first white, yet wee ws w s w s w s w s (‘‘To Mr. Rowland Woodward ’’ 13) (50) a. Perfect French, and Italian; I replyed, w s w s ws w s w s (‘‘Satyre I’’ 103) b. Fishes glide, leaving no print where they passe, w s w s w s w s w s (‘‘To Sr. Henry Wotton: ‘Sir, more then kisses’ ’’ 56) And they may do so whether the preceding syllable is in a nonlexical word as in (49), or a di¤erent lexical word as in (50), parallel to (16) and to (17) respectively, which are all unmetrical for Shakespeare. Thus for Donne as for Shakespeare, strong syllables of lexical words are treated one way, while those of nonlexical words and phrases are treated another way. This pattern follows naturally from the distinction between lexical and postlexical prosodic structure. The di¤erence between the two poets is that for Shakespeare, postlexical prosodic structure is subject to a more relaxed constraint than lexical stress is, while for Donne it is not subject to any. 3.4.4
Shakespeare’s Exceptional Lines Reconsidered
Before leaving Donne, one consequence of this analysis bears mention. Starting as it does from the broad description of Shakespeare’s practice in Kiparsky 1975b, 1977, and the poet-by-poet approach to a map of the English metrical tradition advanced there, these generalizations have been presented as whole-cloth analyses of individual poets’ metrical practice. But as briefly mentioned in section 3.3.1, exceptions to Kiparsky’s generalizations have periodically been cited. Some are spurious, but others are genuine, even involving some of the same words discussed above, like enough and divine (Youmans 1983).14 Most interestingly for the pattern under consideration here, lines like that in (51) become frequent in some of the late tragedies: (51) Upon my secure hour thy uncle stole, w s w s w s ws w s (Hamlet 1.5.61) Here the phrase my secure hour has exactly the same prosodic structure of my divine soul in (28), but is positioned metrically in just the way Kiparsky (1977) claims to
Nonlexical Word Stress in the English Iambic Pentameter
53
be impossible (Golston 1998). Across Shakespeare’s practice, the number of such lines may be small enough to be easily overlooked (though see Youmans 1983), but there is no question that within some plays they become a structure to be reckoned with (Hanson 2006). However, prosodically similar prepositions, to take one example of nonlexical words, are scanned as in (23) throughout Shakespeare’s practice, not only in certain periods or plays (Bartlett 1894). Lines such as (51) therefore do not obviate the need for metrical acknowledgment of the distinction between nonlexical and lexical prosodic structure advanced here. What they have consequences for is rather its history and its aesthetic significance. Donne’s lines in (33) may not be so innovative as they have been made out to be, and the patterns discussed here may reflect style and genre as much as author in ways that merit deeper exploration. 3.5
Shelley
If strong syllables of nonlexical words pattern with those of phrases in the practice of Shakespeare and Donne, however, they do not seem to do so in the practice of all poets. In Shelley’s iambic pentameter, certain observations point to the possibility that words may be treated the same way, whether lexical or nonlexical. 3.5.1
Lexical Words
Hayes (1989b, 251–252) observes that Shelley allows lines like those in (52), in which W contains strong syllables of lexical words that are phrased with a following word: (52)
a.
b. The distinct valley and the vacant woods w s w s w s w s w s (‘‘Alastor’’ 195)
54
Kristin Hanson
c. Was raised by intense pensiveness, . . . two eyes, w s ws w s w s ws (‘‘Alastor’’ 489) d. Its stony jaws, the abrupt mountain breaks, w s ws ws w s w s (‘‘Alastor’’ 551) e. She replied earnestly:—‘It shall be mine, ws w s w s w s w s (‘‘The Revolt of Islam’’ 1000) f. And though my cheek might become pale the while, w s w s w s w s w s (‘‘The Revolt of Islam’’ 1011) g. I remained still for tears—sudden she woke ws w s w s w s w s (‘‘The Revolt of Islam’’ 1087) The configurations are similar to those of Donne in (33) and (34), suggesting that on the approach adopted here, Shelley’s constraint on strong syllables of lexical words must include some condition like the right subordination of (20c), although the apparent absence of lines like (37) and (38) suggests there are also certain di¤erences. The syllable to which the mismatched one is subordinated seems always to follow immediately, and to be in the same phonological phrase (Hayes 1989b).15 The colon also seems to be less salient stylistically. Whereas in Donne’s lines the syllable to which the mismatched one is subordinated seems most often to be a monosyllable in the strongest position of the colon as in (33a–g), in Shelley’s the syllable to which the mismatched syllable is subordinated more often initiates a polysyllabic word as in (48a–e). A significant aesthetic di¤erence follows: whereas Donne’s practice foregrounds the underlying structure at the colon boundary after obscuring it at the foot, creating the sense of resolution of tension noted in connection with lines (33) and even (37)–(38), Shelley’s continuation of words across the colon’s boundary keeps its underlying structure in the background and creates more of a sense of insistence on the natural rhythms of speech. Shelley also allows lexical words to be positioned as in (53), analogous to Donne’s practice in (31):
Nonlexical Word Stress in the English Iambic Pentameter
55
(53)
Thus, as for Donne, for Shelley strong syllables of lexical words seem to be constrained by some version of (20), suitably revised to reflect the di¤erences discussed above. 3.5.2
Phrases
Shelley is also like Donne in allowing syllables that are strong in phrases to occur freely in W (Kiparsky 1977, 210; Hayes 1989b, 251): (54)
a.
b. Like a child, half in tenderness and mirth— w s w s w s w s w s (‘‘The Question’’ 14) c. Then was heard—‘‘He who judged let him be brought w s w s ws w s w s (‘‘The Revolt of Islam’’ 1999) d. Thou darest to speak—senseless are the mountains: w s w s w s w s w s hi (‘‘Hellas’’ 475) 3.5.3
Nonlexical Words
But nonlexical words do not seem to show the latter liberty. Those with final stress listed in Ellis 1892 appear with their stressed syllable in W only in configurations like
56
Kristin Hanson
those of the lexical words in (52).16 In most, the mismatched strong syllable is subordinated to an immediately following strong syllable with which the mismatched one is phrased: (55)
a.
b. And, lo! with gentle motion, between banks w s w s w s w s w s (‘‘Alastor’’ 399) c. Orb above orb, with starry shapes between, w s w s w s w s w s (‘‘The Revolt of Islam’’ 592) d. His reign and dwelling beneath nether skies, w s w s w s w s w s (‘‘The Revolt of Islam’’ 395) e. Rose from their thrones, built beyond mortal thought, w s w s w s w s w s (‘‘Adonais’’ 398) f. How without fear of evil or disguise w s w s w s ws w s (‘‘The Revolt of Islam’’ 950) g. For, before Cythna loved it, had my song w sw s ws w s w s (‘‘The Revolt of Islam’’ 928) h. Horsed upon stumbling jades, carted with dung, w s w s w s w s w s (Charles I I.170) i. The sun will rise not until noon. Apollo ws w s w s w s w s hi (‘‘Prometheus Unbound’’ 2.5.10)
Nonlexical Word Stress in the English Iambic Pentameter
57
Only (56a) shows any clear relaxation of the constraint, while (56b) requires the again-not-implausible assumption of stress on thee if it is to conform to it: (56) a. That among such as these my youth should seek its mate. w s w s w s w s w s w s (‘‘The Revolt of Islam’’ 693) b. But beside thee, where still my heart has ever been. w s w s w s w s w s w s (‘‘The Revolt of Islam’’ 18) Apart from these, there are no lines comparable to Donne’s (43e–f ), (44c), (45b), or (46d), in spite of the exact comparability of Shelley’s phrases in (54) to Donne’s in (49b–e). Thus, the word for Shelley appears to encompass both its lexical and nonlexical forms. 3.6
Conclusion
The distinction between lexical and postlexical prosodic structure thus seems to provide a parameter of variation that is independent of, yet interacts systematically with, the variation already known to exist in how poets treat words as opposed to phrases, even when the phrases are composed of sequences of unstressed and stressed syllables of the type that might be found in words. The overall resultant pattern for the tradition for the poets discussed here, setting aside the historical complexities raised in section 3.4.4, and the need for more careful analysis of (20c) acknowledged in section 3.4.1, is thus as in (57). (57) Distribution of strong syllables in W (9 ¼ prohibited except under initial relaxation; 9 ¼ prohibited under conditions like those in (20); r ¼ free) Lexical
Postlexical
Words
Phrases
Shakespeare Donne Shelley The pattern amply supports Halle and Keyser’s (1966) original observation that the stress of function words can be metrically irrelevant. Accompanying as it did their call for a generative reconceptualization of meter itself, Halle and Keyser’s observation swiftly became the locus of a theoretical reaction to their larger project so impassioned that to this day many literary critics and linguists view themselves as separated from each other by an unbridgable chasm. For
58
Kristin Hanson
example, Magnuson and Ryder (1970, 1971), questioning the assumption that any systematic description of a meter could successfully characterize all lines in a verse corpus, noted that Halle and Keyser’s claim about function words left unexplained Magnuson and Ryder’s judgment that Donne’s (58a) is patently more metrical than their construct in (58b): (58) a.
Behind, before, above, between, below. (Donne, ‘‘Elegie: Going to Bed ’’ 26) b. *Inbetween, before, beneath and beyond
On the analysis presented here, this judgment is reflected in the facts that (58b) is characterized as unmetrical for either Shakespeare or Shelley, and that in consequence, although it is characterized as metrical for Donne, it is also characterized as at the limit of his practice as defined by the relationship of his innovations to the tradition he belongs to. Perhaps more profoundly, Wimsatt (1970, 785), rejecting the assumption that iambic pentameter could lie elsewhere than in an iambic rhythm in the language of a poem, insisted that function words alter their stress under pressure from the meter when the two conflict as in (59): (59) Ther was also a Reve, and a Millere, (Chaucer, The Canterbury Tales, General Prologue 542) Without accepting Wimsatt’s claim that pronunciation actually needs to be changed, his sense of there being less fixity to the stress of such function words can be understood here as a sense of the obligatoriness of the meter’s respect for lexical as opposed to nonlexical stress. But what is more important, perhaps, is that attributing a role in metricality to such an abstract organizational principle of grammar as the distinction between lexical and postlexical components of the phonology need not be taken to cast the experience of meter as an abstract intellectual one rather than as the sensory one critics have traditionally emphasized (Wimsatt and Beardsley 1959). Metrical generalizations drawing on generative linguistics in this way are embedded in a broader assumption that the cognitive faculty that apprehends linguistic structure is as instinctive in its operation as sight, hearing, taste, or touch (Fodor 1983). In this light, the di¤erence that the debate over stress in function words seemed to crystallize can be seen to be spurious, with a sense of the immediacy of metrical experience in fact being shared by literary critics and generative linguists alike. Notes This chapter originated in a course on metrics that Paul Kiparsky taught in my third year as a graduate student, before which I had no intention of pursuing anything remotely related to
Nonlexical Word Stress in the English Iambic Pentameter
59
metrics, and since which it has been at the heart of my working life. I would therefore like to take this opportunity to thank him for that as well as for his comments on earlier versions of the chapter itself. I am also grateful for the suggestions and help of Katy Breen, Tom Cable, Draga Zec, two anonymous reviewers, and especially Sharon Inkelas. 1. Further details of the structures in (2) reflect the assumption that a foot that is not part of a binary grouping is adjoined to one in such a way as to preserve the generalization that strong constituents are heavy, details justified for Romance decasyllabic meters (Piera 1981), but whose correctness for English remains open. In this respect, these structures di¤er somewhat from those proposed by Kiparsky (1977, 230), but nothing in his description of the phenomena at stake here depends on his assumptions about the internal organization of the two cola. The issue is of special interest, however, in light of the importance of colon boundaries suggested below. 2. Specifically, when coupled with the claim advanced below that Donne’s meter treats postlexical prosodic structure as potentially irrelevant, the claim that the limit on the amount of language in any one metrical position is a moraic trochee has the consequence that multiple nonlexical words are freely allowed in single positions, in particular without regard to their weight, a generalization that appears correct for Donne (Hanson 1996). 3. These structures are admittedly rather hairy. Other forms of representation of prosodic structure such as grids are simpler in ways that have been argued to be not only welcome but correct, eliminating representation of levels of constituency and degrees of prominence claimed to have no linguistic relevance, most notably those arising from adjunction. However, the metrical generalizations of Kiparsky (1977) discussed here require reference to some of these levels of constituency (Hammond 1991). Where these have been argued not to be necessary for meter any more than they are for language, the argument has entailed an additional argument that the correct metrical generalizations themselves are in fact rather di¤erent from what Kiparsky (1977) proposes (Hayes 1983, 1989b), an issue I return to briefly in section 3.4. (See also notes 6 and 15.) Note also that while representations using bracketed grids (Halle and Vergnaud 1987) can resolve many of these di¤erences, issues involving prominence in meters other than the English iambic pentameter remain to be worked out. (See Hanson and Kiparsky 1996.) 4. Note that this class of syllables also excludes the strongest syllables of many compounds, because compounds consist of two separate phonological words (Kiparsky 1977, 191–193, though see also Youmans 1983). It should be noted, however, that if compounds were to be analyzed as phonological words themselves, some revision to (6) might be necessary. 5. An exception is object pronouns immediately following their verb, which Inkelas and Zec (1993) therefore classify as true clitics, and assume to be adjoined within their host’s phonological word, as in (13c). 6. In many of these cases Hayes’s (1989) clitic group, a constituent not used in Inkelas and Zec’s (1993) model, would be the constituent within which the relevant strength would be defined. Several issues would need to be resolved in order for the clitic group to be used in restating Kiparsky’s (1977) generalization, however. An explicit account of the distribution of stress on function words themselves would be required, addressing Inkelas and Zec’s (1993) distinction between function words in general and true clitics mentioned in note 5. More important, the question analogous to the one raised in note 3 would need to be answered, namely, whether there is any internal bracketing within clitic groups, as Kiparsky’s approach requires.
60
Kristin Hanson
7. An additional condition not discussed here must also be met: the two syllables preceding the mismatched strong one must themselves be either function words or part of the same word as the mismatched one (Kiparsky 1977; Hayes 1983). 8. The remarks in note 3 regarding grids are thus relevant here as well. See also note 15 regarding the alternative approach suggested by Hayes (1989b). 9. On Hayes’s (1989b) slightly di¤erent assumptions about phrasing, the structures of specific lines would be a little di¤erent, but not the overall pattern. 10. Another example might be the following, but it has an alternative scansion:
(
(i) Else, being alike pure, wee should neither see, w s ws w s w s w s w sw s w s w s w s (‘‘To the Countesse of Huntingdon: ‘Man to Gods image’ ’’ 33) 11. The internal prosodic constituency of words like surrounded is debatable, given that respecting morphological structure as shown here competes with respecting the English preference for left-headed feet (Liberman and Prince 1977); for discussion of parallel cases in Wyatt see Kiparsky 1977. It should also be noted that (37a) opens a poem famous for its games of rhythmic ambiguity, because it mixes tetrameter and pentameter lines. 12. Whether there are any others depends on what assumptions are made about certain other contentious issues. For example, the following line can be scanned on the assumption of intial catalexis: (i) j Making them confesse not only mortall w s w s w s w s w s hi (‘‘Satyre IV’’ 201) 13. Hayes’s (1989) exploration of the metrical consequences of the hierarchical structure of the prosodic hierarchy has this character. So, of course, do current theoretical claims that grammars are rankings of discrete constraints (Prince and Smolensky 1993). 14. For example, Golston’s (1998, 759) list of counterexamples to Kiparsky’s generalizations neglects Kiparsky’s own discussion of many of the exact same lines as involving changes in pronounciation (Kiparsky 1975b, 246), or the operation of the Rhythm Rule (Kiparsky 1975b, 247), or even the historical process by which a poet enters a tradition in which certain practices prevail but winds up excising them from his own (Kiparsky 1975b, 241). 15. On Hayes’s (1989b) own description, Shelley disallows a peak within a word at the right edge of a phrase. This rule, along with relevant observations about Shakespeare’s pattern discussed in section 3.3.2, is formulated within a proposed typology of metrical rules that reflects the general principle that beginnings may be lax while endings are strict, and therefore includes rules that impose special strictness at right edges, with variation among poets arising from the level of the prosodic hierarchy at which such strictness is imposed. The approach di¤ers from that of Kiparsky (1977) in so many interesting and interacting ways that it would require a separate paper to do the comparison justice. As far as the core point of this chapter goes, however, in spite of there being some genuine (though marginal) empirical di¤erences between the two approaches, as well as some di¤erence in aesthetic perspective, the di¤erences do not seem to a¤ect the need described here for distinct behavior of lexical and nonlexical words to be accounted for, nor it is obvious that there would be any impediment to incorporating an
Nonlexical Word Stress in the English Iambic Pentameter
61
account like that o¤ered here within Hayes’s approach. A brief discussion of the issue was excised from this chapter out of considerations of length. 16. Only one di‰cult example presents itself, but the problem involves excess syllables as much as the stress of around; possibly Pile a- occupies, anomalously, a single position. See Gerber 2001 for discussion of such structures in Stevens.
(i) Pile around it, ice and rock; broad vales between w s w s w s w s w s (‘‘Mont Blanc’’ 63)
4
The Prosodic Word as a Unit in Poetic Meter
Draga Zec
4.1
Introduction
The epic decasyllable is a traditional South Slavic folk metrical form. The poems created in this rich oral tradition are of a highly formulaic nature, and possess a discernible metrical organization. My focus will be on the so-called Serbo-Croatian poems within this tradition (see Maretic´ 1907, 1935; Jakobson 1933, 1952), which have been documented in numerous volumes of verse. Among the most notable are the five volumes of Serbian epic poetry collected and published by Vuk Karadzˇic´ between 1814 and 1847, which will be the principal basis of this study.1 Although extensively documented, this verse form has eluded clear characterization. Each line of the epic decasyllable has ten metrical positions, with a caesura after the fourth, which divides it into two cola. A metrical position corresponds, strictly, to a syllable, while colon boundaries coincide with word breaks.2 (1) Mje``se¯c kaˆra¯ k zvije`zdu Da`nicu sssskssssss ‘The moon admonishes the morning star’
Vk, II, 98, 1
These are the obligatory, and just about the only uncontroversial, aspects of the structural organization of this verse. The establishment of further structural characteristics of a decasyllabic line has been fairly intractable. Claims that this is a periodic meter have not been supported by a straightforward prosodic characterization of the division of a line into metrical feet. Yet, any questioning of the periodic nature of the epic decasyllable has left unaccounted for the highly suggestive signs of its trochaic organization. My claim will be that the epic decasyllable is indeed a periodic, more specifically, a trochaic meter. The goal of this chapter is to identify the prosodic indicators of its trochaic organization, which, as will be shown, are of a somewhat unorthodox nature, and crucially include the prosodic word. The theoretical frame of reference will be Hanson and Kiparsky’s (1996) paramaterized account of possible metrical forms,
64
Draga Zec
and the basic premise therein about the dependence of poetic meter on the sound system of language, which sets the range of what is possible, or to be expected, in verse (Jakobson 1960; Kiparsky 1973g; Hanson 1992). Finally, I will rely on Optimality Theory (McCarthy and Prince 1993b; Prince and Smolensky 1993) in presenting the interactions among metrical properties, formalized as constraints, that characterize this metrical form. The organization of the chapter is as follows: In section 4.1, I present aspects of the Serbian prosodic system relevant for understanding this verse. In section 4.2, I establish the structural properties of the epic decasyllable, and in section 4.3, propose a formal account of its metrical organization. In section 4.4 I discuss earlier treatments of this verse. Sections 4.5 and 4.6 address the place of function words in this verse, and section 4.7 revisits the organization of higher order metrical constituency. Concluding remarks are given in section 4.8. 4.1.1
The Prosodic System
This section provides an outline of aspects of the Serbian prosodic system whose understanding is crucial for providing proper characterization of the epic decasyllable. I present the relevant prosodic characteristics of two major dialectal areas, Old and Neo-Sˇtokavian. The prosodic distinctions between these two dialect groups will shed crucial light on the prosodic basis of this verse form, as detailed in section 4.2. I focus first on the distribution of pitch accent in the Old and Neo-Sˇtokavian dialects, then on its prosodic nature, and the minimal prosodic unit within which it is realized. The Neo-Sˇtokavian dialects are traditionally characterized as having four pitch accents, two falling and two rising (Ivic´ 1958; Lehiste and Ivic´ 1986). As illustrated in (2), the falling accents may occur on the only syllable of a monosyllabic word, and on the initial syllable of a polysyllabic one. The rising accents, which are not found in monosyllables, may appear on any syllable of a polysyllabic word, with the exception of the final one. (2) Neo-Sˇtokavian pitch accents Short falling `` n a. Monosyllabic sa ‘dream’ b. Disyllabic me``se¯c ‘moon’ `` buka c. Trisyllabic ja ‘apple’
Long falling daˆn ‘day’ suˆnce ‘sun’ naˆmera ‘intention’
Short rising None
Long rising None
ju`na¯k ‘hero’ da`nica ‘morning star’ planı`na ‘mountain’
lju´ba ‘wife’ prı´lika ‘opportunity’ juna´ci ‘hero, plural’
The Prosodic Word as a Unit in Poetic Meter
65
The prosodic properties of the Old Sˇtokavian dialects are more straightforward: these are the dialects with only the falling accents, which may appear on any syllable within the word, including the final one, as in (3): (3) Old Sˇtokavian pitch accent Short falling `` n a. Monosyllabic sa ‘dream’ b. Disyllabic me``se¯c ‘moon’ `` lju¯ba ‘wife’ `` buka c. Trisyllabic ja ‘apple’ ` danı`ca ‘morning star’
Long falling daˆn ‘day’ suˆnce ‘sun’ junaˆk ‘hero’ naˆmera ‘intention’
The traditional diacritics utilized in (2) and (3) cover a complex of prosodic properties. Each symbol marks the presence of stress, as well as a characteristic pitch, either falling or rising. What the traditional diacritics do not mark is the disyllabic nature of the Neo-Sˇtokavian rising accents, with a high pitch characterizing both the accented syllable and its immediate successor. Moreover, in both dialect groups, the long accents appear on syllables with long vowels, while the short accents appear on syllables whose vowel is short. Vowel length, which is an independent phonological property, is also found on syllables unassociated with pitch accent—for example, on the second syllables of me``se¯c ‘moon’ and ju`na¯k ‘hero’. The Sˇtokavian system of pitch accents is of a complex nature: each pitch accent is manifested as both tone and stress. Although distinct, these two components of pitch accent are integrated through the system of feet, which includes both the standard trochaic set, in (4a), and the set characterized by tonal prominence, in (4b), as argued in detail in Zec 1999. (4) Foot inventory a. Without tone i. [sm m ]F ii. [sm sm ]F
b. With tone i. [sm m ]F H ii. [sm ]F H
Pitch accent is realized within the prosodic word. Each prosodic word obligatorily contains a tonal foot that serves as its head and is the bearer of stress, as shown by the forms parsed into feet in (5) and (6). This foot (designated by underlining) is the
66
Draga Zec
leftmost one in the prosodic word in the Neo-Sˇtokavian, but not necessarily in the Old Sˇtokavian.3 (5) Neo-Sˇtokavian `` n)]po a. [(sa H d. [(lju´ )(ba)]po H H
b. [(daˆn)]po H e. [(da` )(ni)ca]po H H
(6) Old Sˇtokavian `` n)]po a. [(sa H `` )]po d. [(lju¯)(ba H
b. [(daˆn)]po H e. [da(nı``)ca]po H
c. [(me`` )(se¯c)]po H f. [( ju` )(na¯k)]po H H c. [(me`` )(se¯c)]po H f. [ ju(naˆk)]po H
Because the stressed syllable has to be associated with tone, ‘‘standard’’ trochaic feet in (4a) may never serve as heads of prosodic words. However, they may appear in nonhead position: (4ai) appears in nonhead position in (5c) and (6c–d). Nonhead tonal feet are relevant only for the Neo-Sˇtokavian, and are exemplified in (5d–f ). Both ‘‘standard’’ trochaic feet and nonhead tonal feet are active in prosodic morphology, as argued in Zec 1999. The head of the prosodic word is associated with two types of prominence: stress, which is relational, and tonal salience, which is not. The head foot bears the highest degree of prominence by virtue of stress, which is phonetically realized as increase in duration; in (5d–f ) stress alone makes the head foot more prominent than the immediately following nonhead tonal foot. The head foot is also characterized by tonal salience, which distinguishes between feet associated or not associated with a High tone. Tonal salience, as already noted, is the nonrelational sort of prominence, its phonological status being predicated solely on the presence or absence of a High tone. As already noted, pitch accent is realized within the prosodic word. Given the division of the lexicon into lexical and functional elements, only elements in the former class invariably correspond to prosodic words. As argued in Zec 2005, the prosodic properties of functional elements are more varied, their prosodic behavior depending on whether they belong to the class of clitics—that is, bound function words—or free, nonclitic function words. As will be shown in section 4.6, this distinction plays an important role in the organization of the epic decasyllable. While lexical words (WL ) are unexceptionally associated with pitch accent and prosodic word status, as stated in (7), the prosodic status of free, nonclitic function words (WF ) depends on their size. Those that are minimally disyllabic possess a pitch accent and receive the status of a prosodic word. Those, however, that correspond to
The Prosodic Word as a Unit in Poetic Meter
67
a single syllable are accented and endowed with the prosodic word status only if associated with sentential focus, as stated in (8). (7) If WL , then [WL ]po . (8) If WF , then [WF ]po i¤ WF is minimally disyllabic, or associated with focus. In sum, monosyllabic free function words lack the prosodic status and, as such, disrupt the prosodic hierarchy (as defined in Selkirk 1980, Nespor and Vogel 1986, and Hayes 1989b). As argued in Zec 2005, such monosyllabic elements are incorporated directly into the prosodic phrase. The organization of the epic decasyllable is highly sensitive to the prosodic properties of monosyllabic function words: their distribution within a metrical line is considerably restricted in those cases when they lack prosodic salience. Bound function words—those designated as clitics—cannot form a prosodic word on their own under any circumstances. Rather, they subcategorize for a prosodic word, either as proclitics, as in (9a), or as enclitics, as in (9b) (Inkelas 1989); the clitic forms a nested prosodic structure with its host. (9) a. [clitic [host]po ]po ] b. [[host]po clitic]po Both lexical and functional elements may serve as clitic hosts. Moreover, monosyllabic free function words, which are also possible clitic hosts, form with the clitic a disyllabic unit with nested prosodic word structure, which then makes them eligible for pitch accent (Zec 2005). The clitic, in this case, endows its host with the prosodic word status (see Halpern 1992 for other cases of this type). A monosyllabic free function word thus corresponds to a prosodic word in yet another case: not only when the monosyllabic function word is focused, as stated in (8), but also when it is associated with a clitic. At most one pitch accent is associated with a prosodic word, regardless of whether it is simplex in structure, as in (10a), or possesses a nested structure, as in (10b–c): (10) a. [daˆn]po ‘Day’ b. [na [daˆn]po ]po ‘On (the) day’ c. [[daˆn]po je]po ‘(The) day is’ The nested structure created by the clitic and its host is ambiguous in one respect: phonological processes whose domain is the prosodic word may select either the larger or the smaller scansion of the nested structure and, as argued in Zec 1993,
68
Draga Zec
individual dialects, or local idioms, select either the larger or the smaller scansion. However, the larger scansion is relevant for the purposes of prosodic parsing, since it alone yields an exhaustively parsed prosodic constituency. In the epic decasyllable, the larger scansion is selected in the general case, as demonstrated in section 4.6. I take this to be a fact about the linguistic, rather than the metrical, makeup of this verse. 4.2
Prosodic Basis of Metrical Organization
The arguments, both for and against, the periodic organization of the epic decasyllable have relied on the distribution of pitch accent. 4.2.1
Distribution of Pitch Accent
As observed in the literature (Zima 1909a, 1909b; Maretic´ 1907; Jakobson 1933, 1952), the occurrence of pitch accents in this verse appears to be fairly free, with one notable regularity: colon and line-final syllables, the fourth and the tenth, may not bear pitch accent. These distributional properties have been stated with the Neo-Sˇtokavian prosodic system in view, and are illustrated in (11); the accent marks follow the Neo-Sˇtokavian accentuation. ` lo`vi k cı`jel ` daˆn po go`ri, (11) a. Dmı`tar `ˇta ulo`viti: b. i ne` mozˇe k nı`s `` nese c. naˆmjera ga k pred ve``cˇe na d. na ze`leno¯ k u go`ri je``zero, e. u je``zeru k `u`tva zlato`krila; ` k sıˆvo¯ga so `` kola `` sti Dmı`tar f. pu g da u`vati k `u`tvu zlato`krilu, Vk, II, 97, 64–70
‘Dmitar was hunting, all day, in the mountain, but could not catch anything. A purpose brought him, toward the evening, to a green mountain lake. In the lake, there was a gold-winged duck. Dmitar released his gray falcon, to seize the gold-winged duck.’
While the fourth and tenth positions consistently remain unaccented, all other positions in a line, both odd and even, may bear accent: the first syllable is accented in lines (a), (c), and (f ); the second in (b), (d), (e), and (g); the third in (a) and (f ); the fifth in all lines other than (c) and (d); the sixth in lines (c) and (d); the seventh only in line (a); the eighth in all lines other than (a); and the ninth in line (a).4 In light of this, any assumption that the epic decasyllable is a periodic meter opens the issue of the headedness of metrical feet. The problem is illustrated in (12): while in (12a) (repeated from (11a)) each accent is associated with an odd position in the line, in (12d) (repeated from (11d)), each is associated with an even position. (Parentheses
The Prosodic Word as a Unit in Poetic Meter
69
in (12) designate the division into metrical feet; accented syllables are highlighted by underlining.) ` ` (12) a. (Dmı`tar) (lo`vi) k (cı`jel) (daˆn po) (go`ri), b. (na ze` )(leno¯) k (u go` )(ri je`` )(zero),
Cf. (11a) Cf. (11d)
The issue of whether the distribution of pitch accents is metrically significant for the organization of this verse led in fact to divergent views on the nature of this meter. Under one view held, most notably, by Jakobson (1933, 1952), pitch accent exhibits a trochaic tendency, manifested as an overall statistical bias toward accentedness in the odd positions of the meter, and its absence in the even positions.5 In the data he presents, based on two poems (783 lines) by the Montenegrin singer T. Vucˇic´, given in (13), the percentage of accented syllables is generally lower in the even than in the odd positions of the line.6 (13) The distribution of pitch accents in Vucˇic´’s poems (783 lines) (Jakobson 1952, 420) 1
2
3
4
5
6
7
8
9
10
62
31
53
—
75
22
34
35
42
—
Thus, according to Jakobson (1952, 420), ‘‘The even syllables of the line show a marked tendency to be unaccented and the odd syllables a less marked tendency to carry word accents: even syllables are less frequently accented than the following odd syllables.’’ But, with the exception of the fourth and tenth positions, the degree of accentedness in the even positions makes the trochaic tendency fairly weak. In (13), the eighth position has a slight advantage over the preceding seventh position, which according to Jakobson does not lessen the trochaic tendency. But the data from Vuk’s collection show that in certain poems, the eighth position exceeds not only the seventh, but also the ninth position in the line.7 In sum, if the epic decasyllable is indeed a trochaic meter, one would expect much more robust prosodic indicators than is the case with pitch accent. Under the alternative view, this meter is seen as syllabic in nature, and any apparent trochaic tendency is attributed to the phonological properties of the language rather than to the pressures of the meter (Zima 1909a, 1909b; Maretic´ 1907; Vaillant 1932; Matic´ 1964). This view, however, is overly pessimistic in assuming that no prosodic property other than syllable count is subject to metrical patterning. The one robust prosodic regularity that emerges is the absence of pitch accent stands for from the fourth and tenth positions in the line, as stated in (14) (where s a syllable bearing any of the four pitch accents): kssssss (14) a. *s s s s b. *s s s s k s s s s s s
70
Draga Zec
As we will see, the regularity in (14) does reflect the trochaic organization of the epic decasyllable, once the prosodic indicators of its metrical organization are properly identified. 4.2.2
Accents of Mono- and Polysyllables
A clearer metrical characterization of this verse emerges if we narrow the focus on the distribution of the accents of monosyllables, which, as will be argued, play a central role in the trochaic patterning of this verse. Relevant here is the distribution of monosyllables that constitute prosodic words and, as such, bear pitch accent.8 It is important to note that within the bounds of the Neo-Sˇtokavian accentuation, the restriction in (14) is metrically relevant for monosyllabic, but not for polysyllabic prosodic words. Polysyllabic words, whose final syllable may not bear pitch accent in this dialect, as noted in section 4.1, vacuously meet this restriction by virtue of their prosodic makeup. Thus, only the absence of monosyllables from the colon or linefinal syllable has to be viewed as metrically motivated since, by occupying these positions, monosyllables would have endowed them with pitch accent. But what role, if any, is to be ascribed to the accents of polysyllables? While the Neo-Sˇtokavian dialect, whose polysyllables may not bear final accent, provides no useful clues, the Old Sˇtokavian does since, in this dialect, the accent of a polysyllabic word may occur on any syllable, including the final one. Crucially, in the Old Sˇtokavian rendition of the epic decasyllable, accents of polysyllabic words do occur in colon and line-final metrical positions. Evidence for this comes from three Old Sˇtokavian sources of this verse: the Montenegran cycle in the fourth volume of Vuk’s collection, Njegosˇ’s Ogledalo srpsko, and the poems collected in Parry and Lord 1953.9 In (15)–(17), accents in the fourth position are given in the (a) examples, and accents in the tenth position, in the (b) examples. Accents on the fourth syllable occur in about 8 percent of lines, and on the line-final syllable, in about 4 percent.10 (15) Vuk’s collection `` n k ne obra `` c´a gla¯veˆ a. Za toˆ Iva ‘To that, Ivan did not turn his head’ `` k k svojoˆj go¯spı`` kaˆzˇe¯ Onıˆ sa¯na ‘She told that dream to her lady’ `` m rukoˆm k `u`fati oruˆzˇe Jedno ‘With one arm, he seized the gun’ `` du bje¯zˇeˆ b. Svi `o`vcˇa¯ri k uz liva ‘All shepherds ran up along the meadow’ `` nj da¯jeˆ Pa¯ `o`bjema k zˇ ˆıvı¯ oga ‘And with both, he produced live fire’ `` n k ne obra`` c´a gla¯veˆ Za toˆ Iva ‘To that, Ivan did not turn his head’
Vk, IV, 2, 91 Vk, IV, 12, 10 Vk, IV, 12, 55 Vk, IV, 12, 57 Vk, IV, 2, 119 Vk, IV, 2, 91
The Prosodic Word as a Unit in Poetic Meter
(16) Njegosˇ’s Ogledalo srpsko `ı je oraˆj k polomı`o ` krıˆla a. N`t ‘Nor did the eagle break his wing’ `` gle¯ kliku `` je¯ Ali junaˆk k iz ma ‘But the brave man called from the mist’ `` tı¯n dru¯gaˆh male``n indaˆt k `o`sam sto ‘A small squadron, of eight hundred companions’ `` c´e do `` bra¯ k ispod se``be djo¯ga `` b. da ‘He will give the good white horse under him’ `` tı¯n dru¯gaˆh male``n indaˆt k `o`sam sto ‘A small squadron, of eight hundred companions’ `` c´e za me¯ k se``dam dzˇeferdaraˆh da ‘He will give seven rifles for me’ (17) Parry and Lord’s collection11 `` rı¯ serda´ru a. Paˆ ovakoˆ k govo ‘And spoke thus to the sirdar’ Evo serdaˆr k do druzˇineˆ prı´dje ‘Then the sirdar approached the band’ `ˇta ne govo `` rı¯ `` kı¯ sˇu¯tıˆ k nı`s Sva ‘Each was silent and said no word’ ` ¯ da vo¯jskeˆ Spreˆmı¯ sultaˆn k stoˆ hı`lja ‘The sultan sent an army of a hundred thousand men’ `` zbolova¯ k se``dam go `` din’ da¯naˆ b. Ra ‘He was ill for seven years’ `` nom k devedeseˆt dru¯gaˆ Za Osma ‘Behind Osman were the ninety comrades’ ` ¯ da vo¯jskeˆ Spreˆmı¯ sultaˆn k stoˆ hı`lja ‘The sultan sent an army of a hundred thousand men’
71
Nj, 2, 7 Nj, 54, 50 Nj, 54, 86 Nj, 2, 30 Nj, 54, 86 Nj, 2, 26
P&L, 31, 86 P&L, 31, 56 P&L, 1, 43 P&L, 1, 14 P&L, 26, 2 P&L, 31, 17 P&L, 1, 14
Under the plausible assumption that the epic poems composed in the Old and NeoSˇtokavian dialects not only belong to the same poetic tradition, but are also in the same poetic meter, we conclude that the absence of the accents of polysyllables from colon-final positions in the Neo-Sˇtokavian dialect, while being a fact about its phonology, is not a fact about this verse form (contra Matic´ 1964). The accents of monosyllables, however, may not appear in colon- and line-final position in the Old-Sˇtokavian rendition of the epic decasyllable, just as they may not appear in its Neo-Sˇtokavian version. I will take this to be a strong indicator of trochaic patterning, and focus, in the following section, on the distribution of monosyllabic prosodic words in this verse.
72
4.2.3
Draga Zec
Distribution of Monosyllables
As noted in Maretic´ 1907, 53–54, the fourth and tenth positions in the line are not the only metrical positions from which monosyllables are excluded. Monosyllables are also prohibited in the eighth position. However, this metrical position is relatively frequently associated with the accents of polysyllables in both the Old and NeoSˇtokavian renditions of the epic decasyllable. Thus, while the Neo-Sˇtokavian, but not Old Sˇtokavian, may be seen as correlating the absence of the accents of monoand polysyllables in the fourth and tenth positions, the eighth metrical position does not exhibit any such correlation in either dialect group. The exclusion of monosyllabic prosodic words from the fourth, eighth, and tenth positions emerges as a central trait of this meter. Monosyllables thus may occupy a highly restricted set of metrical positions, as shown in (18)–(23). In addition to lexical monosyllables, which include numerals (the most frequent monosyllabic prosodic words in this meter), examples also include focused functional monosyllables. As is generally the case in this highly formulaic verse, specific monosyllables are often restricted to fixed metrical positions, or recur in fixed syntactic collocations.12 In the first colon, monosyllables may occur in the first syllable, as in (18), or the second one, as in (19). For the sake of explicitness, the entire colon is parsed into prosodic words. The prosodic word in the relevant metrical position, as well as its translation, is underlined. (18) First colon: First position [s]po s s s k s s s s s s a. [Grad]po [gradili]po k tri godine dana, ‘They were building the city for three years’ b. [konj]po [do [konja]po ]po k junak do junaka, ‘Horse next to horse, knight next to knight’ c. [tri]po [tovara]po k blaga zadobisˇe ‘Three loads of treasure, they acquired’ d. [svi]po mi [[da]po se]po k u so prometnemo, ‘If we all turned to salt’ (19) First colon: Second position s [s]po s s k s s s s s s a. Vec´ [daj]po [mene]po k tvoju vjeru tvrdu ‘But give me your firm word of honor’ b. Sˇto [car]po [reko]po k to i ucˇinio ‘What the czar said, he made good’ c. Jer [tri]po [fata]po k u visinu skacˇe ‘As he leaps three yards in height’
Vk, II, 25, 7 Vk, II, 49, 22iv Vk, III, 3, 97 Vk, II, 49, 10iv
Vk, II, 63, 45 Vk, III, 12, 167 Vk, III, 60, 26
The Prosodic Word as a Unit in Poetic Meter
73
Under a limited set of circumstances, monosyllabic prosodic words may also occupy the third position: when immediately followed by a functional monosyllable, which has no prosodic status (as detailed in section 4.1): (20) First colon: Third position s s [s]po s k s s s s s s a. [[da]po se]po [ja]po sad k na vojsku opremam ‘That I am now making myself ready for the army’ b. [i [tu]]po [ja]po bih k u boga zlocˇesta ‘There as well I was evil before God’ c. [[Da]po c´esˇ ]po [ti]po nas k kadgodj izbaviti ‘That you will one day rescue us’
Vk, II, 90, 21 Vk, II, 4, 85 Vk, III, 51, 23
In the second colon, a monosyllable may occupy the first syllable, as in (21); the second, as in (22); and the third, as in (23); the entire colon is, again, parsed into prosodic words. (21) Second colon: Fifth position s s s s k [s]po s s s s s a. Ja njoj dado k [struk]po [rumene]po [ruzˇe]po ‘I gave to her a stem of red rose’ b. posˇljite mu k [list]po [knjige]po [bijele]po ‘Send to him a leaf of white book’ c. Grad gradili k [tri]po [godine]po [dana]po ‘They were building the city for three years’ d. pa po svili k [sav]po [izvezen]po [zlatom]po ‘And on silk, embroidered all in gold’ (22) Second colon: Sixth position s s s s k s [s]po s s s s a. ‘‘Sjedi, care k pak [pij]po [rujno]po [vino]po ’’ ‘Sit here, czar, and drink of red wine’ b. i stolove k dje [car]po [sedijasˇe]po ‘And the tables where the czar had sat’ c. bolje nam je k [sva]po [tri]po [poginuti]po ‘It would be better if we all three died’ (23) Second colon: Seventh syllable s s s s k s s [s]po s s s a. Lov lovio k [letnji]po [dan]po [do [podne]po ]po ‘He was a-hunting on a summer day, until noon’ b. ja sam noc´as k [zao]po [san]po [usnio]po ‘Last night, I dreamt an evil dream’
Vk, II, 29, 284 Vk, II, 28, 121 Vk, II, 25, 7 Vk, III, 60, 4
Vk, II, 28, 479 Vk, II, 94, 108 Vk, II, 38, 55
Vk, II, 14, 20 Vk, II, 9, 77
74
Draga Zec
c. kovala mu k [sablju]po [dva]po [kovacˇa]po ‘His sabre was forged by two blacksmiths’ d. Neka sudi k [kao]po [ja]po sˇto [sudim]po ‘May he adjudicate, just as I adjudicate’
Vk, II, 43, 651 Vk, III, 12, l51
In contrast to the third position, however, the ninth position cannot be occupied by a monosyllabic prosodic word. This is because the prosodic configuration that allows for the occurrence of a monosyllabic prosodic word in the third metrical position, as shown in (20), may not be replicated for the ninth position. This will be attributed, in section 4.6, to a special status of the line-final foot, one of its manifestations being a prohibition against functional monosyllables in the tenth position of the meter. To conclude, monosyllabic prosodic words could in principle occupy any odd position in the line, including the third and the ninth; and, among the even positions, only the second and sixth. However, conspicuously absent are the following arrangements of monosyllabic prosodic words: (24) a. *s s s [s]po k s s s s s s [gradili]po [grad]po k tri godine dana (Constructed on the basis of Vk, II, 25, 7) b. *s s s s k s s s [s]po s s Neka sudi k [kao]po sˇto [ja]po [sudim]po (Constructed on the basis of Vk, III, 12, l51) c. *s s s s k s s s s s [s]po ja sam noc´as k [usnio]po [zao]po [san]po (Constructed on the basis of Vk, II, 9, 77) Thus, what remains to be explained is the absence of monosyllables in the fourth, eighth, and tenth positions. Their absence from the ninth position results from strong distributional restrictions associated with the final foot, to be addressed in section 4.7. It should be noted that the distribution of monosyllables in the epic decasyllable sharply contrasts with the ways monosyllables pattern in poetic traditions such as English or Russian, in which they freely occupy weak metrical positions. 4.3
Organization of the Epic Decasyllable into Metrical Feet
The relevant descriptive generalization about the epic decasyllable—that a monosyllabic prosodic word may not occupy the fourth, eighth, and tenth position in the line—clearly indicates a binary pattern of the decasyllabic verse. I will further argue that it also indicates its trochaic nature. With a line organized into five trochaic feet, two in the first and three in the second colon, a monosyllabic prosodic word is excluded from the weak position of the foot. This is represented in (25), with feet built
The Prosodic Word as a Unit in Poetic Meter
75
over syllable-sized metrical positions, and the ‘‘strong’’ (s) and ‘‘weak’’ (w) marking indicating the head/nonhead relation within a foot (Halle and Keyser 1971a; Kiparsky 1977; Prince 1989). (25) *
s w s w s w s w s w (s [s]o ) (s [s]o ) k (s [s]o ) (s [s]o ) (s [s]o )
This restriction is relaxed in colon-initial feet, the first and the third, those commonly associated with the weakening of restrictions imposed by poetic meter (Kiparsky 1977; Hanson 1992). This prohibition thus takes proper e¤ect in the second, fourth, and fifth foot. But if (25) captures the structural organization of the epic decasyllable, how is this metrical pattern prosodically realized? This will be examined by invoking the parameters of poetic organization, proposed by Hanson and Kiparsky (1996, 292). These parameters set the range of variation in periodic meters and crucially distinguish between the structure of metrical feet, in (26b), and their prosodic realization, in (26a). The latter is reflected in the prosodic characterization of metrical position, the type of prominence utilized by the meter, and the site within the foot in which prominence is constrained. (26) General parameters of periodic meters (Hanson and Kiparsky 1996) a. Structure i. Number of feet (1/2/3/4 . . .) ii. Headedness of feet (right/left) b. Realization i. Position size: maximal size of a metrical position (m/s/f/o) ii. Prominence site: where prominence is constrained (S ) sU/ S ) P/ W ) sP/ W ) U) iii. Prominence type: what prosodic category defines prominence (weight, stress, strength, pitch accent) By proposing that the epic decasyllable is a trochaic pentameter, we have already fixed the structural parameters, the number of feet and their headedness. Some of the realization parameter settings are straightforward. As already noted, the size of a metrical position clearly corresponds to a syllable—since a decasyllabic line is invariably exactly ten syllables long. Much less straightforward is the identification of prominence type. The restriction on the distribution of monosyllables in (25) strongly suggests that the prosodic entity relevant for this meter is the prosodic word, which may not exhaustively occupy a weak position in a metrical foot. Whatever prominence type may be represented by the prosodic word, it can at least be stated that prominence is excluded from the weak positions of the meter, which sets the prominence
76
Draga Zec
site to W ) sP (if weak, then not prominent). Thus, in (25), the prosodic word perniciously contributes prominence to metrically weak positions. But what is the phonological basis for the prosodic word, and in particular, for the monosyllabic prosodic word, to serve as the source of prominence in this meter? Any prominence that the prosodic word may possess is entirely due to its a‰liation with pitch accent, a bundle of properties associated with the head of the prosodic word. As noted in section 4.1, pitch accent is characterized both relationally, as signaled by stress, and nonrelationally, by virtue of tone. It is the latter property that is dominant: the entire bundle of properties inherits from tonal salience its nonrelational nature and, as a result, pitch accents in mono- and polysyllabic words will be on an equal footing. Thus, if taken as the basis of prominence, pitch accent will make mono- and polysyllabic forms equally prominent. But taking pitch accent to be the prominence type of the epic decasyllable is at odds with the fact that the accent of monosyllables, but not that of polysyllables, is subject to distributional restrictions in this meter. As already shown in section 4.2, the accents of polysyllables are in no way metrically restricted. Crucially, it is the pitch accents of monosyllables that this meter singles out as its prominence basis. The obvious question, then, is how to grant special status to the accents of monosyllables. The solution proposed is the following: while pitch accent is taken to be the prominence type selected by this meter, metrical restrictions are placed not directly on the distribution of pitch accent, but indirectly on its head domain. Since the syllable bearing pitch accent corresponds, of course, to a prosodic word’s head, the relevant domain will correspond to the prosodic word. The relevant properties regarding the realization of the epic decasyllable’s trochaic structure are summarized in (27): (27) Periodic organization of epic decasyllable a. Structure i. Number of feet (five) ii. Headedness of feet (left) b. Realization i. Position size: s (syllable) ii. Prominence site: W ) sP iii. Prominence type: pitch accent contained within its head domain Our initial empirical generalization has been that monosyllabic prosodic words play a crucial role in propelling the trochaic pattern of this meter. But although, in descriptive terms, monosyllables obviously play a significant role, there is no mention of prosodic word size in the formal account of this meter in (27). This is because the special status of monosyllables is only derivative. It results from the interaction of
The Prosodic Word as a Unit in Poetic Meter
77
two aspects of the realization of this verse: position size and prominence type. Since a metrical position corresponds exactly to a syllable, it follows, trivially, that only a pitch accent’s head domain of precisely that size exhaustively occupies a single metrical position. In sum, monosyllables have no special status among prosodic words, but appear to be playing a special role in this verse because the size of a metrical position invariably corresponds to a single syllable. A further consequence of the characterization in (27) is the asymmetry between the accents of mono- and polysyllabic prosodic words. While the former are directly targeted by the metrical restrictions on this meter, the latter are left unrestricted. Those pitch accents that are heads of polysyllabic prosodic words are free to appear in any metrical position within the line, and they indeed do, as we have seen in section 4.2. In sum, the requirement regarding the head domain of pitch accent takes e¤ect only when the head domain’s size is identical to the size of a metrical position. To conclude, the prosodic realization of the epic decasyllable’s trochaic structure is captured by prohibiting pitch accent, contained within its head domain, from occupying the weak positions of meter. This requirement is relaxed in colon-initial feet, which freely admit monosyllables in weak positions, stated in (28). (28) INITIAL LICENSE 1 In colon-initial feet (first and third), the weak position may be occupied by a pitch accent contained in its head domain—that is, by a monosyllabic prosodic word. This constraint is one of the two colon-initial relaxations of metrical conditions; the second one will be presented in section 4.6. As a result of ranking (28) above (27b), monosyllabic prosodic words are prohibited from the fourth, eighth, and tenth, but not from the second and sixth positions. In the following section I present Jakobson’s influential analysis of this meter, and compare it with my proposal, which, I argue, is superior to Jakobson’s. In sections 4.5 and 4.6, I present the behavior of function words in this meter: the distribution of monosyllabic prosodic words that serve as proclitic hosts in the former, and the peculiar patterning of monosyllabic free function words in the latter. Section 4.7 addresses the higher-level organization into cola and subcola. 4.4
Comparison with Previous Analyses
As already noted, the epic decasyllable has been characterized both as syllabic and as periodic verse. I focus here on Jakobson’s (1932, 1933, 1952) proposal, which is by far the most detailed analysis in the literature. In his studies of metrical forms, Jakobson generally distinguishes between metrical constants and metrical tendencies, and this approach constitutes the core of his analysis of the epic decasyllable.
78
Draga Zec
In his study of Slavic epic verse, Jakobson (1952, 418) grants special status to colon- and line-final positions in terms of13 ‘‘[a] ‘bridge’ or ‘zeugma’ at the end of both colons: the fourth and the tenth syllable belong to the same ‘word unit’ as the third and ninth syllable, respectively. They are consequently accentless in Sˇtokavian dialects, in which the accent has shifted from the final syllable.’’ In other words, the last two colon-final syllables are stipulated to belong to the same word. By virtue of this, monosyllables are excluded from the fourth and tenth positions. Additionally, any polysyllable satisfying the ‘‘bridge’’ will leave the colon-final position unaccented in the Neo-Sˇtokavian dialect; that this is not the case in the Old Sˇtokavian dialect is fully compatible with Jakobson’s definition of ‘‘bridge.’’ This, according to Jakobson, is a crucial metrical constant of this verse, and thus plays a markedly di¤erent role from pitch accent, which in his analysis encodes a metrical tendency toward trochaic rhythm, as mentioned in section 4.2. However, while the ‘‘bridge’’ makes correct predictions about the absence of the Neo-Sˇtokavian accent in colon- and line-final positions, it does not provide a correct empirical generalization about the distribution of monosyllables in this verse. The ‘‘bridge’’ technically excludes monosyllables from the fourth and tenth positions in a line, yet as we have seen, these metrical positions present only a subset of those that do not allow monosyllables. As demonstrated in section 4.2, monosyllables are excluded not only from the fourth and the tenth, but also from the eighth position in the line, and this is fully captured by my analysis in section 4.3. In his earlier work on the epic decasyllable, Jakobson (1932, 1933) does acknowledge restrictions on the distribution of monosyllables:14 Die einsilbigen Wortganzen fallen auf die ungeraden Verssilben. Daher ist die Konstante abgeleitet, dass sich vor der vierten und zehnten Verssilbe keine Wortgrenzen (genauer keine Grenzen zwischen autonomen Wortganzen) finden. In diesem Gesetze u¨ber die einsilbigen Wortganzen macht sich die trocha¨ische Tendenz des deseterac geltend, d.h. die Tendenz zum Alternieren ‘‘Steigung-Senkung’’: die Senkung darf hier nicht mit einem einsilbigen Wortganzen belastet sein. (Jakobson 1933, 54)
Note that only the absence of monosyllables from the fourth and tenth metrical positions is interpreted as a constant property of this verse and expressed as a constraint on the placement of word boundaries. The absence of monosyllables from the eighth position in the line is seen as simply reinforcing the general trochaic tendency of this meter. This interpretation is consistent with Jakobson’s (1952) analysis: while an explicit metrical device, notably, the ‘‘bridge,’’ is posited to exclude monosyllables from the fourth and tenth positions, no device is posited to exclude monosyllables from the eighth position. In other words, while the exclusion of monosyllables from the fourth and tenth positions is treated as a constant property of the meter, their absence from the eighth position is seen as a metrical tendency. That monosyllables are consistently excluded from the eighth position is
The Prosodic Word as a Unit in Poetic Meter
79
subsumed under the distribution of pitch accents in this and other weak metrical positions, which, according to Jakobson, tends toward trochaic patterning. Yet, as argued in section 4.2.1, the overall distribution of pitch accents in this verse is not su‰ciently systematic to reflect a trochaic tendency. Moreover, even if the accentedness of the eighth metrical position could indeed be subsumed under a trochaic tendency, this move would obscure a crucial fact about this meter: that monosyllables are absent from the eighth position, which can hardly be interpreted as a tendency. In sum, Jakobson’s analysis fails to capture the distribution of monosyllables in this meter, which is its defining property. My approach crucially di¤ers from Jakobson’s in treating the distribution of monosyllables as a defining property of the meter, which, in a robust fashion, reflects its periodic organization. This calls for a single metrical device that would capture all aspects of the distribution of monosyllables. This of course cannot be accomplished by Jakobson’s ‘‘bridge,’’ nor can it be accomplished by any other account in that same spirit. Positing two ‘‘bridge’’ points for the second half line—before the tenth and, additionally, before the eighth syllable—would incorrectly exclude lines that end with a trisyllabic word, and present a very common line type. The ‘‘bridge’’ is thus not an empirically adequate metrical device for capturing the distribution of monosyllables within a decasyllabic line. It does not cover the full range of relevant facts about this verse, nor does it provide an appropriate conceptualization of its metrical structure. As I have argued here, the distribution of monosyllables strongly suggests the organization of a line into metrical feet; in sum, it is a clear indicator of the trochaic pattern of this verse. 4.5
Monosyllabic Prosodic Words Combined with Clitics
A prosodic word that includes clitics has a nested structure, as in (29), and due to this may potentially have an ambiguous status: either the smaller scansion (i.e., poi ) or the larger (i.e., poj ) could be taken as prior: (29) [clitic [host]poi ]poj Generally, it is the larger scansion that is relevant for singling out prosodic words, both in the phonological system and in this meter. Thus, the four syllables in the first colon may be filled by a single prosodic word, which in turn may correspond to a single morphological word, as in (30a), or may have a nested structure: due to the presence of an enclitic, in (30b), or a proclitic, in (30c). (30) a. [Andje`lija]po k [mo`ja]po [vjeˆrna¯]po [ljuˆbo]po ‘Andjelija, my faithful wife’ `` nese]po b. [[naˆmjera]po ga]po k [pred [ve``cˇe]po ]po [na ‘A purpose brought him, toward the evening’
Vk, II, 97, 98 Vk, II, 97, 66
80
Draga Zec
`` tva]po [zlato`krila]po c. [u [ je``zeru]po ]po k [u ‘In the lake, there was a gold-winged duck’
Vk, II, 97, 68
In all three cases, the entire first colon is taken to be occupied by a single prosodic word—that is, by the prosodic word that corresponds to the maximal scansion. Maximal scansion is relevant also for monosyllables, which, combined with clitics, form configurations that behave like polysyllabic prosodic words. Monosyllabic prosodic words, as we have seen, are excluded from the weak positions of this meter, but when serving as proclitic hosts, they are admitted into weak metrical positions, as exemplified in (31)–(34); in fact, they are allowed in all weak positions other than the tenth—that is, the one that belongs to the final foot. A monosyllable combined with a proclitic may occur in colon-initial feet; this of course is expected, since the first and third foot are those in which metrical prohibitions against monosyllabic prosodic words are generally relaxed, due to (28). Examples for the first foot are given in (31) and for the third foot in (32); lexical monosyllables figure in the (a) and (b) examples, with a numeral in the latter, and monosyllabic function words, in the (c) examples.15 (31) First foot [s [s]po ]po s s k s s s s s s a. [u [sto]po ]po sjede k Crnojevic´ Ivo ‘At the table sat Crnojevic´ Ivo’ b. [za [tri]po ]po dana k i tri noc´i tavne ‘For three days, and three dark nights’ c. [od [njih]po ]po pusto k oduzesˇe blago ‘From them, they took away great treasures’ (32) Third foot s s s s k [s [s]po ]po s s s s a. svi mi da se k [u [so]po ]po prometnemo, ‘If we all turned to salt’ b. pa napravi k [do [tri]po ]po cˇasˇe zlatne, ‘And made three gold goblets’ c. il’ c´esˇ vatru k [sa [tim]po ]po ugasiti ‘Or will you extinguish the fire with it’
Vk, II, 88, 97 Vk, II, 52, 69 Vk, III, 5, 159
Vk, II, 49, 10iv Vk, II, 26, 29 Vk, IV, 24, 341
A monosyllable combined with a proclitic may also occupy the weak positions of the second and fourth foot, those in which metrical constraints rigorously take e¤ect. In (33), a monosyllable combined with a proclitic occurs in the weak position of the second foot: the noun lov ‘hunt’ figures in (33a), the numeral dva ‘two’ in (33b), and the function word njoj ‘it-feminine’, which, when combined with a clitic, is endowed with a prosodic word status, in (33c).
The Prosodic Word as a Unit in Poetic Meter
(33) Second foot s s [s [s]po ]po k s s s s s s a. vec´ on [u [lov]po ]po k u planinu podje ‘But he went hunting in the mountain’ b. dje se [do [dva]po ]po k druma rastavljaju ‘Where two roads fork’ c. kralj ga [u [njoj]po ]po k u svatove zove, ‘In it, the king invites him to his wedding’
81
Vk, II, 7, 53 Vk, III, 49, 178 Vk, II, 61, 14
And, in (34), a monosyllable preceded by a proclitic occupies the weak position of the fourth foot, with the noun red ‘line’ in (34a), the numeral tri ‘three’ in (34b), and the function word njoj ‘it-feminine’ in (34c). (34) s s s s k s s [s [s]po ]po s s a. Kako dodje, k s njima [u [red]po ]po stade. ‘As he came, he got in line with them’ b. te uprezˇe k konje [u [tri]po ]po reda; ‘And he harnessed horses in three rows’ c. te vidjeo k sˇto se [u [njoj]po ]po pisˇe ‘And he saw what was written in it’
Vk, II, 28, 233 Vk, II, 90, 25 Vk, III, 20, 48
This clearly shows that it is the larger scansion that is relevant—that is, the weak position is not construed as filled by a monosyllabic prosodic word. It should be noted that lines with the fourth and eighth positions filled with a monosyllable that hosts a proclitic, as in (33) and (34), are encountered much less frequently than those in which the second and the sixth positions are filled in this fashion, as in (31) and (32). Still, monosyllables hosting proclitics consistently occur in the fourth and eighth positions, albeit with a very low percentage, which is below 1 percent. Their presence is stable, with several cases emerging within any randomly picked set of one thousand lines. Thus, although lines like (33) and (34) are rare, I will consider them metrical—in other words, in full compliance with the constraints that regulate this verse. This is endorsed by the distribution of formulaic expressions across metrical positions. As already noted, the epic decasyllable is a highly formulaic verse in which certain fixed expressions recur with only minimal variation. Of interest at this point is the range of metrical positions compatible with a given formulaic expression. Thus, the phrase na um ‘to mind’, consisting of a proclitic followed by its host, appears in the fixed expression in (35), with either the order in (35a) or in (35b). (35) a. [na [um]po ]po pade b. pade [na [um]po ]po ‘Came to mind’
82
Draga Zec
The examples below show that the phrase na um is compatible not only with the first and third foot, as in (36), but also with the second and fourth foot, as in (37)— in fact, with all feet other than the final one. (36) a. [na [um]po ]po pade k begu Radul-begu ‘It occurred to beg Radul-beg’ b. Igumanu k [na [um]po ]po pade Simo, ‘The abbot recalled Simo’ (37) a. Pade [na [um]po ]po k Crnojevic´ Ivu, ‘It occurred to Crnojevic´ Ivo’ b. Ali njemu k odmah [na [um]po ]po pade, ‘But it occurred to him instantly’
Vk, II, 74, 40 Vk, II, 13, 190 Vk, II, 88, 116 Vk, II, 38, 79
Moreover, numerals often co-occur with the preposition do ‘up to’, which in this context serves merely as an empty filler. In (38), do trista ‘up to three hundred’ can only be interpreted as exactly ‘three hundred’, which suggests that do contributes no meaning of its own, and that its sole function is to ensure correct syllable count within the line. (38) a. te pokupi k do trista svatova ‘And he gathered three hundred wedding guests’ b. te nadjosˇe k do trista izvora ‘And they found three hundred springs of water’
Vk, III, 73, 42 Vk, II, 26, 26
Combinations of do with a monosyllabic numeral are found not only in the first and third foot, as in (39) and (40), but also in the second and fourth foot, as in (41) and (42): (39) First foot [do [dva]po ]po brata, dva Nedic´a mlada ‘Two brothers, two young Nedic´es’ (40) Third foot a. Vino piju k [do [dva]po ]po pobratima ‘Two bosom friends were drinking wine’ b. eto otud k [do [tri]po ]po gorske vile, ‘From there came three mountain sprites’ (41) Second foot a. kada [do [dva]po ]po k kneza pogibosˇe ‘When two princes perished’ b. dje se [do [dva]po ]po k druma rastavljaju ‘Where two roads fork’
Vk, IV, 26, 69
Vk, II, 16, 1 Vk, II, 11, 15
Vk, IV, 24, 441 Vk, III, 49, 178
The Prosodic Word as a Unit in Poetic Meter
(42) Fourth foot a. da ti kazˇem k do dve [do [tri]po ]po recˇi: ‘Let me say to you two or three words’ b. Na Balacˇku k jesu [do [tri]po ]po glave: ‘On Balacˇko, there were three heads’
83
Vk, II, 14, 95 Vk, II, 28, 608
Thus, formulaic expressions, which, in a sense, constitute the building blocks of this verse, when taking the form of a proclitic hosted by a monosyllable, are compatible with both the fourth and the eighth position within the line. However, this generalization does not extend to the final foot: a monosyllable combined with a proclitic may not occupy the tenth metrical position. While a few lines have been attested with a functional monosyllable combined with a proclitic occupying the tenth metrical position, not one case has been found with a lexical monosyllable occurring in this configuration.16 We may thus conclude that, in the final foot, it is the smaller scansion that counts, and that therefore, lines such as (43) are not within the range of metrical possibilities in this verse. (43) *s s s s k s s s s [s [s]po ]po Ali njemu k odmah pade [na [um]po ]po (Constructed on the basis of Vk, II, 38, 79) I will attribute the unmetricality of lines such as (43) to the special metrical status of the line-final foot, an important property of this verse that is addressed in section 4.7. 4.6
Monosyllabic Free Function Words
A monosyllabic free function word has no pitch accent of its own, and is thus incapable of endowing a metrical position with prominence, as noted in section 4.1. As such, it should in principle be compatible with any metrical position, and therefore also with the weak position in the foot. But while the distribution of monosyllabic function words is not constrained by any of the metrical conditions in (27), it is not free, either. In fact, it directly reflects, and is subsumed under, the set of constraints that govern the arrangement of hierarchically organized prosodic constituents (as in Nespor and Vogel 1986; Hayes 1989) within a metrical line. Both the line and the colon are subject to alignment with prosodic constituents. The right edge of the line is generally associated with syntactic boundaries, either phrasal or clausal, which strongly suggests that a metrical line is minimally aligned, at its left and right edges, with an intonational phrase boundary. The colon, on the
84
Draga Zec
other hand, is minimally aligned with the edges of a prosodic word, as stated in (44) and (45) (following McCarthy and Prince 1993a). (44) COLON LEFT The left edge of the colon coincides with the left edge of a prosodic word. (45) COLON RIGHT The right edge of the colon coincides with the right edge of a prosodic word. The e¤ect of (44) and (45) is best evidenced at the end of the first and the beginning of the second colon, and captures the already-stated structural property of this verse—that the decasyllabic line is divided into two cola by a fixed caesura, coinciding with a word boundary. To illustrate this, while colon boundaries may align with higher-level prosodic constituents—for example, the prosodic phrase, as in (46)— they minimally align with prosodic word boundaries, as in (47): (46) [[[na [Krusˇevcu]po ]po ]pf k [[gradu]po [bijelome]po ]pf ‘In Krusˇevac, the white city’
Vk, III, 10, 45
(47) [[[na [bijelu]po ]po k [gradu]po ]pf [[[Smederevu]po ]pf ‘In the white city of Smederevo’
Vk, III, 10, 53
Constraints (44) and (45) have an important impact on the distribution of free functional monosyllables. Because its morphological edges do not coincide with any prosodic edges, including the edges of a prosodic word, a monosyllabic function word that appears at the left edge of a colon violates (44), and one appearing at its right edge violates (45). In sum, a monosyllabic function word (here mwF ) prevents prosodic edge alignment both at the left edge of a colon, as in (48a), and at its right edge, as in (48b) (where k signifies colon edge): (48) a. * k mwF [po . . . ]po k b. * k [po . . . ]po mwF k A further aspect of the mapping between metrical and prosodic constituents is the general requirement that the poetic material within a metrical line be exhaustively parsed into prosodic words, as stated in (49): (49) CONTIGUITY Any two prosodic words within a colon, poi and poj , are either contiguous, or separated by another prosodic word, pok : * k . . . ]po o [po . . . k where o is a morphological word with no prosodic word status. The presence of a monosyllabic function word anywhere in the line directly conflicts with this requirement, as schematized in (50):
The Prosodic Word as a Unit in Poetic Meter
(50) * k . . . ]po mwF
po [ . . .
85
k
Monosyllabic function words are not, of course, entirely banned from a decasyllabic line, as strict adherence to Colon Left, Colon Right, and Contiguity would require. In fact, monosyllabic function words occur relatively freely at the left edges of both cola, due to another case of colon-initial relaxation of metrical conditions, Initial License 2, stated in (51); and its overriding e¤ect over Colon Left, expressed by the ranking in (52): (51) INITIAL LICENSE 2 In colon-initial feet (first and third), the weak position may be occupied by a free monosyllabic function word. (52) INITIAL LICENSE 2 COLON LEFT The occurrence of monosyllabic function words at the left edge of the first colon is illustrated in (53), and their occurrence at the left of the second colon, in (54). (53) First colon a. vec´ [isprosi]po k lijepu devojku ‘But he was accepted by a beautiful maiden’ b. on [sazida]po k kulu od kamena ‘He built a tower of stone’ (54) Second colon a. pa se svome k on [poklanja]po [kralju]po ‘And he bowed before his king’ b. ‘‘Sjedi, care k pak [pij]po [rujno]po [vino]po ’’ ‘Sit here, czar, and drink of red wine’
Vk, II, 9, 18 Vk, II, 14, 111
Vk, II, 80, 97 Vk, II, 28, 479
As the examples in (55) and (56) show, the entire colon-initial foot may be occupied solely by functional monosyllables, and thus left outside the prosodic word domain; this collocation is found much more frequently at the left edge of the first colon than of the second. (55) First colon a. vec´ on [u [lov]po ]po k u planinu podje ‘But he went hunting in the mountain’ b. al’ josˇ [sestra]po k u dusˇeku spava, ‘But the sister was still in bed asleep’ (56) Second colon a. Ne sluzˇi ga k sˇto on [blaga]po [nema]po ‘He is not serving him because he has no treasure’
Vk, II, 7, 53 Vk, II, 5, 64
Vk, II, 75, 4
86
Draga Zec
b. nek je haracˇ k ko sˇto [Murat]po [recˇe]po ‘May the taxes be as Murat had said’
Vk, IV, 24, 360
Finally, with Contiguity ranked over Initial License 2, as in (57), configuration (50) is prohibited in colon-initial feet. (57) CONTIGUITY INITIAL LICENSE 2 To conclude, constraints and constraint rankings posited thus far have the e¤ect of excluding functional monosyllables from noninitial feet. In a narrowly defined discourse context, however, monosyllabic function words may exceptionally occur in the second and the fourth foot, as illustrated in (58)– (60). These and similar lines occur predominantly within dialogue exchanges, and are hardly ever found within the narrative portion of the poem. This strongly suggests that lines of this type are licensed by virtue of a discourse-based relaxation of both Contiguity and Colon Right in portions of the poem that call for a more discursive style: (58) Second foot: Fourth position a. [Koja [ je]po ]po tu k Roksanda djevojka ‘Which one here is maiden Roksanda’ b. [Kada [c´u]po ]po ja k i tebeka prec´i ‘When will I cross you, as well’ (59) Fourth foot: Seventh position a. Ako l’ bi se k [u [nju]po ]po kad [vratio]po ‘If I ever returned to it’ b. Al’ da vidisˇ k [cˇuda]po josˇ [vec´ega]po ‘If only you saw this yet greater miracle’ (60) Fourth foot: Eighth position a. Nit’ sam bio, k [niti [c´u]po ]po kad [biti]po ‘I never was, nor will ever be’ b. Da ogledam k [mozˇe [li]po ]po sˇto [biti]po ‘To see if anything could occur’
Vk, II, 28, 571 Vk, IV, 24, 628
Vk, IV, 28, 415 Vk, III, 38, 181
Vk, II, 49 (iii), 45 Vk, II, 66, 108
Thus, monosyllabic function words are exceptionally licensed within dialogue, possibly to render such exchanges more lively and more ‘‘natural,’’ as expressed by the following constraint: (61) DISCOURSE LICENSE Functional monosyllables are licensed in dialogue exchanges. In sum, the overall ranking of constraints that regulate the arrangement of prosodic words within metrical constituents is as in (62):
The Prosodic Word as a Unit in Poetic Meter
87
(62)
This ranking allows for the suboptimal configuration (50) to be instantiated in dialogue exchanges in colon-initial feet, and this indeed is the case, as illustrated in (63) for the first colon. (63) First colon a. [daj]po ti [nama]po k Sekulu nec´aka, ‘Give (you) to us your nephew Sekula’ b. [konj]po moj [nije]po k a djevojka tudja, ‘The horse is not mine, and the maiden belongs to another’
Vk, II, 84, 30 Vk, II, 86, 162
However, a functional monosyllable may not occur in the fifth foot, specifically, in line-final position. This distributional gap is not captured by the analysis presented thus far, as summarized in (62). One solution could be to ensure that Colon Right is enforced more rigorously in the second than in the first colon, thus capturing greater restrictiveness imposed at the end of the line. However, greater restrictiveness at the line closure has already been evidenced in section 4.5, which calls for a more general mechanism. In what follows, special restrictions associated with the line-final foot will be attributed to structural subdivisions within the line, beyond those posited thus far. 4.7
Higher-Order Metrical Units
The line-final foot is subject to greater restrictions than other positions in the line, as shown both by the distribution of monosyllabic clitic hosts (section 4.5), and of functional monosyllables (section 4.6). Rather than positing a constraint with the sole purpose of further codifying this metrical position, I will attribute these distributional restrictions to the higher-level organization of a metrical line. A notable structural property of the epic decasyllable is that a line is divided into two cola. The caesura imposes an asymmetric structure, with two feet in the first and three in the second colon, rendering the second colon weightier than the first, as represented in (64):
88
Draga Zec
(64)
Line Colon Left (s s) (s s)
k
Colon Right (s s) (s s) (s s)
As we have seen, the arrangement of prosodic words in a line crucially observes the caesura, invariably positioned after the fourth syllable; a line with a prosodic word occupying the fourth and fifth metrical positions would be unmetrical. This, however, is not the only structural division within a decasyllabic line. The heightened metrical restrictions in the line-final foot will be accounted for by positing an additional break, henceforth the subcaesura, which divides the second colon into two subcola. Unlike the caesura, the subcaesura is only partially fixed: the second subcolon corresponds to a single prosodic word and minimally includes the final metrical foot, as stated in (65): (65) The rightmost subcolon a. Subsumes the line-final foot (metrical requirement) b. Corresponds to exactly one prosodic word (prosodic requirement) Both cases of heightened metrical restrictions take e¤ect within the rightmost subcolon. Its special status follows from the structural organization of higher-level metrical constituents and their relative strength, as represented in (66): (66)
The strong/weak labeling in (66) (following Prince 1989) designates the right colon, and likewise, the right subcolon, as stronger than their left counterparts. The rightmost subcolon thus emerges as the strongest higher-level metrical subconstituent and, as such, is associated with greater metrical restrictiveness. The special status of the line-final foot follows from its obligatory inclusion into the rightmost subcolon.17 By contrast, the weak subconstituents, leftmost colon and leftmost subcolon, allow for relaxations of metrical conditions. The two Initial License constraints, stated in (28) and (51), could profitably be restated as taking e¤ect at the left edge of metrical subconstituents labeled weak. The absence of functional monosyllables follows directly from the characterization of the rightmost subcolon. Note that (65) imposes an obligatory prosodic word boundary at the end of the line. If (65) is rigorously enforced—that is, if it ranks
The Prosodic Word as a Unit in Poetic Meter
89
above Discourse License—then the final foot will not contain any functional monosyllables, which captures the heightened restriction in this metrical position, stated in section 4.6. The metrical restriction stated in section 4.5, according to which a monosyllabic proclitic host may not occupy the tenth position in a line, calls for further refinement. The rightmost subcolon is prosodically transparent: it forms a narrow window in which all prosodic structure, including prosodic word nesting, is visible to, and evaluated by, metrical constraints. As a result, any prosodic structure that it contains has to conform to the metrical requirements of this verse, including all layers of prosodic word structure. (67) STRONG SUBCOLON TRANSPARENCY Every prosodic word included in the rightmost subcolon has to comply with metrical constraints (in particular, (27b)). Thus, the clitic combination na um in the unmetrical line (30), repeated here as (68), violates Strong Subcolon Transparency because one of the prosodic words in the rightmost subcolon exhaustively occupies a weak position of the meter. The fact that, in addition to a monosyllabic prosodic word, there is also a disyllabic scansion available, does not salvage this configuration. (68)
s s s s k s s s s [s [s]po ]po *Ali njemu k odmah pade [na [um]po ]po
The special status of the second subcolon can account for another metrical peculiarity of this verse. As schematized in (69), the second subcolon—which, as already stated, corresponds to the final prosodic word in a line and subsumes the final foot— may vary considerably in size. The extreme case is (69a), in which the subcaesura coincides with the caesura, and (69b), in which the subcaesura is one syllable apart from the caesura. (69) a. b. c. d. e.
k k k k k
Colon Right Subcolon Left Subcolon Right j [s s s s s s]po s j [s s s s s]po ss j [s s s s]po sss j [s s s]po ssss j [s s]po
By far the most common are the types of lines in (69c–e), which jointly account for over 95 percent. In only one of these cases is there a further requirement imposed on the second subcolon: when it corresponds to a single foot, as in (69e). There is a strong tendency for disyllables in line-final position to possess a heavy initial syllable
90
Draga Zec
(Maretic´ 1903, 1907). This is the case in about 75 percent of the total number of disyllables in the line-final foot, compared to the situation in the third foot, in which no more than about 44 percent of disyllables have a long initial syllable. This will be interpreted as a tendency toward prominence required in the strong position when the second subcolon equals a foot, as stated in (70):18 (70) Final-foot prominence (tendency) If the rightmost subcolon corresponds to a single foot, then a. The strong position of the foot has to be prominent (S ) P). b. The prominence type is syllable quantity (P ¼ vowel length). While this is not an absolute requirement but rather a metrical tendency, this tendency is su‰ciently strong—that is, su‰ciently above chance—to deserve a place in the formal account of this verse. Moreover, the property of this verse to avoid vowel length in the seventh and eighth metrical positions, observed by Jakobson, could well be construed as a counterpart of the marked case in (70). In the unmarked case, when the rightmost subcolon’s size exceeds the size of a foot, as in (69c–d), the corresponding word unit preferably begins with a short voweled syllable.19 Cases of greater restrictiveness in line-final position, which have been captured here as metrical conditions on the rightmost subcolon, while apparently disparate in nature, do possess a unifying property: they impose greater regularity, both metrical and prosodic, within the metrically strongest constituent in the line. Constraint (70) could be motivated in this same spirit. The preferred size of the rightmost subcolon is to be greater than a metrical foot, and constraint (70) encodes what amounts to compensatory prominence when this is not the case. Thus, the unmarked size for the second subcolon is to be greater than a metrical foot, which in e¤ect maximizes the window within which a higher degree of metrical regularity is imposed. But the preference for the rightmost subcolon to be greater than a foot could be interpreted in yet another fashion. As already noted, the subdivision into cola is asymmetric, with the second colon being weightier than the first. In light of variation in the size of the rightmost subcolon in (69), the following generalization obtains: The second subcolon is weightier than the first in (69a–c), and equally weighty as the first in (69d). Only in (69e), which, as I have argued, is a marked case, is the first subcolon weightier than the second. Thus, while at the colon level, the second constituent is categorically weightier than the first, at the subcolon level, due to the mobility of the subcaesura, there is a strong preference for the second subcolon to be at least as weighty, or weightier than the first; and constraint (69) encodes what amounts to compensatory prominence when this is not the case. In other words, lines such as (69e) are marked by special prominence because they counter what appears to be the overall asymmetric organization of this verse, calling for any rightmost metrical subconstituent to be weightier than its left counterpart.
The Prosodic Word as a Unit in Poetic Meter
4.8
91
Concluding Remarks
The principal claim of this study is that the epic decasyllable is a periodic meter, corresponding to a trochaic pentameter, with the prosodic word—or rather, pitch accent contained in its head domain—serving as its crucial prominence source. Because of this, the monosyllabic prosodic word emerges as this meter’s sole prominence source, and is, as such, excluded from weak metrical positions. A further important property is the exclusion of monosyllabic free function words, those that possess no prosodic status, from all positions within the line other than colon initial. This is due to this meter’s strict adherence to the principles of prosodic parsing, much stricter in fact that in the ‘‘ordinary’’ language. The general consequence of these two principles is an overall preference in this meter for words that are minimally disyllabic. Since both lexical and free functional polysyllables invariably correspond to prosodic words, their distribution within a metrical line is fairly free, restricted only indirectly, by constraints on monosyllables. Moreover, the proposed higher-level organization of this verse into cola and, in particular, into subcola, and the metrical constraints associated with these structural constituents, further reinforce this meter’s preference for polysyllabic lexical units. The epic decasyllable thus possesses a set of metrical properties that di¤er in crucial ways from the well-known situation found, for example, in the English or Russian iambic meters, in which the distribution of monosyllables is relatively free, while that of polysyllables is heavily constrained. Notes I am grateful to Kristin Hanson and an anonymous referee for valuable comments and sugges´ upic´, former director of the Institute tions on an earlier version of this chapter, and to Drago C for the Serbian Language, for his immense help with the interpretation of the Old Sˇtokavian verse. This study would not have been possible without access to, and extensive use of, Djordje Kostic´’s Serbian Language Corpus, an invaluable resource that includes Vuk’s entire collection of epic poetry. 1. This study is based in fact on the second, third, and fourth volumes (Karadzˇic´ 1958). The first volume includes only the lyric poetry, and only some of the poems are decasyllabic; the fifth volume is not nearly as polished as the others, and is thus a less reliable source of this verse (Maretic´ 1907). Other sources include Njegosˇ’s 1845 collection (Njegosˇ 1972) and the poems collected by Milman Parry in the thirties and later published by Lord (Parry and Lord 1953). 2. In the examples, Karadzˇic´’s collection is denoted by Vk, Njegosˇ’s by Nj, and Parry and Lord’s by P&L. Specified for each example is the volume number (if relevant), the number of the poem within the volume, and that of the line within the poem. 3. The position of pitch accents in the two dialect groups diverges in a systematic fashion. Diachronically, the Old Sˇtokavian is a historical predecessor of the Neo-Sˇtokavian, which, in the fifteenth century, underwent the process of leftward accent shift whereby all noninitial falling
92
Draga Zec
accents became rising accents on the immediately preceding syllable (Belic´ 1956). While the place of the Neo-Sˇtokavian falling accents is much the same as in the Old Sˇtokavian, the Neo-Sˇtokavian rising accents correspond in the Old Sˇtokavian to the falling accents one syllable to the right. The retracted accent, and as a result, the bisyllabic status of the rising accents, are the central prosodic properties that set apart the Neo-Sˇtokavian group of dialects from the Old Sˇtokavian ones, those that had not been subject to this change (see Ivic´ 1958; Inkelas and Zec 1988; Zec 1993). 4. The type of pitch accent does not play a role in the composition of this verse, either. Again, any of the four pitch accents may appear on any of the accent-bearing positions within a line, although there are certain dispreferences. Long accents are rare in the seventh, and even more so, in the eighth position in the line. For example, in the poem from which the excerpt in (11) was taken, and that has 102 lines, we do not find a long rising accent in the seventh, nor a long falling accent in the eighth position in the line; but there are four long falling accents in the seventh, and two long rising accents in the eighth position. Jakobson (1952, 418–419) treats this as a metrical constant, but in fact it is no more than a tendency. 5. Jakobson (1933, 1952) distinguishes between metrical constants and metrical tendencies, and considers the prosodic demarcation of feet in the epic decasyllable to be of the latter type. He attributed the periodic nature of the epic decasyllable not only to the distribution of pitch accent, but also to the positioning of word boundaries. In the latter case, he identified a tendency toward the placement of word boundaries before the odd, rather than before the even verse positions. This however sheds no light on the headedness of feet—that is, on whether this is a trochaic meter. 6. According to Jakobson (1933, 53), Vucˇic´’s performance had been recorded by G. Gesemman at the Sound Division of the Prussian State Library. 7. According to Ruzˇic´ (1975, 147), the poem Marko Kraljevic´ i Arapin (II, 65), in Vuk’s second volume, which has 435 lines, exhibits the following distribution of pitch accents: 1
2
3
4
5
6
7
8
9
10
50.6
36.3
53.3
0.2
80.2
20.7
33.1
41.1
38.4
—
Note that here, the eighth position outweighs both the seventh and the ninth. High accentedness of the eighth position is characteristic of one of Vuk’s best singers, Tesˇan Podrugovic´, who preferred trisyllabic to disyllabic words in line-final position (Ruzˇic´ 1975, 147). For possible reasons, see section 4.7. 8. As detailed in section 4.1, this class includes all lexical monosyllables, as well as those functional monosyllables that possess prosodic salience by virtue of the association with focus. 9. It has been tacitly assumed that the epic decasyllable was composed in the Neo-Sˇtokavian dialect. This is mostly due to Vuk’s own e¤ort to bestow on the Neo-Sˇtokavian dialect, or rather, on one of its regional variants, the status of linguistic norm. However, while it is true that some of Vuk’s best singers are from a Neo-Sˇtokavian background, epic decasyllable has been composed by singers from more than one dialectal area, including the Old Sˇtokavian, as is clear from Vuk’s own commentary. Parry and Lord’s singers come from dialectal regions with marked Old Sˇtokavian traits, as documented by the accentuation they provide, while the Montenegrin origin of Njegosˇ’s collection associates it clearly with a specific Old Sˇtokavian idiom.
The Prosodic Word as a Unit in Poetic Meter
93
10. This is based on 506 accented lines in Vuk’s fourth volume (IV, 2, 12, 13, 17), 471 accented lines in Njegosˇ’s Ogledalo srpsko (poems 2, 50, 54), and more than 1,000 accented lines in Parry and Lord. Accentuation provided in Parry and Lord’s (1953) collection clearly reveals that the dialect of the singers is of the Old Sˇtokavian type (with some Neo-Sˇtokavian ´ upic´ for his generous help with the selection and accentuation traits). I am grateful to Drago C of poems from Vuk’s and Njegosˇ’s collections. 11. Translations of these examples are based on Parry and Lord 1954. 12. The latter property may well follow from the scarcity of grammatical roles that lexical monosyllables could perform: monosyllabic nominals and adjectives may be either in the nominative singular or, if inanimate, also in the accusative singular form. Monosyllabic verbal forms are very few, and restricted to a handful of verbs. 13. A footnote is omitted from the Jakobson quote. 14. ‘‘Monosyllabic word units are found in odd syllables of the verse. This reflects the metrical constant that before the fourth and tenth syllables in verse no word boundary (or rather, no boundary between autonomous word units) may occur. This rule regarding monosyllabic word units gives validity to the trochaic tendency of the deseterac, the tendency toward risefall alternations: the fall may not coincide with a monosyllabic word unit.’’ 15. The preposition/pronoun combinations za me ‘for me’, za te ‘for you’, za se ‘for oneself ’ are treated here as lexicalized units, rather than as sequences of a proclitic followed by its host, for the following reasons. First, the vowel of the preposition is long in this collocation but short in all other cases. Second, these expressions have ‘‘full-pronoun’’ counterparts za mene, za tebe, za sebe; the pronoun is focusable in the latter, but not in the former set. Likewise, expressions such as sa mnom ‘with me’, za mnom ‘after me’, which contain the instrumental first-person pronoun, will be treated as lexicalized, since here as well the pronoun cannot be focused. For examples, see note 16. 16. The following lines contain functional monosyllables combined with a proclitic in the tenth position: (i) da oko sˇta, k vec´e [ni [oko [sˇta]po ]po ]po ‘What about, if not about a trifle’ (ii) a ne mogu k ni da cˇuju [za [nju]po ]po ‘But they cannot even hear about her’ (iii) il ne cˇujesˇ, il’ ne hajesˇ [za [nas]po ]po ‘You either don’t hear, or don’t care about us’ (iv) hoc´e li se k nasmijati [na [nju]po ]po ‘Will he smile at her?’
Vk, II, 74, 6; Vk, II, 97, 24 Vk, II, 96, 34 Vk, III, 16, 60 Vk, III, 51, 77
These are the only lines of this type that I have found; they may well be an exhaustive list of such cases. If lines like these are to be permitted into the canon of this meter, then their privileged status should follow from a special metrical license granted to functional monosyllables that goes beyond those proposed in section 4.6. Lexicalized expressions discussed in note 15, such as za me ‘for me’, za te ‘for you’, za se ‘for oneself ’, as well as sa mnom ‘with me’, za mnom ‘after me’, do not violate any constraints on this meter when they appear in line-final foot, as in (v) Jal’ ne mare, k jal’ ne cˇuju za te? ‘Is it that they don’t care, or don’t hear about you?’
Vk, II, 96, 126
94
(vi) Stani, Andjo, ti ne idi za mnom ‘Wait, Andja, you should not follow after me’
Draga Zec
Vk, III, 54, 205
17. There are a small number of metrically infelicitous lines with lexical monosyllables in the fourth and eighth positions. Below are listed the known examples in Vuk’s three volumes: (i) Ili si [lud]po k i nisˇta ne znadesˇ, ‘You are either mad, or know nothing’ (ii) kamo ti [macˇ ]po k i megdan junacˇki ‘Where is your sword, and the heroic duel’ (iii) najprva [kc´er]po k Mihaila bana, ‘The first daughter of Mihailo ban’ (iv) pak mu sve [tri]po k otsijecˇe glave, ‘And he cut o¤ all his three heads’ (v) Meni se [zna]po k i otac i majka ‘It is known who my father and mother are’ (vi) Da se, jadna, k za zelen [bor]po vatim, ‘If, in my misery, I touch a green fir tree’ (vii) c´erasˇe ga k za puno [dva]po sata ‘They chased him for two whole hours’
Vk, II, 65, 356 Vk, III, 60, 47 Vk, II, 12, 41 Vk, II, 28, 665
Vk, III, 55, 120 Vk, II, 50, 135 Vk, IV, 28, 381
Crucially, however, no cases have been found with lexical monosyllables in line-final position, which is consistent with maximal metrical restrictiveness associated with the line-final foot. 18. This condition appears to be subject to singers’ individual strategies, which would lead to a reduced number of disyllables with a short initial syllable in line-final position (Ruzˇic´ 1975, 147). Jakobson (1952, 418–419) treats the quantitative clause as a metrical constant, on the basis of Vucˇic´’s two poems (783 lines), in which ‘‘an accented short is avoided in the penult (ninth) syllable.’’ Thus in this case penultimate syllables are reportedly long, regardless of the size of the word unit they belong to. It could be that, in Vucˇic´’s style, length in the ninth position is a condition relevant for any final foot, rather than just for those singled out by a specific shape of the second subcolon. 19. See note 4.
5
The Word in Tiberian Hebrew
Bezalel Elan Dresher
5.1
Introduction
The notion of a ‘‘word’’ is subject to various ambiguities. The text of the Hebrew Bible (Elliger and Rudolph 1977) provides us with several conflicting notions of word. The Biblical text consists of two main layers, and the written word—in the sense of letters surrounded by blank space—di¤ers in each layer. The earlier layer contains a consonantal text, devoid of almost all indications of voweling and punctuation. I will show that the word in this consonantal layer corresponds to a potential prosodic word, that is, a unit that could be an independent word for purposes of phrasing, whether or not it actually functions as such in any particular context. Since such words are not necessarily prosodic words in every context, I will call the word in this layer an orthographic word. Ordinary written Hebrew makes similar word divisions. To more precisely indicate correct pronunciation and phrasing, various diacritic marks, or ‘‘points,’’ were later added to the consonantal text, producing a ‘‘pointed’’ text. These marks include vowel signs, some allophonic consonantal distinctions, and an elaborate system of ‘‘accents’’ that indicates position of stress, division into verses, and a highly articulated prosodic parse of each verse. Orthographic words of the consonantal text (potential prosodic words) can be joined together by hyphens to create a larger unit, the (actual) prosodic word. As we shall see, these constituents are prosodic words in the sense that they count as words for purposes of phrasing. I will show that the principles for forming prosodic words—rules of cliticization—are quite complex and interact in intricate ways with other aspects of prosodic structure, such as the phonological phrase and the intonational phrase. Turning to the evidence of the phonology, I will distinguish between the prosodic word and the phonological word, which is the notion of word referred to by the phonology proper (segmental processes, syllabification, stress), as opposed to the phrasing. Though the phonological word necessarily has some relation to the prosodic word, the two concepts are not identical. Thus, a study of the word in Biblical
96
Bezalel Elan Dresher
Hebrew bears on issues of the syntax-phonology mapping in contemporary linguistic theory, as well as on the notion of levels in Lexical Morphology and Phonology (LMP). 5.2
The Consonantal Text: The Orthographic Word
In the consonantal text, all content words, such as nouns, verbs, adjectives, and numerals, are separate words separated by a space from adjacent words. Most prepositions are also written as independent words. Prepositions that consist of only a single consonant (or consonant plus schwa, depending on whether the schwa is analyzed as inserted by rule or part of the underlying form), however, are written as bound prefixes, with no space separating them from what follows. It is clear that word status is not connected to semantics in this case, because all these prefixes have variants or synonyms consisting of more phonological material, and these are invariably written as independent words. Independent word b mo (poet.), im l mo (poet.), el k mo (poet.) e e e
(1) Prepositions Prefix a. b( )b. l( )c. k( )-
Gloss ‘in, at, with, by’ ‘to, for’ ‘like, as’
e e e
Morphemes of the form C( ) do not make up a full syllable. Therefore, the above observations suggest the generalization that morphemes that consist of less than a full syllable are not written as independent words. The prepositions in (1) have variants of the form Ci- when prefixed to words that would otherwise have an initial syllable with a schwa. e
(2) Variants of prefixed prepositions a. C b Dava´r ‘in a word, matter’ (absolute) b. CibiDvar (unprefixed: d var) ‘in the matter of ’ (construct) e e
e
The @ i alternation in these prepositions is phonological, with i being the default vowel in closed syllables.1 The replacement of schwa by the full vowel i does not change the bound prefixal status of the prepositions. In this category we can also place the conjunctive w( )- ‘and’, called waw after the Hebrew letter used to represent it. This morpheme is written u- before a syllable containing schwa or a labial consonant but remains a prefix. e
e
The Word in Tiberian Hebrew
97
(3) Conjunctive waw a. Elsewhere w Dawı´D ‘and David’ b. Before schwa ux na´an ‘and Canaan’ c. Before labial umirya´m ‘and Miriam’ e
e
The u- allomorph of the conjunctive forms an exception to an otherwise regular rule that Hebrew syllables, and hence words, begin with a consonant. Thus, u- does not form a proper syllable and so also falls under the generalization that morphemes that consist of less than a full syllable are not written as independent words. Phonological subminimality is a su‰cient condition for a morpheme to be written as an a‰x, but it is not a necessary condition. The preposition min ‘from, out of, than’ frequently occurs in the form mi-, where the final n has been historically assimilated to the following consonant. Synchronically, the allomorph mi- causes gemination of a following consonant and is always written as a prefix. (4) The preposition min a. min min hae´s ‘from the tree’ b. mimidd va´sˇ ‘than honey’ c. memee´s ‘from a tree’ e
Before gutturals, which do not normally geminate, the vowel is lengthened to e. However, this lengthening does not restore the prefix to the status of independent word.2 Another morpheme with the shape CV that is always written as a prefix is the definite article ha-. Like mi-, ha- causes gemination of the following consonant; when gemination is not possible, the vowel of the article is lengthened to a.3 (5) The definite article haa. hahamme´lex ‘the king’
98
Bezalel Elan Dresher
b. hahae´veD ‘the servant’ The interrogative ma ‘what’ occurs in a number of variants: ma, ma, me. These variants occur under various segmental and prosodic conditions. This morpheme is almost always written as a separate word, though there are a few cases in which it is joined to a following word. In Isa 3:15, mlkm is to be read malla xe´m lit. ‘what to you (m.pl.)’, (i.e., ‘how dare you’); in Ex 4:2, mzh is to be read ma-zze´ ‘what is this?’ It is interesting that in both these cases the morpheme ma- is attached to a function word (a pronoun in the first case and a demonstrative in the second). Function words are more prone to be fused with other morphemes into single words.4 Thus, orthographic words are potential prosodic words. To qualify, a form must meet the minimum criterion of having at least a full syllable CV, where V is not schwa. Full vowels created by phonological processes do not count. On top of that, a morpheme must exhibit a certain syntactic-semantic independence—hence, ma is a potential prosodic word, ha- is not. 5.3
The Pointed Text: The Prosodic Word
The Biblical text was gradually stabilized and fixed in the centuries leading up to the first century of the common era (Cross and Talmon 1975; Sa´enz-Badillos 1993). At a certain point, no further changes were permitted to be made to this text. Therefore, to this day Torah scrolls that are used for public readings consist only of a consonantal text, with no indications of verse divisions, stress, or other prosodic markers. The prohibition against adding markings did not apply to texts intended for private or nonliturgical purposes, and symbols for vowels, consonant diacritics, and an elaborate system of ‘‘accents’’ to mark phrasing began to be introduced in the 6th and 7th centuries c.e., presumably to preserve the pronunciation of the traditional reading of the text (Goshen-Gottstein 1963). This activity was carried on for a number of generations by scholars known as Masoretes. A number of distinct but related schools arose; the best known was associated with a group working around the city of Tiberias and so is known as the Tiberian system (Dotan 1971; Yeivin 1980). The Tiberian system of accents represents a highly elaborated prosodic representation that, among other aspects of pronunciation, organizes the text into hierarchical groupings of verses, phonological phrases, and prosodic words (Dresher 1994). The Masoretes were not free to tamper with the consonantal text itself. While they could add diacritic marks over, under, beside, or even inside letters, they could not change or transpose letters or add or remove spaces between words. To indicate that two or more orthographic words are to be considered as a single prosodic word, the Masoretes connected the words in question by a hyphen, called maqqef.
The Word in Tiberian Hebrew
99
Whether an orthographic word in the consonantal text is cliticized to a following word or heads its own prosodic word has phonological consequences. A prosodic word has a single main word stress on the final or penultimate syllable. One rule that applies only to syllables bearing main word stress is Tone Lengthening. (6) Tone Lengthening Lengthen a vowel bearing main stress in its prosodic word. Conditions: a. The rule does not apply to the low vowel /a/ when followed by two consonants; b. The rule does not apply to verbs. (Prince 1975) This rule is exemplified by the accusative particle ey. In the majority of cases, this particle is attached by maqqef to the following word, indicating that it is cliticized to it and does not have its own word stress. In these cases the particle is pointed with the vowel e, as in (7a). When it is an independent prosodic word (7b), it is pointed with the vowel e. (7) The accusative particle a. As clitic eyey-hao´r acc-the.light (Gen 1:4) b. As independent word e´y e´y hasˇsˇama´yim acc the.heavens (Gen 1:1) 5.4
Cliticization in the Tiberian Text
Whether an orthographic word is cliticized or not depends on a complex set of prosodic, phonological, and syntactic conditions, some of which are reviewed in the following sections. It turns out that cliticization is tightly tied in with the entire Tiberian prosodic system and cannot be understood without taking into account the principles of phrasing. 5.4.1
Rudiments of the Tiberian System of Accents
For purposes of the current discussion, it is necessary to know that the Tiberian diacritics known as ‘‘accents’’ fall into two groups. A conjunctive accent on a word indicates that the word is in the same phonological phrase as the following word; a disjunctive accent indicates that its word is phrase final. Disjunctive accents, in turn, are arranged into four hierarchical classes, conventionally designated D0, D1, D2, and D3, where D0 represents the strongest disjunction (coming only at the end of a verse and at the end of the half-verse) and D3 represents the weakest (Cohen 1969).
100
Bezalel Elan Dresher
A phrase ending in an accent of level Di is divided by an accent of level Diþ1, until the D3 level. A D3 phrase is divided by another D3. Therefore, unlike much contemporary work that assumes strict layering of phonological phrases, the Tiberian prosodic representation divides each verse into nested phonological phrases.5 The principles governing the division of a verse into phrases, and hence the distribution of the accents, are extremely complex, and though some of the leading principles and rules for particular circumstances are now known, much remains to be discovered (see Arono¤ 1985; Breuer 1982; Dresher 1994; Janis 1987; Price 1990; Wickes 1887; among others). Cliticization is integral to the entire system, because phrasing is sensitive to the number of words and to the prosodic weight of words, and cliticization a¤ects both: cliticization can change two short words into one long word, for example.6 Therefore, cliticization is woven into the phrasing algorithm; it cannot be regarded as a preliminary step that takes place prior to the division into nested phrases or, conversely, as a late fix up that follows the division of words into phrases. The principles governing cliticization are therefore particularly complex, because, being situated at the interface between word and phrase, they involve general principles of phrasing as well as particular idiosyncrasies of lexical items. The most detailed discussion of cliticization in the Biblical text that I know of is that of Breuer (1982, chapter 7). Breuer proposes a series of descriptive generalizations that set out conditions under which cliticization is facilitated or blocked. These generalizations take the form of conditions akin to the ‘‘preference laws’’ of Vennemann (1988) or the constraints of much current phonological theory, notably Optimality Theory (OT; Prince and Smolensky 1993). OT proposes a theory of how conflicts among constraints can be accommodated. Generally speaking, the principal categories of cliticization are the following: 1. Small words: Some small words have an inherent tendency to be cliticized.7 2. Simplification of phrasing: Cliticization simplifies the phrasing, either by reducing the number of conjunctive accents in a phonological phrase or by reducing the number of phonological phrases. 3. Clash avoidance: Clash avoidance averts a stress clash by relieving the cliticized word of its clashing main stress. 5.4.2
Small Words
Cliticization occurs most readily to small monosyllabic words that have a short vowel in a closed syllable. Breuer divides these words into two classes: those that are generally cliticized to any word, short or long, and those that are regularly cliticized only to short words. Breuer (1982, 167) gives the following list of words of the first class.
The Word in Tiberian Hebrew
101
(8) Small function words that can be cliticized to any word ey ‘accusative particle’, al ‘on’, el ‘to’, min ‘from’, aD ‘until’, im ‘with’, im ‘if ’, al ‘not’, bal ‘not’, pen ‘lest’, af ‘also’, ma ‘what’, kol ‘all’, ben ‘son’, bay ‘daughter’, ey ‘time’ Most of these words are straightforwardly function words—the accusative particle, prepositions, negative particles, various subordinating complementizers, and quantifiers. The nouns ben ‘son’ and bay ‘daughter’ might appear to be content words; however, they are also used in contexts where their lexical meanings are attenuated or lost and take on a more functional cast. The word ben, for example, can designate a quality (ben-h a´yil lit. ‘son of valour’ ¼ ‘valiant’) or mean ‘deserving of ’ (ben mo´y lit. ‘son of death’ ¼ ‘he shall surely die’) or be part of an expression indicating age (b ne-sˇana´ lit. ‘sons of a year’ ¼ ‘of the first year’, i.e., ‘less than a year old’), and so on. Similar considerations apply to bay (e.g., e´z bay-sˇ naya´h ‘a she-goat in its first year’, bay-b liyya´al lit. ‘daughter of baseness’, ‘a worthless woman’) and less obviously also to ey ‘time’, perhaps because of its association with time (l e´y ziqnayo´ ‘in his old age’). The interrogative ma ‘what’, though it fits semantically, appears to be out of place because it has an open syllable. However, Breuer points out that it functions as a closed syllable (maC ) because it causes gemination of a following consonant. However, it is not cliticized when followed by a guttural, which does not geminate. Hence, we have the following pair. e
e
e
e
(9) Cliticization of ma when a geminating consonant follows (ma-ppisˇı´)D2 (ma´ hattayı´)D1 what-my.trespass what my.sin ‘What is my trespass? What is my sin?’ (Gen 31:36) The first instance of the word ma in this verse is cliticized to a word with a geminating consonant. In the second phrase, the word following ma is long and has an initial consonant that is a nongeminating guttural; therefore, ma acts like a word ending in an open syllable and does not cliticize in this phrase. In addition to the small words in (8), Breuer identifies another set of small words that are more restricted in their tendency to cliticize. In general, these words tend to cliticize only to short words. Breuer divides these words into two lists: the words in the first list (10a) cliticize more readily than those in the second (10b). (10) Small (mostly) content words that can be cliticized to short words a. am ‘also’, ax ‘but’, raq ‘only’, yaD ‘hand’, kaf ‘palm’, am ‘people’, dam ‘blood’, d var ‘word (const.)’, har ‘mountain’, sar ‘o‰cer’, an ‘garden’, rav ‘great’, ha' ‘holiday’, rax ‘soft’, n um ‘speech’ e
e
102
Bezalel Elan Dresher
b. af ‘anger’, mas ‘tax’, al ‘heap’, qasˇ ‘straw’, pay ‘morsel’, ay ‘winepress’, bar ‘son (aram.)’, haD ‘one (aram.)’, sˇen ‘tooth (const.)’, hoq ‘statute (const.)’, mor ‘myrrh (const.)’, tom ‘integrity (const.)’, tam ‘complete (const.)’, sˇal ‘do¤ ’, raD ‘subdue’, hay ‘live’, at ‘you (f.sg.)’, ze ‘this’, bah ‘in that (aram.)’, b aD ‘for’, n qam ‘revenge (const.)’, sˇ 'ar ‘young of an animal (const.)’, p sˇar ‘interpretation (const.) (aram.)’, l ven ‘white (const.)’, m lox ‘reign’ (const.) e e
e
e e
e
These lists consist mainly of content words, though the first three words in (10a) are function words with syntactic/semantic values comparable to those in (8). Thus, it is not clear why these words are grouped with (10) rather than (8). Some of the content words in (10), like those in (8), also have wider uses that could arguably put them into the function word class. An example is yaD ‘hand’, which combined with various prepositions can mean ‘by the side of ’, ‘next to’, ‘at the disposal of ’, and so on. Note that af ‘anger’ in (10b) is homonymous with af ‘also’ in (8). Besides the content word function word distinction, these words are also distinguishable phonologically. The final consonant in the noun af derives from an underlying geminate that surfaces in su‰xed forms, such as possessive appo´ ‘his anger’. Many of the other words in (10) likewise have underlying final geminates; thus, these words are not just semantically ‘‘heavier’’ than those in (8) but phonologically heavier, also, though the phonological distinction is neutralized at the surface in unsuffixed forms. The nouns in (10a) di¤er from those in (10b) mainly in that the former are more common. The words in (10b) consist of Hebrew nouns together with a mixed bag of other parts of speech, including the odd verb, preposition, demonstrative, and pronoun, and even some words in Aramaic. Thus, tendency to cliticize depends on a variety of factors, including phonological weight, morphological/syntactic class, semantic function, and commonness.8 Many words in (10) are construct forms. The construct raises special problems for the definition of word, but we cannot pursue this topic here. Breuer (1982, 168) points to pairs such as those in (11) as showing how cliticization of words in group (10) is sensitive to the length of the following word; cliticization of am ‘also’ applies before a short word (11a) but not before a long word (11b).9 (11) Cliticization of am depends on the length of the following word a. As clitic (w 'am-oyı´)D2 and.also-acc.me (2 Sam 2:7) b. As independent word (w 'a´m anoxı´)D2 and.also I (2 Sam 2:6) e e
The Word in Tiberian Hebrew
103
Given what I have reported on here, one might expect that the words listed in (8) and (10) are freely cliticizable—everywhere, in the case of the words in (8), and before short words for (10). This, however, is not the case. As mentioned above, cliticization interacts with other aspects of phrasing. Other constraints on the phrasing algorithm can conflict with cliticization in certain situations, and in these configurations, cliticization is systematically blocked. For example, there is a very strong constraint that the half-verse, which ends with a D0 accent, should consist of at least two phrases. In some verses, the main division is such that one of the half-verses contains only two words, one of which is a small cliticizable word. In such a case, the small word almost always remains an independent word in its own phrase, marked with a disjunctive accent. (12) Half-verse contains only two words (w e´y)D1 (b yue´l)D0 and.acc Bethuel (Gen 22:22) e
e
Another constraint that applies to D0 phrases is that a long word does not easily coexist with another word. This phenomenon can be understood as due to a slowing down of the reading in prominent positions, so that a long word in such a position counts as if it were two words, hence already enough to fill a whole phrase (Dresher and van der Hulst 1998). Thus, a small word is generally not cliticized to a long word in a D0 phrase but again is placed in its own phrase with a disjunctive accent. This phenomenon is illustrated by the following verses. (13) Cliticization of a small word to a long word a. In a D1 phrase (w ey-hairasˇ´ı)D1 and.acc-the.Girgashites (Gen 15:21) b. In a D0 phrase (w e´y)D1 (hairasˇ´ı)D0 and.acc the.Girgashites (Gen 10:16) e e
5.4.3
Simplification of Phrasing
Cliticization can also occur to reduce the number of disjunctive accents in order to create a smoother phrasing. In particular, expected phrasings of the form (14a) and (14b) below may be simplified as shown. (14) Simplification a. (w)Dnþ1 (w w)Dn ! (w w-w)Dn b. (w w)Dnþ1 (w)Dn ! (w-w w)Dn The likelihood of a word being cliticized in these contexts increases with increased shortness of the word. Of course, the small words in (8) cliticize to any following
104
Bezalel Elan Dresher
word unless blocked by the phrasing principles discussed in the previous section, so they would be cliticized in (14) without any further stipulation. The small words in (10) cliticize as a matter of course only to short words; recall (11) above. The contrast in (11) arises where the conditions of (14) do not obtain; however, such words do cliticize to a long word when Simplification is possible. (15) Cliticization of am in Simplification context a. (w ibbane´ 'am-anoxı´)D1 and.I.shall.be.built.up also-I (Gen 30:3) b. *(w ibbane´)D2 (a´m anoxı´)D1 e e
Apart from the small words discussed above, cliticization applies most commonly, according to Breuer, to the subordinating complementizers ki ‘that, for, when’, a˘sˇer ‘that’, and the negative morpheme lo. Cliticization of this type is illustrated in the following example. (16) Example of Simplification via cliticization of lo(w 'e´r lo-yone´)D1 (w lo´ yilhase´nnu)D0 and.stranger not-vex and.not oppress.him ‘Thou shalt neither vex a stranger, nor oppress him’ (Ex 22:20) e
e
The first instance of the word lo in this verse is cliticized, thereby reducing the number of words from three to two and allowing the preceding word to be phrased in the same phrase, rather than forcing it into its own phrase. This is thus an example of (14a). The second instance of this word is not cliticized, because doing so would serve no simplifying purpose, nor is there a stress clash in this phrase. This example shows that, at least in the case of this lexeme, it is preferable to have two words in a phonological phrase than for it to be cliticized, leaving just one word in the phrase, in the absence of other factors favoring cliticization. Another form of Simplification is cliticization to reduce the number of words in a phrase by replacing a conjunctive accent. Such cases can be represented schematically as in (17). (17) Reduction a. (w w w)Dn ! (w w-w)Dn b. (w w w)Dn ! (w-w w)Dn The situations in (17) are the minimal ones in which Reduction can occur. That is, a word that is not inherently cliticizable by the criteria discussed above will not cliticize to reduce a phrase from two words to one word. Cliticization of this type can occur, however, in larger phrases containing more than two words. It follows that Reduction is relevant only in phrases ending in disjunctive accents that support more than one preceding conjunctive accent, that is, that allow more than two words in a phrase. As a rule, the less prominent the phrase, the more words
The Word in Tiberian Hebrew
105
can be fit into it. Therefore, Reduction situations arise often in D3 and D2 phrases but only rarely in D1 or D0 phrases. D0 phrases, for example, normally permit a maximum of two words. This generalization is never violated before a verse-final D0 accent. However, before the D0 that ends the first half-verse (an accent called atnah ), there is a particular situation in which more than two words can occur. Breuer (1982, 156) observes that this configuration arises when the word ki is followed by a word with initial stress that is in turn followed by the D0 word (18b). When ki is followed by a word with noninitial stress in a comparable sequence, it is cliticized, resulting in an ordinary two-word D0 phrase (18a). (18) Cliticization of ki in a D0 phrase depends on the following word a. As clitic (ki-yele´x imma´nu)D0 rather-you.will.go with.us (Nu 10:32) b. As independent word (kı´ va´ vila´m)D0 that came Balaam (Nu 22:36) The pattern exemplified in (18) is the opposite of what we might have expected; since words are cliticized as a way of averting a stress clash (see the following section), the failure of ki to cliticize in just such a position is unexpected. Thus, (18b) is anomalous in two ways: the existence of a three-word phrase ending in D0 and the failure of cliticization. These two anomalies can be connected by supposing that ki in (18b) is treated by the accent system as in fact being cliticized at some abstract level; for reasons that remain unclear, the cliticization is suspended in this particular configuration. 5.4.4
Clash Avoidance
Cliticization can occur to prevent a stress clash between words in the same phonological phrase. In Tiberian Hebrew, a stress clash occurs between two words in the same phonological phrase when the first word has final stress and the second word has initial stress. If the first word ends in a superheavy syllable (a phonologically long vowel in a closed syllable), no clash is considered to occur. The cliticized word in (19a) has final stress when independent (19b). The e¤ect of cliticization is to deprive the word of its main word stress, thus averting a stress clash. Notice that a secondary stress appears on the initial syllable of the cliticized word in (19a). This secondary stress arises through the ordinary rule of secondary stress assignment, counting back two full syllables from the main stress of the entire prosodic word (Dresher 1981b).
106
Bezalel Elan Dresher
(19) Cliticization to avert a stress clash a. (va`yhi-e´rev . . .)D1 and.was-evening ‘And there was evening’ (Gen 1:8) b. (vayhı´ vae´rev)D1 and.was in.the.evening ‘When evening came’ (Gen 29:23) Another means of averting a stress clash is by stress retraction (McCarthy 1979; Rappaport 1984; Revell 1987), examples of which are shown in (20); the first word in each phrase is normally stressed on the final syllable. (20) Stress retraction to avert a stress clash a. (qa´ra la´yla)D0 he.called night (Gen 1:5) b. (to´xal le´hem)D1 you.will.eat bread (Gen 3:19) The vowel onto which stress is retracted must normally be long. For words that do not meet the conditions for stress retraction, cliticization is the only option for avoiding a stress clash, as in (19a). Where retraction can occur, it appears to be the preferred option; where retraction is not permitted, cliticization can occur. Sometimes the stress clash is left unresolved, for reasons explained in detail by Revell (1987). An example illustrating these two options side by side in the same phonological phrase is the following. (21) Stress retraction and cliticization to avert stress clashes (mo´zne se´Deq avne-se´Deq)D2 balances honest weights-honest ‘an honest balance, honest weights’ (Lev 19:36) In the word mo´zne, stress can retract onto the phonologically long vowel o; but the initial vowel of avne is short, so cliticization is the only available option (short of leaving the clash unresolved). 5.4.5
Summary
The preceding remarks on cliticization present only a partial picture of this complex phenomenon. They should su‰ce to show, however, that the principles governing the distribution of prosodic words in the text are bound up with constraints on phrasing that operate at higher levels of the prosodic hierarchy. Earlier proposals for mapping the prosodic structures indicated by the accents from the syntax assumed a derivational approach whereby prosodic structure is built up in a series of steps (Dresher 1994; Janis 1987). The above survey suggests that evaluation of candidate forms by
The Word in Tiberian Hebrew
107
means of ranked constraints, as proposed by OT, o¤ers a promising alternative. I will not, however, attempt such an analysis here (though the reader is invited to begin to construct one from the materials presented above). 5.5
The Word Level in Phonology
Up to now we have considered the notion of the word as represented orthographically in the Tiberian text and have found two types of words associated with the consonantal and pointed text, respectively. Both notions relate to the prosodic word of linguistic theory. The orthographic word of the consonantal text corresponds to potential prosodic words, that is, words that can stand as independent prosodic words in some context. The pointed text indicates which of these potential prosodic words are actually realized as such and which are cliticized. There is another notion of ‘‘word’’ that is relevant to phonological theory in the sense of a level at which certain phonological and morphological processes apply. In the theory of LMP (Kiparsky 1982d, 1985; Mohanan 1982, 1986; Pesetsky 1979), phonology and morphology apply in stages to a series of levels, such as the stem, the word, and postlexical levels. Though a prosodic word is necessarily a domain for word-level processes in LMP, the word level of LMP is not exhaustively characterized by the prosodic words indicated in the Tiberian pointed text. That is, word-level processes apply also to certain subconstituents of prosodic words. In Dresher 1983, I argued that Biblical Hebrew displays some level ordering, but the levels are not exactly what we might expect from the results of other studies. In brief, I argued that there is very little evidence for stem-level phonology, apart from some minor rules that apply to particular morphemes. I connected this fact to the nonconcatenative nature of much Semitic morphology, which does not create a suitable environment for the operation of typical phonological processes. Thus, there is no evidence that su‰xes, for example, need to be distinguished as being stem level or word level. Unlike many dialects of Arabic (Brame 1974; Broselow 1976; Kenstowicz 1981; Kiparsky 2007), for example, object su‰xes, though attached outside of subject-agreement su‰xes as in other Semitic languages, do not appear to be attached at a di¤erent level than other su‰xes. For purposes of stress, syllabification, and segmental phonology, a word like yisˇmorxa´ ‘he will guard you’, from underlying /yaþsˇmorþeþka/, can be subjected to a‰xation and word-level phonology in one pass, without any internal cycles. It follows from LMP that we should not expect the morphology to be sensitive to derived phonological properties, and this prediction is borne out; unlike English, for example, Biblical Hebrew has no wordformation rules that make reference to stress. I also argued that the word level itself must be conceived of di¤erently than in most studies of LMP. We can think of a word as existing paradigmatically in the
108
Bezalel Elan Dresher
lexicon or else as being syntagmatically placed within a phrase. In most studies of LMP, the word is thought of, perhaps correctly, as being in the lexicon. Hence, languages like English have category-changing word-level a‰xes that are best thought of as being assigned in the lexicon prior to insertion in the syntax (e.g., singaer, sadaness, nationalaize). Consistent with this is the fact that word-level phonology in English is not sensitive to the position a word has in its phrase. In Biblical Hebrew, however, these phenomena point in a di¤erent direction: here, word-level phonology applies not to words in the lexicon but to words already placed in a phrase. Evidence for this is that word-level phonology in Hebrew is sensitive to the position of a word in a phrase. This evidence comes from the so-called pausal forms, a ubiquitous feature of Tiberian Hebrew prosody. In Tiberian Hebrew, many words have one form when they are phrase internal—the contextual form—and another form—the pausal form—when they are final in a major phrase (which I take to be the intonational phrase of the contemporary prosodic hierarchy; see DeCaen 2005; Dresher 1994; Goerwitz 1993; Revell 1980, 1981). In most cases, both the contextual and pausal forms can be derived from a common underlying source by the same regular rules of the phonology. The source of the di¤erence can be located in the way rules of stress and reduction apply in pause and in context. Starting from /yaþsˇmorþeþka/, for example, the penultimate vowel is reduced when the word is in context, causing stress to appear on the final vowel and preserving the stem vowel, the result being yisˇmorxa´, as cited above. When the same word is in pause, however, the penultimate vowel is retained and stressed, resulting in the reduction of the stem vowel, yielding yisˇm re´xa (see Malone 1993; Prince 1975 for details). If word-level phonology waits until the whole word has been put together and inserted into its phrase, it follows from LMP that there should be no word-level category-changing a‰xes, and this appears to be correct. Although there exist no word-level a‰xes that apply in the lexicon, there does exist a class of word-level prefixes that create a word-level cycle in the phrase. e
(22) Word-level prefixes a. Prepositions: b( )- ‘in’, l( )- ‘to’, k( )- ‘like’, mi- ‘from’ b. Conjunctive: w( )- ‘and’ c. Definite article: ha- ‘the’ e
e
e e
The prefixes in (22) are just those discussed in section 5.2 as being noteworthy in that they are not written as independent words in the consonantal text, though their syntactic-semantic status might qualify them as being grammatical words. In the case of the prepositions (22a), we observed that they each have variants or synonyms that are independent orthographic words. From the point of view of the syntax, too, these prepositions are best viewed as being introduced into the syntax as independent morphemes with their own syntactic positions. They are then obligatorily cliticized
The Word in Tiberian Hebrew
109
but not before certain word-level phonological processes have applied. One of these is a rule changing /a/ to i in a word-initial closed syllable (Prince 1975).10 (23) A-to-I a ! i/aC
CC
This rule applies to (24a), where there is no prefix. The underlying /a/ surfaces in prefixed forms, such as (24b). However, the rule applies despite the presence of one of the prefixes in (22), as shown in (24c). (24) Examples of A-to-I a. No prefix: A-to-I applies /addel/ ! idde´l ‘he brought up’ b. Lexical prefix: A-to-I does not apply /yaþaddel/ ! y adde´l ‘he will bring up’ c. Word-level prefix: A-to-I applies /w( )aaddel/ ! w idde´l ‘and he brought up’ e
e
e
Another phenomenon that attests to the distinctive status of the word-level prefixes is spirantization. Spirantization normally applies to a (nonemphatic and nongeminate) stop that immediately follows a vowel (see Idsardi 1998 for a detailed discussion). Hence, in (25a) the /k/ and /b/ of the root /ktb/ are spirantized, but the /t/ is not. Following a word-level prefix, however, the /t/ is also spirantized (25b).11 (25) Contrast in spirantization a. Lexical prefix: Medial C of root not spirantized /laþktob/ ! lixto´v ‘to write’ b. Word-level prefix: Medial C of root spirantized /b( )aktob/ ! bixyo´v ‘when writing’ e
This spirantization attests to the presence of a vowel between the k and the t in (25b). Such a vowel would arise on the inner cycle in /ktob/ to break up the illicit initial consonant cluster, spirantizing the /t/ (26b). On the outer cycle, the prefix vowel (whether underlying or inserted to break up the CC cluster) spirantizes the /k/. The addition of the prefix also puts the derived vowel of the inner cycle in the context VC CV; this is a context in which short vowels are deleted, hence the output bixyo´v. In (26a), the prefix plus stem form a single word domain, and so there is no reason to insert a vowel between the first two root consonants (Borer 1979; Idsardi 1998; Rappaport 1984). (26) Word domains Inner cycle Input
a. Lexical prefix la-ktob (laktob)w
b. Word-level prefix ktob (ktob)w
110
Bezalel Elan Dresher
Output Outer cycle Input Output
(lixto´v)w — —
(kVyo´v)w b-(kVyo´v)w (b(kVyo´v)w )w bixyo´v
In sum, Biblical Hebrew supports the general picture of phonology-morphology interaction posited by LMP. It also shows that word-level phonology may have access to the position of a word in a phrase and that there exist postlexical levels that have many of the properties of lexical levels (see Dresher 1983 for further discussion). Of course, since phrasing itself depends on some derived phonological properties (notably, the position of stress), the nature of the iteraction between level-ordered phonology and the phrasing algorithm is not entirely clear. Putting together the results of this section and the previous one, it appears that an adequate analysis of Biblical Hebrew phonology and prosody may require a derivational component as well as parallel constraint evaluation, perhaps along the lines sketched by Kiparsky (2002).12 5.6
Conclusion
For hundreds of years, Biblical Hebrew has been at the center of important developments in linguistics. In this brief survey of some aspects of the word in Tiberian Hebrew, I hope to have shown that this position is entirely merited and that the Masoretic text continues to raise interesting and complex problems that are relevant to current issues in linguistic theory. Notes It is an honor and a pleasure to dedicate this article to Paul Kiparsky, who has contributed so much to our understanding of the word and its place in phonology and morphology. For various kinds of valuable help and illuminating discussions of Biblical Hebrew, I would like to thank Jean Balcaen, Vincent DeCaen, and Bill Idsardi. This research was supported in part by Social Sciences and Humanities Research Council of Canada research grant 410-96-0842. 1. Jou¨on (1947, section 8), against the opinion of Kautzsch (Bergstra¨sser 1962; Gesenius 1910), considers that a stem-initial schwa was actually pronounced, though in a weakened form that does not amount to a normal reduced syllable: biD var. His arguments for this assumption are first, that the schwa corresponds to a vowel that was historically present, and second, that the rule of spirantization applies to the following consonant, indicating the presence of a vowel (*biD bar). However, neither of these arguments is compelling: the historical existence of a vowel does not necessarily bear on its synchronic status; and though spirantization does point to the synchronic presence of a vowel, it does not necessarily indicate that this vowel is present at the surface—see Idsardi (1998) and section 5.5. e
2. A remark is needed concerning the transcription of vowels used here and the issue of vowel quantity. The Tiberian transcription distinguishes seven vowel signs, and the current consensus
The Word in Tiberian Hebrew
111
is that these vowels are distinguished by quality, not quantity, with values approximating to [i, e, , a, , o, u] (Bergstra¨sser 1962; Jou¨on 1947; Khan 1987). Despite the apparent seven-vowel system of the Tiberian transcription, there is a long-standing tradition (Chomsky 1952) of considering the underlying vowel system of Biblical Hebrew to comprise 10 vowels, symmetrically divided into five long and five short: /i, i, e, e, a, a, o, o, u, u/. Even while indicating vowel quantity in transcriptions, writers such as Jou¨on 1947 and Lamdin 1971 are noncommittal as to the phonetic reality of this scheme. However, there is no doubt that the quantitative interpretation makes much better sense of the phonological alternations of Biblical Hebrew than does the purely quantitative interpretation. Thus, I will refer to lengthening and long and short vowels, understanding these terms to refer to a genuine phonological reality in the grammar of Tiberian Hebrew, though not necessarily at the surface phonetic level. c
3. I omit other variants of the definite article that arise in various environments having to do with the position of stress and other peculiarities of the gutturals. 4. For example, prepositions cannot occur with independent pronoun forms, but only with su‰xal forms of the pronoun: l xa´ or ele´xa ‘to you’, never *e´l atta´, where atta´ is the independent form of the second person masculine singular pronoun. e
5. See Dresher 1994 for further discussion of the rationale behind this nesting and its connection with contemporary approaches to prosodic structure. 6. A long word has at least two full syllables before the main stress; a short word does not meet this condition. See Dresher 1981a for discussion of the theoretical basis underlying these definitions. 7. A small word is a word with only one syllable (not counting schwa). 8. Breuer (1982, 171) writes that he includes n um ‘speech’ in this list even though it has a long vowel in a closed syllable because it cliticizes frequently, particularly in certain fixed phrases. ze ‘this’ appears to be out of place here because it ends in an open syllable; however, like ma ‘what’ it causes gemination of a following consonant when it cliticizes, thereby closing its syllable. e
9. The initial /g/ is spirantized after a vowel in both examples in (11). 10. See Prince (1975, 157) for a refinement of this rule. 11. Note that the infinitive prefix laþ is distinct from the preposition l( )a. e
12. See also Dresher (in press) for further discussion of this issue.
6
Faithfulness and Componentiality in Metrics
Bruce Hayes
6.1
Introduction
The field of generative metrics attempts to characterize the tacit knowledge of fluent participants in a metrical tradition. An adequate metrical analysis will characterize the set of phonological structures constituting well-formed verse in a particular tradition and meter. Structures that meet this criterion are termed ‘‘metrical.’’ An adequate analysis will also specify di¤erences of complexity or tension among the metrical lines. Example (1) illustrates these distinctions with instances of (in order) a canonical line, a complex line, and an unmetrical line for English iambic pentameter. (1) a. The li- / on dy- / ing thrust- / eth forth / his paw (Shakespeare, Richard II, V, 1, 29) b. Let me / not to / the mar- / riage of / true minds (Shakespeare, Sonnet 116) c. Ode to / the West / Wind by / Percy / Bysshe Shelley (Halle and Keyser 1971a, 139) The goals of providing explicit accounts of metricality and complexity were laid out in the work of Halle and Keyser (1966) and have been pursued in various ways since then. From its inception, generative metrics has been constraint based: formal analyses consist of static conditions on well formedness that determine the closeness of match between a phonological representation and a rhythmic pattern. The idea that the principles of metrics are static constraints rather than derivational rules has been supported by Kiparsky (1977), who demonstrated that paradoxes arise under a view of metrics that somehow derives the phonological representation from the rhythmic one or vice versa. The idea that grammars consist of well-formedness constraints has become widespread in linguistic theory. An important approach to constraint-based grammars in current work is Optimality Theory (¼ OT, Prince and Smolensky 1993), whose basic
114
Bruce Hayes
ideas have been applied with success in several areas of linguistics. One might expect that metrics would be easier to accommodate in the OT worldview than any other area, given that metrics has been constraint based for over 35 years. Surprisingly, problems arise when one attempts to do this. To begin, OT is, at least at first blush, a derivational theory: it provides a means to derive outputs from inputs. But in metrics, the idea of inputs and outputs has no obvious role to play; rather, we want to classify lines and other structures according to their metricality and complexity. Second, there is the problem of ‘‘marked winners’’: as we will see, many existing lines or other verse structures violate Markedness constraints. Why shouldn’t these marked winners lose out to less marked alternatives? Hayes and MacEachern (1998) attempt to explain this by supposing that whenever a winning candidate violates a Markedness constraint, there are still higher-ranking Markedness constraints that are violated by all of the rival candidates. However, as we will see, this cannot be true in general. In phonology, the reason marked winners can occur is plain: they obey Faithfulness constraints that are violated by all of their less-marked rivals. But it is not immediately clear how Faithfulness can be implemented in metrics: in a patently nonderivational system, where are the underlying forms that surface candidates can be faithful to? Third, the problem of marked winners arises again when we consider metrical complexity. Intuitively, in certain cases we want to say that the Markedness violations of a winner give rise to a complexity penalty. However, as we will see, in many other cases, Markedness violations can occur without inducing any penalty at all. What distinguishes the two cases? Last, there is a problem of the ‘‘missing remedy.’’1 OT defines the output of any derivation as the most harmonic candidate, the form created by gen that wins the candidate competition. Thus, in principle, every unmetrical form ought to have a well-formed counterpart, an alternative that wins the competition that the unmetrical form loses. But this fails to correspond to the experience of poets and listeners; unmetrical forms like (1c) usually sound wrong without suggesting any specific alternative. All of these problems would have a quick and easy solution under a recent proposal made by Golston (1998); see also Golston and Riad (2000). These authors suggest that the unmetrical lines are simply those that violate high-ranked Markedness constraints, and complex lines are those that violate medium-ranked Markedness constraints. This solution is a radical one, since it claims that in metrics—unlike any other component of grammar—there are no e¤ects of constraint conflict. In other grammatical components, it is commonplace for a candidate to win (and sound perfect) even when it violates a high-ranked constraint when all rivals violate even
Faithfulness and Componentiality in Metrics
115
higher-ranked constraints. Moreover, Hayes and Kaun (1996) and Hayes and MacEachern (1998) give evidence for constraint-conflict e¤ects in metrics, so I believe that the strategy of a special version of OT just for metrics would not work in any event. My own proposal for solving the problems outlined above draws from several sources. Following the principle of the ‘‘Richness of the Base’’ (Prince and Smolensky 1993, 191; Smolensky 1996), an OT grammar can be used to delimit a set of well-formed representations, rather than derive one set of representations from another. To derive marked winners, I adopt metrical ‘‘Faithfulness constraints,’’ which are ranked against Markedness constraints and determine which forms emerge as metrical in spite of their Markedness violations. The problem of finding the required underlying representations can be solved by fiat, simply by adopting the surface form of each metrical entity as its underlying form (Keer and Bakovic´ 1997; Bakovic´ and Keer 2001). With Faithfulness constraints in place, the problem of metrical complexity can be addressed by using the stochastic approach to gradient well-formedness developed in Hayes and MacEachern (1998), Hayes (2000), and Boersma and Hayes (2001). Finally, to solve the missing-remedy problem, I assume (following Kiparsky 1977) that the metrical grammar is componential and that candidate representations should be evaluated independently in each component. To be well formed, an output must win the competition for every component. This permits grammars that rule out forms absolutely, without suggesting an alternative.
The data with which I will test my proposals involve two problems that (in my opinion) received only partial solutions in earlier work: free variation in quatrain structure (Hayes and MacEachern 1998) and the distribution of mismatched lexical stress in sung verse (Hayes and Kaun 1996). 6.2
Basics
I assume that a meter forms an abstract rhythmic pattern and that there exists for each tradition a system of principles that determine when phonological material properly embodies a pattern in verse. The verse examined here will be the sung verse of traditional Anglo-American folk songs. For many such songs, the rhythmic pattern of each line can be represented as in (2). (2) [ [ [x [x x]
x x x x ][x x [x x][x x][x
][ ][x x][x
x x x x ][x x x][x x][x x][x x
] ] ] ]
line hemistichs dipods feet
116
Bruce Hayes
This is a ‘‘bracketed grid’’ (Lerdahl and Jackendo¤ 1983; Halle and Vergnaud 1987), which embodies information about the relative prominence of its terminal positions (height of grid columns) and about grouping (constituency at various levels, labeled here at the right side of the grid). The anonymous poet/composers who collectively created the body of Anglo-American folk song sought (tacitly) to provide phonological embodiments of this and similar structures. They did so by matching the rhythmic beats (grid structure) with syllables and stress—and by matching the constituent structure with phonological phrasing. A simple example is the following.2 (3) [ [ [x [x
x x x x][x
][x x][x
It was late
x x ][ x x x ][x x ][x x x][x x][x x][x x][x x][x x
in the night when the squire came
] ] ] ]
line hemistichs dipods feet
home (Karpeles 1932, a33A)
Inspection of this line shows a good match on several grounds: the tallest grid columns are filled with stressed syllables; most of the shortest grid columns initiate no syllable at all; the syllables are fairly well matched with their natural durations; and the main prosodic break of the sentence (after night) coincides with the division of the line into two hemistichs. For discussion and exemplification of these phenomena, see Hayes and Kaun (1996). In English folk songs, it is not just lines that are metrically regulated but also higher-level structures like quatrains. Hayes and MacEachern 1998 (hereafter HM) is a study of quatrain structure, focused in particular on the sequencing of line types within quatrains. For what follows, it will be crucial to make use of HM’s typology of line types, which is reviewed below. A line type that HM call ‘‘3’’ places its final syllable on the 11th grid position, which is the third strong position of the line. The extensive empty grid structure that follows this syllable is detectable in the timing of performance. An example, with its grid structure, is given in (4). (4) [ [ [x [x As
x x x x][x
][x x][x
bright
x x ][ x x x ][x x ][x x x][x x][x x][x x][x x][x x
as the sum-
mer
sun
] ] ] ]
line hemistichs dipods feet
(Ritchie 1965, 36)
G (mnemonically ‘‘Green-O’’) has elongation of the syllable occupying position 11, with no further syllable initiated until the fourth strong position in 15:
Faithfulness and Componentiality in Metrics
(5) [ [ [x [x x]
117
x x x ][ x ][x x ][x [x x][x x][x x][x
A- mong
the
leaves
so
x x x x ][x x x][x x][x x][x x green
O
] ] ] ]
line hemistichs dipods feet
(Sharp 1916, a79)
3f (‘‘three-feminine’’3) has one weakly stressed syllable after position 11 and leaves position 15 unfilled. (6) [ [ [x [x
x x x x][x
She’s
x ] x ][ x x ] ][x x ][x x ][x x ] x][x x][x x][x x][x x][x x][x x]
gone
with the gyp-
sen
Da-
vy
line hemistichs dipods feet
(Karpeles 1932, a33A)
4 is free from any of these gaps; all of the four strong metrical positions are overtly filled and there are no elongations. (7) [ [ [x [x
x x x x][x
The
x ] x ][ x x ] ][x x ][x x ][x x ] x][x x][x x][x x][x x][x x][x x]
keep-
er
did
a
shoot-
ing
line hemistichs dipods feet
go (Sharp 1916, a79)
The distribution of these line types within quatrains is restricted. Inspecting a corpus of 1,028 Appalachian folk songs and other material, HM determined that only certain sequences of 3, G, 3f , and 4 lines can constitute a well-formed quatrain. The list of types that are well attested and assumed to be well formed appears below. For examples of these quatrain types, see HM 478–482. (8)
4444 GGGG 3 f 3f 3f 3f 3333
4G4G 43f 43f 4343 G3G3 3f 33f 3
444G 4443f 4443 GGG3 3f 3f 3f 3
GG4G 3343
G343 3f 343 3f 3G3
HM also lay out and defend an Optimality-theoretic analysis of their data, which is based on a set of 10 metrical Markedness constraints. The idea is that each quatrain type results from a song-specific ranking. The gen function is assumed to provide all of the conceivable schematic quatrain forms, each represented simply as a sequence of line types, for example, 4343. In verse composition, the poet is assumed
118
Bruce Hayes
to adopt a particular ranking of the Markedness constraints so that a single quatrain type wins the Optimality-theoretic competition. The set of possible quatrain types in (8) is modeled by assuming that the poet may freely rank the constraints for purposes of composing any particular song but adheres to that ranking for all of the song’s quatrains. Therefore, the set of quatrains that are predicted to be metrical are those that can be derived by ranking HM’s constraints. The HM analysis predicts the inventory of (8) (or something reasonably close to it) as the factorial typology (Prince and Smolensky 1993) of the constraint set.4 6.3
Problem I: Free Variation in Quatrains
The first empirical problem to be discussed here stems from an apparent inadequacy in the HM analysis, namely, its treatment of free variation. Poets do not always use the same quatrain scheme throughout a multiquatrain song. The most common pattern of variation is one in which the poet uses 3 in the even-numbered lines of each quatrain but either 4 or G for the odd-numbered lines, thus (4/G)3(4/G)3. An example of (4/G)3(4/G)3 is given below in (9), which includes four quatrains taken from the same song. The four strongest metrical beats are marked with underlining and (for silent beats) /q/. (9) 4 3 4 3
Young Edward came to Em-i-ly His gold all for to show, q That he has made all on the la´nds, All on the lowlands low. q
G 3 4 3
Young Emily in her cha´m——ber She dreamed an awful dream; q She dreamed she saw young Edward’s blo´od Go flowing like the stream. q
G 3 G 3
O father, where’s that stra´n——ger Came here last night to dwell? q His body’s in the o´ ——cean And you no tales must tell. q
4 3 G 3
Away then to some councillor To let the deeds be known. q The jury found him guı´l——ty His trial to come on. q
(Karpeles 1932, A56A)
HM note that the purpose of this variation is almost certainly to permit a wider variety of word choice on the poet’s part. The poet’s choice of 4 versus G is based
Faithfulness and Componentiality in Metrics
119
s / and 4 for on the stress pattern of the last two syllables of the line: G for / . . . s other line endings. This dependency is illustrated by the boldface material in (9). Moreover, this pattern is the expected one, since it provides the best match of linguistic stress to rhythm grid: G provides a falling sequence to match a falling stress pattern, and 4 provides a rising sequence to match a rising one (see (5) and (7) above). The issue of how to derive the free variation in (4/G)3(4/G)3 is deferred by HM. As a stopgap, they propose ‘‘F’’ as a fifth line type, defined specifically as involving free variation between 4 and G. To this they add a Match Stress constraint, whose e¤ect is specifically to favor F. Under this arrangement, it is possible to derive quatrain types like (4/G)3(4/G)3, viewed as ‘‘F3F3,’’ simply by ranking Match Stress high enough. A more principled account would allow each of the types in {4343, G343, 43G3, G3G3} to emerge as a winner of the candidate competition under appropriate circumstances relating to the stress pattern of the line ending and hence ultimately to the poet’s choice of words. However, such a capacity is beyond the HM system, since that system evaluates only schematic representations like ‘‘4343,’’ without regard to their linguistic content. At this point, we can state the problem to be solved: to set up a grammatical system that avoids artificial constructs like F but can nevertheless derive variable quatrain types like (4/G)3(4/G)3. This will require us first to develop the formal apparatus. 6.4
Theory
To begin, it is helpful to consider what Optimality-theoretic grammars can do. 6.4.1
Defining Inventories with OT Grammars
The most familiar function of an Optimality-theoretic grammar is that of derivation; for instance, from a phonological underlying representation, we seek to derive the surface representation. In derivation, the gen function creates all conceivable surface representations, and the output is selected from among them by successively winnowing down the candidate set through a ranked set of constraints until one winner emerges. A second thing that OT grammars can do is inventory definition: the definition of a fixed (though possibly infinite) set of legal structures. The method of inventory definition described here is from Prince and Smolensky (1993) and Smolensky (1996). Let there be an additional gen, called genrb (‘‘gen of the Rich Base’’) that defines the full set (possibly infinite) of underlying representations. Submit each member of genrb to an OT grammar. When this is done, it will often be the case that distinct
120
Bruce Hayes
members of genrb will be mapped onto the same surface form. Assume further a process of collation: we remove duplicate outputs and thus collect the full set of forms that are derived as an output from at least one input. This is the inventory that the grammar defines. I will call this inventory the ‘‘output set,’’ and I will refer to an Optimality-theoretic grammar intended for defining an output set as an ‘‘inventory grammar.’’ In this view, it is not crucial for the ‘‘derivation’’ to add any new material at all. Assume in particular that genrb is su‰ciently unconstrained that it includes all possible surface representations. In this case, we can assume, as proposed by Keer and Bakovic´ 1997 and Bakovic´ and Keer 2001, that the underlying form for any candidate surface form is simply itself. We can test a form for well formedness by employing it as an input then determining whether the grammar permits it to survive into the output set.5 More precisely, to test a form F: let IF be an input form identical to F and OF be an output candidate identical to F. If OF defeats all rival candidates when IF is the underlying form, then F belongs to the output set and is legal.6 Whether OF can win the competition will depend in large degree on the ranking of the Faithfulness constraints. OF is, by definition, more Faithful to IF than any other candidate. When Faithfulness is ranked high, OF will be able defeat rival forms that perform better than OF on competing Markedness constraints. Thus, in general, inventory grammars with high-ranking Faithfulness constraints permit larger output sets (Smolensky 1996). 6.4.2
Metrics with Inventory Grammars
HM was an attempt to do metrics with an inventory grammar. However, all the constraints in their grammar were Markedness constraints, so the concepts of input forms and Faithfulness were irrelevant. To solve the problem laid out in section 6.3, we need to use inventory grammars that include Faithfulness constraints. I propose that the set of metrical quatrains, under a particular constraint ranking, should be defined as the output set for that constraint ranking. Moreover, the candidate set does not consist of schematic quatrain forms like ‘‘4343,’’ as in HM, but rather quatrains fully embodied in phonological material. To give an instance, the first quatrain in (10) can be taken to be a representative input form.7 (10)
[ [ [x [x x] Young
x x x ][ x ][x x ][x [x x][x x][x x][x x] E- mi- ly
in
her
x x x ][x [x x][x cham-
] x ] x ] x][x x] ber
line hemistichs dipods feet
Faithfulness and Componentiality in Metrics
[ [ [x [x
x x x ][ x ][x x ][x x][x x][x x][x x][x x]
She dreamed [ [ [x [x
Go
an
aw-
ful
she
saw
young
ing
like
the
x x x x ][x x [x x][x x][x x Ed-
x x x ][ x ][x x ][x x][x x][x x][x x][x x] flow-
x x x x ][x x [x x][x x][x x
] line ] hemistichs ] dipods ] feet
dream
x x x ][ x ][x x ][x x][x x][x x][x x][x x]
She dreamed [ [ [x [x
121
ward’s blood
x x x x ][x x [x x][x x][x x
stream
] line ] hemistichs ] dipods ] feet
] line ] hemistichs ] dipods ] feet
(Karpeles 1932, a56A)
Assuming that metrics is transparent, this candidate will count as well formed (i.e., metrical) if it passes the well-formedness test for inventory grammars. Specifically, if IF ¼ OF ¼ (10) and OF wins the Optimality-theoretic competition against all distinct output candidates, then (10) is predicted to be metrical. 6.4.3
Componentiality in Metrics
Before examining the candidate competition, we must add one more ingredient to the analysis: the role of components in candidate evaluation. The issue of componentiality in metrics is addressed by Kiparsky (1977), whose conception is adopted here. Kiparsky proposes that metrics is tricomponential: there is a ‘‘pattern generator,’’ which accounts for the meter; a ‘‘paraphonology,’’ which establishes the metrically relevant phonological representation; and a ‘‘comparator,’’ which evaluates the paraphonological representation against the meter to determine metricality and complexity. These three components are discussed in turn below. Pattern Generator The pattern generator for the verse described here is rather simple. In OT it can be characterized with a set of undominated constraints.
6.4.3.1
(11) a. Quatrain ¼ Couplet Couplet b. Couplet ¼ Line Line c. Line ¼ Hemistich Hemistich
122
Bruce Hayes
x d. Hemistich ¼ [x x] x e. Dipod ¼ [x x] x f. Foot ¼ [x x]
where terminals are dipod heads where terminals are foot heads where terminals are metrical positions
These constraints yield the following structure: [Quatrain [Couplet Line Line] [Couplet Line Line]], where each Line has the internal structure given above in (2). Paraphonology The paraphonology defines the phonological representations that are used in composing verse; it ‘‘constitute[s] a paralinguistic system that specifies the poetic language as a derivative of the system. . .of ordinary language’’ (Kiparsky 1977, 241). Kiparsky used a rule-based paraphonology, which included among others a paraphonological rule for John Milton’s verse that deletes stressless vowels postvocalically.
6.4.3.2
(12) Postvocalic Syncope (Kiparsky 1977, 240–241) V !q/V stress Postvocalic Syncope can derive monosyllabic [’raIt] from underlyingly disyllabic riot /’raI. t/ and trisyllabic [v .’raI.ti] from quadrisyllabic variety /v .’raI. ti/. Because this rule applies optionally, riot can be scanned in Paradise Lost (PL) as either one or two syllables (13a), and variety as either three or four (13b). e
e
e
e
(13) a. Of riot / ascends / above / thir lof- / tiest Towrs To lux- / urie / and ri- / ot, feast / and dance b. Varie- / tie with- / out end; / but of / the Tree For Earth / hath this / vari- / ety / from Heav’n
(PL 1.499) (PL 11.715) (PL 7.542) (PL 6.640)
Examples of Postvocalic Syncope are found in Shakespeare as well. As Kiparsky points out, no examples occur in Pope’s verse, which shows that the rule is poet specific. It is quite straightforward to translate paraphonological rules into Optimalitytheoretic terms. Postvocalic Syncope reduces to the free ranking of the constraint þsyllabic (McCarthy and Onset (Prince and Smolensky 1993) against Max stress Prince 1995). Where Onset dominates, the stressless vowel is dropped from riot in þsyllabic order to avoid the onsetless second syllable; where Max dominates, stress the vowel is retained. A point that will be crucial below is that, at least in English, paraphonology has only modest e¤ects: schwas are lost in hiatus, nonlow vowels become glides; but
Faithfulness and Componentiality in Metrics
123
major insertions and deletions (say, of whole syllables) are not found. References on English verse paraphonology supporting this point include Bridges (1921) and Tarlinskaja (1973). I will also stipulate that paraphonology cannot alter the stress patterns of words.8 This means that when the data show a mismatch of stress against the grid, I will be assuming that this involves a Markedness violation in the comparator component, not a Faithfulness violation in the paraphonology. As Kiparsky (1977) points out, the paraphonology is independent of the mechanisms (whatever they may be) that govern the oral performance of verse. In particular, a performer is usually free not to realize paraphonological changes, even those that are crucial to metricality. The paraphonological level of the metrical grammar is, therefore, an abstract one that serves to define the phonological representations against which metricality and complexity are computed. The last part of the componential organization that Kiparsky assumes is the comparator. This is the core of the metrical system. It consists of a set of metrical filters (essentially, constraints) that examine a phonological representation from the paraphonology and a rhythmic representation from the patterngenerating component and determine whether and how they can be matched to form a unit of metrical verse. Further principles adumbrated by Kiparsky assign differences of complexity.
6.4.3.3
Comparator
Componential Organization and the Evidence Supporting It Kiparsky o¤ers empirical arguments that metrics is organized componentially. His most crucial point is that the paraphonology always provides the same representation to every metrical constraint: for example, constraints matching stress cannot regard riot as monosyllabic, while constraints matching syllable count to rhythmic positions regard it as disyllabic. In an Optimality-theoretic account of the paraphonology, there is an additional reason the system must be componential. In OT, structural changes (e.g., vowel loss) are decoupled from their phonotactic causes (e.g., the requirement that syllables have onsets). A noncomponential theory of the paraphonology would wrongly claim that structural changes could be triggered in order to obey the requirements of the metrics. A hypothetical example of this type would be as follows. Imagine we are dealing with the verse of a poet like Milton who licenses loss of stressless vowels in postvocalic position. It is necessary under any account of iambic pentameter to assume a constraint that prevents extra syllables from cropping up in random locations in the line; let us call this constraint *Ungridded s. If the system is not componential, we would expect to find lines like (14). 6.4.3.4
124
Bruce Hayes
(14) *Vivacity without end; but of the Tree
(construct)
This would follow from the constraint ranking given in (15a), with representations as in (15b) (hi marks ungridded material). þsyllabic (15) a. *Ungridded s Max stress [metrical constraint] [paraphonological constraint] b. [ x x ] line [ x x][ x x x ] hemistichs [x x] [x x][x x][x x][x x ] feet [vI ’væsh iti] with- out end; but of the tree e
To my knowledge, such cases do not exist. Paraphonological phenomena in metrics always have authentic phonological structural descriptions. Indeed, as Kiparsky points out, they look just like ordinary language phonology and are often grounded in the fast-speech phonology of the poet’s language. A componential organization of the metrical system implies, correctly, that paraphonological processes apply only when their phonological structural descriptions are met. With the concepts of inventory grammar and componentiality in place, we can now provide a definition of metricality. The leading idea, found in earlier work such as Inkelas and Zec (1990), is that di¤erent components simultaneously evaluate the well-formedness of the same representation.9 The components are separate, not because they apply in sequence but because they evaluate di¤erent aspects of the representation. Following this approach, we can say that a quatrain of verse is metrical when the following conditions are met.
6.4.3.5
Defining Metricality in a Componential Inventory Grammar
(16) A verse form is metrical i¤ a. The metrical pattern belongs to the output set for the pattern generator (section 6.4.3.1). b. The phonological material belongs to the output set for the paraphonology (section 6.4.3.2). c. The complete representation, with phonological material aligned to the metrical pattern, belongs to the output set for the comparator (section 6.4.3.3). Componentiality guarantees the result observed in the previous section: paraphonology cannot be conditioned metrically, because the candidate set against which forms are paraphonologically evaluated consists solely of phonological representations without regard to their metrical setting. Thus, if a schwa is lost paraphonologi-
Faithfulness and Componentiality in Metrics
125
cally, it must be lost in order to avoid a hiatus, rather than to avoid a mismatch in the scansion—if the latter holds true as well, that is a felicitous result for the poet, but the components of the metrical grammar act blindly to each other’s purposes. As will become clear below, the componential approach is also crucial to explaining metricality itself. It introduces the possibility that the rival candidates that defeat an input are sometimes themselves ill formed with respect to some other component, specifically the paraphonology. This is the crucial means by which lines may be classified as unmetrical. In the sections that follow, I demonstrate how lines are classified as metrical or unmetrical in this system. Once this is done, we can return to the problem that was stated in section 6.3, the (4/G)3(4/G)3 quatrain. 6.5
Unmetricality in 4343 Quatrains
Many English folk songs (particularly ballads) are composed in quatrains of the form 4343. The odd-numbered lines in such quatrains are consistently of the type 4, and never G. Since ballads can go on for many stanzas, we can be confident that in such cases the quatrain structure is not the (4/G)3(4/G)3 discussed in section 6.3 above. A quatrain of the type 43G3, G343, or G3G3 introduced into a strict 4343 song would count as an unlicensed deviation from the established meter, that is, as unmetrical. To illustrate the ideas above, I will construct a tricomponential, Optimality-theoretic inventory grammar that permits only 4343. 6.5.1
Constraints
The bulk of the work will be done in the comparator, and all constraints mentioned should be assumed to belong to this component unless otherwise stated. Four Markedness constraints from Hayes and MacEachern (1998) are relevant. Couplets are Salient requires that a couplet consist of an internally connected, pause-demarcated unit, roughly as follows. (17) [ ] [ x x ][ x x ] [x x ][x x ][x x ][x x ] [x x ][x x ][x x ][x x ][x x ][x x ][x x ][x x ] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] [ x x] ! ! su‰ciently dense spacing of gaps syllables required here required here
couplet lines hemistichs dipods feet
This constraint is obeyed fully only by the couplets 43, G3, and 3f 3. For a precise statement of the constraint and its rationale, see HM (485–486, 492–493).
126
Bruce Hayes
Lines are Salient (HM, 483–486, 493) is the analogous constraint at the line level. (18) [ x x ] [ x x ][ x x ] [x x ][x x ][x x ][x x ] [x x][x x][x x][x x][x x][x x][x x][x x] ! ! su‰ciently dense spacing of gaps syllables required here required here
line hemistichs dipods feet
This constraint is violated gradiently in a way explained in HM (p. 493) and note 15; the optimal line type defined by this constraint is 3. Fill Strong (HM, 490) requires that the four strongest positions of the line be filled by a syllable, as shown in (19). (19) [ [ [x [x
x x x ][x x][x x][x l s
x x ][ x x x ][x x ][x x x][x x][x x][x x][x x][x x l l l s s s
] ] ] ]
line hemistichs dipods feet
It is obeyed by 4 and G lines and violated by 3 and 3f . Finally, *Lapse (HM, 490) penalizes failure to place a syllable between any two of the four strongest positions, as in (20). (20) [ [ [x [x
x x x x][x
x ] ][x x ] x][x x][x x] q q q q
x [ x x [x x ][x x [x x][x x][x x][x x q q q q q
] ] ] ]
line hemistichs dipods feet
* * * *Lapse is violated by 3 and G lines but not 3f or 4. I will also assume two Faithfulness constraints, which are stated in the language of Correspondence Theory (McCarthy and Prince 1995). (21) a. MAX(s) Assess a violation for every syllable in the underlying form that is not in correspondence with a syllable in the surface form. b. DEP(s) Assess a violation for every syllable in the surface form that is not in correspondence with a syllable in the underlying form.10
Faithfulness and Componentiality in Metrics
127
It is straightforward to determine the Max(s) and Dep(s) violations for any pair of input and output when they are classified by their line type (see (4)–(7) above). For example, any line of type 4 has an extra syllable with respect to a similar line of type G and therefore incurs (at least) one Dep(s) violation when the G line occurs at its underlying representation. The full candidate set that is input to these constraints is enormous, since it comprises all phonological representations placed in correspondence with the grid (see section 6.4.2). However, for initial purposes it su‰ces to use formulae like 4343 to designate any quatrain that would be classified as 4343. Idealized in this way, the candidate set numbers 256, which is the set of logical possibilities implied by choosing from among four line types, four times per quatrain (256 ¼ 4 4 ). 6.5.2
Ranking
The ranking needed to derive 4343 is given in (22a). As tableau (22b) shows, all but 9 of the 256 candidate types are ruled out by Couplets are Salient.11 Of these 9, only 4343 maximally satisfies both Fill Strong and *Lapse. (22) a. Couplets are Salient {Fill Strong, *Lapse} {Max(s), Dep(s), all others} Other constraints
Dep(s)
Max(s)
*
Lapse
+
Fill Strong
/4343/
Couplets are Salient
b.
[4343]
**
**
*[G343]
**
***!
*
(*)
*[43G3]
**
***!
*
(*)
*[G3G3]
**
***!*
**
(*)
*[3f 343]
***!
**
*
(*)
*[433f 3]
***!
**
*
(*)
*[G33f 3]
***!
***
**
(*)
*[3f 33f 3]
***!*
**
**
(*)
*[3f 3G3]
***!
***
**
(*)
(*)
(*)
(*)
247 candidates
*!(*)
(*)
(*)
(*)
128
Bruce Hayes
It can be seen that under this ranking, Faithfulness plays no role in determining the output; the decision is already made once we have culled candidates with the three top-ranked Markedness constraints. 6.5.3
Ruling out Unmetrical Forms
What must now be demonstrated is that this schematic analysis is e¤ective in ruling out not just schemata but actual unmetrical quatrains under the conception of metrical grammar laid out in section 6.4. Let us suppose that the anonymous folk poet is making up a new stanza for a ballad. Assume that this ballad (like hundreds of others) is composed in strict 4343 quatrains. We must demonstrate that under the grammar with the ranking of (22a), any other quatrain type would emerge as unmetrical. In particular, consider the task of ruling out *G343. To be concrete, let us suppose that the G line of this hypothetical *G343 quatrain happens to be (10) above, repeated for convenience in (23). (23)
[ [ [x [x x] Young
x x x [x
][x x][x
E- mi- ly
x x ][ x x x ][x x ][x x x][x x][x x][x x][x x][x x in
her
cham-
] ] ] ]
line hemistichs dipods feet
ber
Although this line is thus metrical in its own context (i.e., a real song composed in (4/G)3(4/G)3), it could not metrically appear in a song composed in strict 4343; this is what we want the analysis to predict. The folk poet is assumed to have (tacitly) internalized an inventory grammar of which (22a) is a partial sketch. I assert without proof that this grammar is transparent. Under this assumption, one can show that (23) is unmetrical in the given context by using the ‘‘grammaticality test’’ laid out in section 6.4.1. Specifically, one must demonstrate that when (23) is adopted as an underlying representation, there will be a rival candidate that defeats (23) in the candidate competition. In fact, there are many such candidates, one of which is shown in (24). [ [ [x [x x] Young
x x x ][ x ][x x ][x [x x][x x][x x][x x] Em- i-
ly
in
her
x x x ][x [x x][x
[’t§eIIm
b
e
(24)
] x ] x ] x][x x]
line hemistichs dipods feet
lÔ]
The reader is asked for the moment to ignore the absurdity of the word chambeler and concentrate solely on the candidate competition. Candidate (24) is an unfaithful candidate, since it possesses a syllable [b ] where the input form (23) has a null. Specifically, (24) violates Dep(s) once. However, there is also a Markedness constraint e
Faithfulness and Componentiality in Metrics
129
that is violated by (23) but not (24), namely *Lapse. Since in grammar (22a) *Lapse dominates Dep(s), (24) will emerge as more harmonic than (23). (25)
(23) Young Emily in her cham——ber +
*Lapse
(24) Young Emily in her chambeler *(23) Young Emily in her cham——ber
Dep(s) *
*!
As (25) shows, (23) is defeated in the candidate competition, despite its obvious Faithfulness virtues. It is thus excluded from the output set of the grammar defined by this constraint ranking and therefore is unmetrical in its context. The reader will have noticed that candidate (24) is itself absurd from a di¤erent point of view: nothing in the paraphonology of English folk verse licenses the extra syllable or the inserted segmental material [ l]. Here, componentiality plays a crucial role. Candidate (24) does not directly compete with (23) with regard to its phonological content; that competition unfolds within the paraphonology—where (24) most definitely loses.12 But under the componential definition of metricality in (16), the winner must belong to the output set of each component separately and, therefore, must defeat all rivals in each component separately. In the grammar under discussion, (23) fails to defeat (24) in the comparator. The fact that its phonological material wins in the paraphonology cannot rescue (23). Consequently, for the grammar of (22a) and the underlying representation (23), there is no candidate that wins in all components. For this reason, the derivation ‘‘crashes,’’ and (23) is classified as unmetrical. One might appropriately call the ‘‘chambeler’’ line (24) a ‘‘suicide candidate.’’ In a componential inventory grammar, a suicide candidate is one that defeats the maximally faithful candidate in one component while losing to the faithful candidate in a di¤erent component. Suicide candidates usually cause derivations to crash.13 Crashing derivations provide an answer to one of the questions asked in the introduction, namely, how a grammar can predict a line to be ill formed without making any claims about which more optimal form should putatively take its place. In the case of (23), there simply is no alternative line that emerges from the grammar as the appropriate ‘‘corrected’’ version: any alternative that beats (23) metrically will be paraphonologically impossible. A folk poet dissatisfied with the unmetrical (23) must think up a di¤erent line, and the grammar does not tell her what to compose.14 e
6.6
Deriving Free Variation and Marked Winners
Having illustrated the apparatus of Faithfulness and componentiality, I will now return to the problem laid out in section 6.3, that is, of deriving the (4/G)3(4/G)3
130
Bruce Hayes
quatrain. The grammar needed for (4/G)3(4/G)3 turns out to be rather similar to that for strict 4343, except that the Faithfulness constraints Max(s) and Dep(s) are ranked higher. This permits a larger output set, which encompasses the free variants. The specific ranking needed is the one given in (26). This ranking di¤ers from the 4343 ranking in (22a) in that Max(s) and Dep(s) outrank *Lapse. (26) Couplets are Salient Fill Strong {Max(s), Dep(s)} {*Lapse, remaining constraints} A tableau for the variant 43G3 is given in (27). Lines are Salient is explained in section 6.5.1 above, and Long-last is explained in HM (488–490, 493); they are included here merely to show that there are Markedness constraints in the system that would have favored di¤erent outcomes had they been ranked higher. *Lapse
Lines are Salient
Long-Last
110 15
*
**
200
*
****
20
*
*
***
110
*
**
101
***
11
*
**
101
*
**
*
***
11
*
***!*
**
*
**
2
*
(*)
(*)
(*)
(*)
(*)
(*)
[43G3]
**
*[4343]
**
*[G3G3]
**
*!
*[G343]
**
*!
*[3f 343]
***!
*
*[3f 3G3]
***!
*
*[433f 3]
***!
*
*[G33f 3]
***!
*[3f 33f 3] 247 candidates
*!(*)
Dep(s)
***
Max(s)
+
Fill Strong
/43G3/
Couplets are Salient
(27)
*!
Grammar (26) culls the rival candidates in a way similar to the 4343 grammar (22a): Couplets are Salient removes all but 9 of the 256 candidate types, and Fill Strong removes 5 more. Thus, at this point we know that the output cannot contain any quatrains other than the four targets; what is at issue is whether all four will make it into the output set.
Faithfulness and Componentiality in Metrics
131
Grammar (26) di¤ers from grammar (22a) in that whereas (22a) places the Faithfulness constraints at the bottom of the hierarchy, (26) places them just below Fill Strong, so that they are active in selecting from among the surviving four candidates. When the underlying representation is a 43G3 quatrain, any rival candidates of the types 4343, G3G3, or G343 will incur Faithfulness violations. The 43G3 form, being totally faithful to itself, thus emerges as a winner and qualifies as a member of the output set. Not surprisingly, the three other types embodied in the formula (4/G)3(4/G)3 also emerge as part of the output set of (26), since each is free of Faithfulness violations when it is selected as the underlying form. Tableaux showing how each one beats out its three best rivals are given below. *Lapse
Lines are Salient
Long-Last
200
*
*!
***
110
*
**
**!
****
20
*
**
*!
***
110
***
110
**
200
*
****
20
*
***
110
*
****
20
*
**
*[43G3]
**
*[G3G3] *[G343]
Dep(s)
b.
**
[4343]
Max(s)
+
Fill Strong
/43G3/
Couplets are Salient
(28) a.
/G343/ +
c.
[G343]
**
*[4343]
**
*[G3G3]
**
*!
*[43G3]
**
*!
*!
*
/G3G3/ +
[G3G3]
**
*[4343]
**
*!*
**
200
*
*[43G3]
**
*!
***
110
*
*[G343]
**
*!
***
110
132
6.6.1
Bruce Hayes
Distribution of 43G3
The 43G3 quatrain type has an interesting property: although there are songs composed in ‘‘strict’’ (nonvarying) 4343, strict G343, and strict G3G3 (see, HM 478– 481), to my knowledge there are no songs composed in strict 43G3. 43G3 occurs only as a free variant in songs that also allow 4343, G343, and G3G3 in other stanzas. This is predicted by the system given here. There exist rankings (see HM, 495) that permit only 4343, only G343, and only G3G3 to survive into the output set. For these rankings, the output set is culled down to a single quatrain type by Markedness constraints placed at the top of the ranking. But for 43G3 to survive into the output set, the Faithfulness constraints Dep(s) and Max(s) must be ranked relatively high. When such a ranking holds, the other three quatrain types will also be allowed into the output set, since the Faithfulness constraints will not penalize them when they are selected as the underlying representation. Note that since 43G3 cannot be selected by Markedness constraints alone, it is a kind of ‘‘marked winner’’ in the sense defined in section 6.1. The example of this section thus illustrates how Faithfulness constraints can be used in metrics to derive marked winners. This concludes the analysis of free variation in quatrain structure. The crucial idea has been that when Faithfulness is ranked higher in the grammar, a variety of underlying forms are able to defeat all their rivals and emerge as outputs of the inventory grammar. In this way, the analysis is able to derive free variation, without (as in HM) stipulating constraints that actively require it. 6.7
Lexical Inversion
As a second illustration of Faithfulness and componentiality in metrics, I will discuss a problem that was addressed but not fully solved by Hayes and Kaun (1996; hereafter HK). I define a ‘‘lexical inversion,’’ following earlier work, as a configuration in which the syllables of a simplex polysyllabic word with falling stress are placed in a metrical position that calls for rising stress. Here is an example, highlighted in bold. (29) [ [ [x [x He
x x x x][x
][x x][x
pulled
rings
x x x][x o¤
x ][ x ][x x ][x x][x x][x x][x of
his
fin-
x x x][x
] ] ] x]
line hemistichs dipods feet
gers (Karpeles 1974, 7G)
Faithfulness and Componentiality in Metrics
133
HK noticed that lexical inversions in folk verse have an asymmetrical distribution, which is strikingly di¤erent from what occurs in iambic pentameter. For folk verse, the great majority of inversions occur at the end of a line, as in (30).16 (30) a. b. c. d. e. f. g.
Who should ride by but Knight William I’ll bet you twenty pound, master I fear she will be taken by some proud young enemy There lived an old lady in the north country And two of your father’s best horses But he had more mind of the fair women Lived in the west country q
(Karpeles 1974, 27A) (Karpeles 1974, 7F) (Karpeles 1974, 45A) (Karpeles 1932, 5B) (Karpeles 1932, 5B) (Ritchie 1965, 36) (Karpeles 1974, 43E)
Most of these are of line type 4, as in (30a–f ), with a few cases of 3, as in (30g). A further asymmetry that HK note is that in quatrain types where 4 or 3 occur in free variation with G or 3f , lexical inversion is quite unusual. Thus, songs in strict 4343 include lexical inversions far more often than songs in (4/G)3(4/G)3 quatrains.17 HK propose an intuitive explanation for these facts, which I will here employ as the basis of an OT analysis. There are three points at issue: (i) why lexical inversion is disfavored in general, (ii) why it has a special privilege of occurring at the end of the line, and (iii) why this privilege should be so rarely exercised in line positions that permit (4/G) free variation. 6.7.1
Ruling out Inversion in General
For the first point, HK observe that unlike pentameter, folk verse virtually always leaves a certain number of grid positions unfilled. Therefore, when the syllables of a line are such that an inversion might arise, it is usually the case that a minor shift in the location of the syllables would make inversion unnecessary. For example, instead of producing the inversion in (31a) (mismatched stress and grid columns shown in boldface), the folk poet can sidestep the problem simply by moving William over a bit, as in (31b). (31) a. [ [ [x [x x]
x x x ][ x ][x x ][x [x x][x x][x x][x x]
x x x x ][x x [x x][x x][x x
*Wilb. [ [ [x [x x]
liam he was a x x x ][ x ][x x ][x [x x][x x][x x][x x]
noble knight (construct) x ] line x x ] hemistichs x ][x x ] dipods [ x x ] [ x x ] [ x x ] feet
Wil-
liam
he
was a
no-
ble
knight
] ] ] ]
line hemistichs dipods feet
(construct)
134
Bruce Hayes
To formalize this idea, we need to state two constraints: a Markedness constraint that bans lexical inversion and the Faithfulness constraint violated by (31b) when (31a) is the underlying representation. This is the subject of the next two sections. 6.7.1.1 MATCH STRESS In formulating a constraint to exclude lexical inversions, we are on well-explored territory. Kiparsky (1975b), Bjorklund (1978), and other scholars have shown that poets and poetic traditions often require a particularly strict match to the meter for sequences of stressed and unstressed syllables that fall within a single simplex word. Let us assume such a constraint here.18
(32) MATCH STRESS Assess a violation if
si and sj (in either order) are linked to grid positions Gi and Gj , respectively; si has stronger stress than sj ; Gj is stronger than Gi ; and si and sj occupy the same simplex word.
In a full grammar, there would be other constraints requiring stress matching in other contexts as well; but for present purposes, (32) will su‰ce. IDENT(location) We must also formulate the Faithfulness constraint that is violated when, for example, (31b) is taken to be a candidate surface form for underlying (31a). Here, the contents of grid and phonological representation are identical, but the temporal association of syllables and grid marks is di¤erent. I will assume that this violates the Faithfulness constraint (33).
6.7.1.2
(33) IDENT(location) If si is linked to grid position G in the input and to grid position G 0 in the output candidate, assess a number of violations equal to the distance in grid positions between G and G 0 . To make (33) explicit, we need to say how violations are assessed when more than one syllable is shifted over. Various possibilities exist; since nothing matters here in how this issue is resolved, I will assume for concreteness that the violations are simply summed. Thus (31b), taken as a surface candidate for underlying (31a), incurs nine violations of Ident(location): two for Wil-, two for -liam, two for he, two for was, and one for a, as shown below.
Faithfulness and Componentiality in Metrics
(34) [ [ [x [x x]
135
x x x ][ x ][x x ][x [ :x x ] [ ::x x ] [ ::x x ] [ ::x x ]
2 ::::::::: :2:::::::::: 2 :::::::: 2::::::::::: 1:::::: ::::::: ::: :::: ::::::: ::::
*Wil-
liam Wil-
he liam
was he
x x x x ][x x [x x][x x][x x
a nowas a no-
ble ble
knight knight
] ] ] ]
line hemistichs dipods feet (31a) (31b)
6.7.1.3 Ruling out Nonfinal Lexical Inversion The goal is to construct an analysis in which (31), with a lexical inversion in nonfinal position, is unmetrical. Assume that (31a) is the underlying form and that (31b) is a rival candidate. If Match Stress outranks Ident(location), (31b) will be the winner, as (35) shows.
(35) (31a) William he was a noble knight
Match Stress
(31b) William he was a noble knight *(31a) William he was a noble knight
Ident (location) *
*!
Because it is beaten by (31b), (31a) is excluded from the output set and is unmetrical. There is independent reason to think that the ranking Match Stress Ident(location) will prevail, because Ident(location) is generally a weak constraint and Match Stress a strong one. Here is the evidence for these two claims. The experimental data gathered by HK indicate that folk song lines are composed in a way such that the grid locations of the syllables are relatively predictable. In particular, HK’s consultants, given only text and grid, showed fair agreement among themselves as to what the proper alignment of syllables to grid should be. They could not have achieved this unless the song texts gave them clues as to where to locate the syllables. This implies that where syllables go in the grid is, to a fair degree, noncontrastive information. In Optimality-theoretic terms, noncontrastive structural information is that which is protected by low-ranking Faithfulness constraints; hence, Ident(location) must be ranked low. On the other hand, Match Stress is expected, based on our general knowledge, to be ranked rather high. There are poetic traditions (e.g., classical German and Russian verse) in which it is undominated; and even where Match Stress is violated there are usually strict limitations on where the violations may occur (Kiparsky 1977). Moreover, quite a few folk songs have no lexical inversions at all, suggesting that they are composed under a ranking in which Match Stress is undominated (see section 6.7.3 below). If rankings are relatively constrained even across di¤erent metrical forms, we expect Match Stress to be ranked relatively high in general. It
136
Bruce Hayes
would certainly be expected to outrank a characteristically feeble constraint like Ident(location). The upshot is that in the general case, candidates that match stress by ‘‘sliding’’ the syllables will be favored over candidates that mismatch stress. This provides an across-the-board pressure against lexical inversion. 6.7.2
Ruling Out Nonfinal Inversion
To explain why inversion can occur at the end of the line (in certain quatrain types), HK note that in this location, additional constraints are active. These constraints rule out all of the available ‘‘slid over’’ candidates, leaving lexical inversion behind as the best remaining option.19 The way this works can be seen if we ponder what kind of syllable sliding in principle could rescue the lexical inversion in (36). (36) [ [ [x [x
x x x x][x
Fair
][x x][x
El- li- nor
x x x][x
][ ][x x][x
she
x x x x][x
][x x][x
was a gay
la-
x x x][x
] ] ] x]
line hemistichs dipods feet
dy (Karpeles 1974, 15I)
In order to avoid a violation of Match Stress, the crucial stressed syllable la- must migrate to a stronger grid position than mismatched -dy. The migration in principle could be to the left (where gay sits in (36)) or to the right (the location of -dy in (36)). For each of these two possibilities, there are two reasonable20 possibilities for where to put -dy, making a total of four, shown in (37). The dotted arrows show where lahas been ‘‘moved’’ in each of the four possibilities.
El- li- nor
Fair Fair Fair Fair -dy . . .
ElElElEl-
lililili-
she
nor she was a nor she was a
nor nor
she she
was a gay gay was was
gay
:::: X
Fair
][x x][x la-
::::
x ][ ][x x ][x x][x x][x x][x
x x x x][x
x x x][x
lalaa a
dy gay gay
] ] ] x]
line hemistichs dipods feet
dy
original, with inversion
dy
slide la- left
: X :: X :::: ::::: ::::::::: ::::::::: ::::::::: ::::: :::: :: : : : :::: ::: X
a. b. c. d.
x x x x] [ x
::::
(37) [ [ [x [x
la- dy la-
slide la-, -dy left slide la-, -dy right slide la- right; -dy into next line
Faithfulness and Componentiality in Metrics
137
Below, we consider the four possibilities in turn and show that under appropriate constraint rankings, they are excluded. The lexical inversion setting remains as the most viable option. Sliding to G Example (37a), repeated below as (38), is a candidate in which Match Stress is obeyed by moving the penult of lady into the third strong position of the line, crowding some of the other syllables to fit it in. The syllable -dy is kept in the fourth position, so a G line results. 6.7.2.1
(38) [ [ [x [x
x x x x][x
Fair
x x ][ x ][x x ][x x x][x x][x x][x x][x
El- li- nor she was a
gay
x x x][x
][x x][x
la-
] ] ] x]
line hemistichs dipods feet
dy
This candidate will fail to defeat the lexical inversion candidate (36), provided that *Lapse, the constraint that forbids G, is ranked above Match Stress. (39) (36) Fair Ellinor she was a gay lady +
*Lapse
Match Stress
(36) Fair Ellinor she was a gay lady
Ident (location)
*
*(38) Fair Ellinor she was a gay la——dy
*!
9 *’s
As we will see in section 6.7.4 below, this ranking will necessarily hold, for independent reasons, in the quatrain types that allow lexical inversion. 6.7.2.2 Sliding to 3f Example (37b), repeated as (40), is similar to (37a): la- is again placed in the third strong position of the line, but in this case -dy is also slid over, so that a 3f line results.
(40) [ [ [x [x x] Fair
x x x ][ x ][x x ][x [x x][x x][x x][x x] El- li- nor she was a
gay
x x x [x la-
][x x][x
x x x][x
] ] ] x]
line hemistichs dipods feet
dy
If Fill Strong, which forbids 3 and 3f lines, is ranked above Match Stress, then this candidate will also fail to defeat the lexical inversion candidate (36).
138
Bruce Hayes
(41)
Fill Strong
(36) Fair Ellinor she was a gay lady +
(36) Fair Ellinor she was a gay lady
Match Stress
Ident (location)
*
*(40) Fair Ellinor she was a gay lady q
*!
11 *’s
6.7.2.3 The Fast-Syllable Candidate Example (37c), repeated as (42), is a candidate in which la- has been moved rightward to the fourth strong position of the line; -dy occupies the weak terminal position.
(42) [ [ [x [x Fair
x x x x][x
][x x][x
El- li- nor
x ][ x ][x x][x x][x she
x x x ][x x][x x][x
was
a
] x ] x ] x][x x]
gay
line hemistichs dipods feet
la- dy
Impressionistically, the e¤ect is of an uncomfortably fast rendition of lady at the end of the line, suggesting unmetricality. In fact, lines like this are quite rare in real verse. Moreover, in an experiment conducted by HK in which 670 lines of folk verse were chanted from the written text by 10 speakers of English, the consultants fairly generally avoided this kind of rendition. HK suggest that this line type is ill formed because it involves a gross mismatch of the sung duration of the final syllables versus their natural duration. (For discussion of the evidence that supports duration matching in folk verse, see HK section 6.1.) In the present case, because of the e¤ects of phrase-final lengthening (Wightman et al. 1992) and the concurrence of line and intonation phrase boundaries, the line-final syllable is normally quite long. It is therefore ill fitted to fill a single grid slot. I will therefore assume a constraint to be called Match Duration that penalizes intonational phrase-final syllables21 that are squeezed into the final grid slot of the line. I assume further that Match Duration is quite highly ranked and in particular that it outranks Match Stress. Therefore, the ‘‘fast syllable’’ candidate (42) (shown here iconically with condensed type) must lose out to the inverted stress candidate (36). (43) (36) Fair Ellinor she was a gay lady +
Match Duration
(36) Fair Ellinor she was a gay lady *(42) Fair Ellinor she was a gay lady
Match Stress
Ident (location)
* *!
6 *’s
Faithfulness and Componentiality in Metrics
139
6.7.2.4 The Overflow Candidate The fourth and last reasonable possibility for placing lady in metrically matched position was (37d), repeated below as (44).
(44) [ [ [x [x
x x x x][x
Fair [ [ [x [x
x x ][ x x ][x x ][x x ][x x x][x x][x x][x x][x x][x x][x x
El- li- nor x x x x][x
she
was
a
gay
] ] ] ]
line hemistichs dipods feet
] ] ] ]
line hemistichs dipods feet
la-
x x ][ x x ][x x ][x x ][x x x][x x][x x][x x][x x][x x][x x
dy . . . Here, the second syllable of lady spills over into the grid of the next line. This is unusual in folk verse and arguably is so because it violates general principles of alignment for phonological phrasing and metrical constituents. Evidence in support of such alignment principles is given in Kiparsky (1975b), Hayes (1989b), Hayes and MacEachern (1996), and HK section 6.2. I assume that such lines violate Alignment constraints (McCarthy and Prince 1993a), which in this case require line breaks to coincide with major phonological phrase breaks. For concreteness, we will assume that the relevant type of phrase break is the Intonational Phrase, though in a full grammar additional constraints would be needed that refer to higher and lower prosodic domains as well (Selkirk 1980; Nespor and Vogel 1986; Pierrehumbert and Beckman 1988). In McCarthy and Prince’s system, the relevant constraint is Align(Line, L, Intonational Phrase, L): ‘‘the left edge of every Line must coincide with the left edge of an Intonational Phrase.’’ Align(Line, L, Intonational Phrase, L) is a characteristically strong constraint in folk verse, and I assume it generally outranks Match Stress. Under this ranking, candidate (44) must lose out to the lexical inversion candidate. (45) (36) Fair Ellinor she was a gay lady +
Align
(36) Fair Ellinor she was a gay lady *(44) Fair Ellinor she was a gay la-]Line [dy
Match Stress
Ident (location)
* *!
7 *’s
140
Bruce Hayes
To summarize, candidates with lexical inversion will win when the inversion is in final position and lose when the inversion is in other positions, if the following rankings hold. (46) {*Lapse, Fill Strong, Match Duration, Align(Line, L, Intonational Phrase, L)} Match Stress Ident(location) The bottommost ranking of Ident(location) means that nonfinal inversion is prevented, because a shifted candidate will defeat any input with nonfinal inversion. Final inversion is possible because Match Stress is ranked below a group of constraints that collectively prevent the victory of any shifted candidates. As HK note, this explanation relies crucially on the notion of constraint conflict that lies at the heart of OT. 6.7.3
Songs Where Inversion Is Unmetrical
Under the ranking of (46), any line-final inversion comes out as acceptable on a faute de mieux basis. Further data, however, suggest that this pattern is not always found but rather holds true only for certain verse forms. In folk verse, many songs include no lexical inversions at all,22 and it is reasonable to suppose that such songs are composed under a ranking that classifies inversion candidates as ill formed. I posit that this ranking is Match Stress Max(s). Under this ranking, a lexical inversion, even in final position, is defeated by a suicide candidate in which the stressless syllable is lost. One suicide candidate of this type is given in (47). (47)
[ [ [x [x
x x x x][x
*Fair
x x ][ x x ][x x ][x x ][x x x][x x][x x][x x][x x][x x][x x
El- li- nor
she
was
a
gay
] line ] hemistichs ] dipods ] feet
[leId]
The following tableau shows how the suicide candidate defeats the inversion candidate. (48)
(36) Fair Ellinor she was a gay lady +
Other constraints
Match Stress
(47) Fair Ellinor she was a gay [leI d ] *(36) Fair Ellinor she was a gay lady
Max (s) *
*!
Ident (location)
Faithfulness and Componentiality in Metrics
141
As before, the suicide candidate does not embody a metrical line of verse, since it cannot win in the paraphonology: nothing in English metrical paraphonology permits arbitrary dropping of whole syllables. In contrast, in the verse type discussed in the preceding section, where line-final inversion is allowed, Max(s) must dominate Match Stress. Under this ranking, (36) would be the winner; it would therefore belong to the output set and count as metrical.23 This result may be related to the discussion above in section 6.1. In their original account, HK assumed a naı¨ve OT approach in which the best candidate always emerges as well formed. In verse varieties that avoid lexical inversion, this assumption turns out to be wrong. The more articulated version of OT used here, incorporating Faithfulness and componentiality, is able to make the correct prediction of outright unmetricality. 6.7.4
Linking Inversion to Quatrain Type
It remains to account for one more of HK’s observations: that inversion is unusual in the odd-numbered lines of (4/G)3(4/G)3 quatrains. The argument works as follows. (1) Consider any pair of lines L4 and LG that have the same text but di¤er in that L4 is a 4 line with a final lexical inversion and LG is a G line. For example, L4 could be (36) and LG could be (38). L4 and LG di¤er in their crucial Markedness violations: L4 violates Match Stress (and LG does not); LG violates *Lapse (and L4 does not). Moreover, since L4 and LG have the same text, there will be no Faithfulness violations other than Ident(location), when L4 is taken as a candidate surface form for underlying /LG / or vice versa. In particular, although Max(s) outranks Match Stress in any verse that allows lexical inversion, it cannot a¤ect the outcome here, since both competing candidates obey it. (2) By hypothesis, the quatrain type is (4/G)3(4/G)3. Therefore, underlying G lines in the first and third lines must be able to defeat all alternative settings. For LG , this includes the rival candidate L4 . Since L4 violates only the feeble Ident(location) among the Faithfulness constraints, LG must defeat L4 on the basis of Markedness. Given the Markedness constraints that L4 and LG violate, it follows that Match Stress must dominate *Lapse. (3) Now consider what happens when L4 is the underlying form. Given what has just been said, L4 cannot survive into the output set because LG will defeat it. Specifically, the Faithfulness constraint Ident(location) is too weak to save L4 , and the Markedness constraints Match Stress and *Lapse have just been shown to be ranked in a way that causes LG to defeat L4 . The argument is summarized in tableau (49).
142
Bruce Hayes
(49) /L4 /
Max(s)
Match Stress
+ LG L4
Lapse
Ident (location)
*
*
*!
The result is that if the constraints are ranked in a way that permits G lines to occur in free variation with 4 lines, then lexical inversion candidates cannot make it into the output set and are thus unmetrical. The only exception will be when Match Stress and *Lapse are specially designated to be freely ranked. I assume that the relatively few cases where lexical inversion occurs in the odd lines of a (4/G)3(4/G)3 quatrain fall under this heading. However, even in this circumstance, there will be a complexity penalty for both inversion and G lines, for reasons discussed in the next section. 6.8
Metrical Complexity
In the final section, I will extend the Faithfulness/componentiality proposal so that it can account for metrical complexity. For earlier work on complexity, see Halle and Keyser (1966, 1971a), Kiparsky (1975b, 1977), Youmans (1989), and Golston (1998). I adopt from Youmans’s work the position that metrical complexity should be analyzed in the same terms as metricality; namely, that absolute metricality and unmetricality are only the ends of a continuum. One reason to believe this is that when one examines a variety of poets and traditions, complexity turns out to respond to the same factors that govern metricality. For instance, the coincidence of a prosodic break with the post-fourth position hemistich break in iambic pentameter is a strong normative tendency for many English poets (Oras 1960). For the verse of George Gascoigne, however, or French pentameter (the ‘‘decasyllabe’’), it is obligatory. Such cases are easily multiplied. The shared basis of metricality and complexity has a natural interpretation under OT: the traditions and poets di¤er, not in constraints that guide verse composition but only in their ranking. If this view is correct, then what is needed for analyzing gradient well-formedness is a conception under which ranking is a gradient phenomenon. For this purpose, I adopt the apparatus developed in HM, Hayes (2000), and Boersma and Hayes (2001). The version of Boersma and Hayes is quantitatively explicit and will be employed here. In this model, constraints are assigned ‘‘ranking values’’ on a continuous numerical scale. Grammars are stochastic, in that at any one application of the grammar,
Faithfulness and Componentiality in Metrics
143
the values employed for constraint strictness are determined at random. This is done by selecting a point for each constraint from a normal probability distribution, centered on its ranking value. Grammars of this type can generate a range of outcomes, with di¤erent probabilities a‰liated with each outcome depending on the ranking values of the constraints. However, such grammars can also generate outcomes that are essentially categorical; this occurs when the ranking values of the relevant constraints are extremely far apart.24 A further assumption of the model is that, at least in the crucial class of cases, gradient well-formedness can be treated in terms of probability (Boersma and Hayes 2001; Frisch and Zawaydeh 2001): forms that could be derived only under a somewhat unlikely choice of selection points are assumed to be somewhat ill formed, forms that could be derived only under a highly unlikely choice of selection points are assumed to be almost entirely ill formed, and so on. This model of gradient well-formedness has been tested by Boersma and Hayes against data on English /l/, taken from Hayes (2000), and against data involving the nasal mutation process of Tagalog by Zuraw (2000). In both cases, the model achieves a good match against scalar well-formedness ratings gathered from a panel of native speakers. In the present case, we need to adapt the stochastic apparatus to inventory grammars, which designate whether a representation is or is not in an output set. Adapting the probability-based strategy just described, I posit that the appropriate definition of complexity is as follows. (50) The ‘‘metrical complexity’’ of a line (couplet, etc.) is the probability that the constraints of a stochastic OT grammar will be ranked in a way that excludes it from the output set. On this scale, the complexity of a line or other structure varies from zero (under all possible rankings of the grammar, the line will be allowed in the output set) to one (there is no possible ranking that allows it in the output set). Zero is equivalent to perfect metricality, and one to total unmetricality. Obviously, (50) is a theory-internal definition of complexity. Complexity is also a term that is defined empirically, relating to the gradient intuitions people have about verse structures. My hypothesis is that with appropriate constraints and rankings, the probability-based theoretical values defined by (50) can be mapped onto human intuitive judgments by some monotonic function. Plainly, extensive research is needed to test this claim. However, definition (50) can already be seen to have three advantages. First, it is quantitatively explicit. Second, it is compatible with the criterion set above for an adequate theory of complexity, that is, that metricality and unmetricality should be characterized as extremes on the complexity continuum. Finally, under
144
Bruce Hayes
this approach, the same constraint inventory can be used to characterize both metricality and complexity. 6.8.1
The Complexity of Inversion
A good example of metrical complexity in folk verse is lexical inversion, analyzed in section 6.7. I think most listeners share a sense that lines with inversion are complex; certainly it has attracted the attention of scholars who have examined folk verse (Hendren 1936, 137; Karpeles 1973, 24). Experimental evidence also supports the complexity of inversion: the consultants for Hayes and Kaun (1996), asked to chant texts that in the original song included an inversion, often responded with alternative noninverted settings but seldom did the reverse.25 I will now attempt to characterize inversion quantitatively as complex. Specifically, I will develop a grammar in which lines containing lexical inversion will emerge with a complexity value no lower than .8. This value is arbitrary, being unanchored in experimental data; the point here is to show that the grammatical apparatus is capable of providing such values. I will assume that the quatrain type under examination is always of the type 4343. This quatrain type can be guaranteed by the (essentially) strict rankings given below. (51)
The units along the strictness scale are as defined in Boersma and Hayes (2001). The value 13.49 is chosen as that which results in a one-in-a-million probability against a reversed ranking for the arrows shown; hence, these rankings are essentially obligatory.26 The ranking arguments for (51) are as follows: Couplets are Salient must outrank Fill Strong if 4343 is to defeat 4G4G or any other quatrain type that fills the last strong position of a couplet; Couplets are Salient must outrank *Lapse if 4343 is to defeat 43f 43f ; Fill Strong must outrank Lines are Salient if 4343 is to defeat 3f 33f 3; and *Lapse must outrank Lines are Salient if 4343 is to defeat G3G3. A second, independent group of rankings is the following (their relative placement along the scale with respect to the constraints of (51) would not matter).
Faithfulness and Componentiality in Metrics
145
(52)
The (essentially) categorical rankings Match Duration Match Stress, Align Match Stress, and Match Stress Ident(location) are defended above in sections 6.7.2.3, 6.7.2.4, and 6.7.1.3, respectively. The crucially gradient ranking is Match Stress Max(s). When the math is done, it emerges that with the di¤erence of 2.38 in ranking values shown, there is an 80% probability that Match Stress will dominate Max(s) at any given evaluation time. As we saw in section 6.7.3, when Match Stress dominates Max(s), inverted lines are defeated in the candidate competition by suicide candidates that remove the final unstressed syllable of the line. Therefore, under the gradient ranking of (52) there is an 80% probability that a candidate with final inversion will not make it into the output set. The metrical complexity of lines with inversion (all else in the line being perfect) is thus .8, which is what we sought originally to describe. 6.8.2
Complexity in General
This analytic strategy can be extended as a treatment of metrical complexity in general. Suppose we want to characterize the complexity of a given metrical structure S in the grammar. We locate first a Markedness constraint M violated by lines (quatrains, etc.) containing S. We also locate a distinct structure S 0 , such that lines containing S 0 instead of S obey M but violate a Faithfulness constraint F and moreover are paraphonologically illegal. Under these circumstances (all else being equal), lines containing S will be unmetrical if M outranks F by a wide margin. metrical but complex if M and F have relatively close ranking values. The degree of complexity will depend on the size and direction of the di¤erence. fully metrical if F outranks M by a wide margin.
These rankings determine the likelihood of whether lines containing S can survive the competition with a suicide candidate that contains S 0 instead of S. In the case discussed in section 6.8.1, S was the mismatched lady in (36), S 0 was the corresponding material ([leId]) in (47), M was Match Stress, and F was Max(s).
146
6.9
Bruce Hayes
Conclusion
OT appears to have major potential advantages as an approach to metrics. It achieves explanatory force by letting the ‘‘ingredients’’ of metrical grammars be general, typologically motivated constraints, with idiosyncrasy resulting from genre- or tradition-specific rankings. Moreover, existing case studies, such as the analysis of quatrains in Hayes and MacEachern (1998) or of lexical inversion in Hayes and Kaun (1996), indicate that the data patterns seen in metrics really do reflect constraint conflict, which is the central idea of OT. The goal of this paper has been to make possible a full-fledged OT metrics by resolving the problems that were noted in the introduction. The crucial ingredients have been inventory grammars (section 6.4.1), in which surface forms are derived from identical underlying forms faithfulness constraints (section 6.4.1) components (section 6.4.3) that independently evaluate di¤erent aspects of the same representation stochastic grammars and the probabilistic definition of metricality (section 6.8)
In the data examined here, these proposals have su‰ced to explain the existence of marked outputs in metrics (section 6.6), to account for free variation (section 6.6), to permit forms to be ruled out without suggesting an alternative (sections 6.5.3, 6.7.3), and to make explicit predictions about complexity (section 6.8.1). It remains a much larger task to assess the validity of these proposals for other areas of metrics and of linguistics. Notes Thanks to Chris Golston, Patricia Keating, Gerhard Ja¨ger, Elliott Moreton, Paul Smolensky, Donca Steriade, Colin Wilson, the volume reviewers, and the participants in a winter 2000 UCLA seminar class for helpful input in the preparation of this chapter. They are absolved, in the usual way, of responsibility for defects. 1. For work on the ‘‘missing remedy’’ problem elsewhere in linguistics, see Prince and Smolensky (1993, 47–51, 175–178), Orgun and Sprouse (1999), Ra¤elsiefen (1999), and To¨rkenczy (2002). 2. Readers seeking help in interpreting the gridded examples may download chanted versions of them (in .wav format) from http://www.linguistics.ucla.edu/people/hayes/Faithfulness InMetrics/. Example (3) is rendered in musical notation in Hayes and MacEachern (1998, 475). 3. In traditional metrics, a ‘‘feminine ending’’ is one in which the penultimate syllable of the line bears stress and the final syllable is unstressed. Most 3f lines do indeed have this stress pattern in their final two syllables.
Faithfulness and Componentiality in Metrics
147
4. There is a slight complication: it is necessary to stipulate that certain constraints are undominated, so the actual set of predicted outputs is smaller than the full factorial typology. 5. For this to work, the inventory grammar must be transparent, in the sense of Kiparsky (1971b). In an opaque grammar, there are often outputs that are legal but cannot be derived from themselves. The introduction of ‘‘crashing’’ derivations below, in which some inputs yield no output, does not alter the situation with regard to testing the grammaticality of a form; it remains the case that an entity will be well formed only if it is mapped onto itself by the grammar. 6. The idea that underlying representations can be completely as rich as surface representations is proposed earlier in Inkelas (1995). 7. To give an expository simplification, phonological representations are depicted as orthography, without stress or phonological phrasing. 8. Empirically, this stipulation has massive support: if paraphonology could alter stress patterns, then words could be mismatched against the meter in all contexts; in actual fact, such mismatches are tightly constrained in a way that requires a metrical rather than a paraphonological analysis. For details, see section 6.7. 9. For other conceptions of componentiality in OT, see Pesetsky (1997, 1998), Ja¨ger and Blutner (1999), Blutner (2000), and Wilson (2001a). 10. As a reviewer points out, Max(s) and Dep(s) seem to play little role in phonology (McCarthy and Prince 1999) and might be better replaced with Max(V) and Dep(V). However, since syllables are the central elements counted in metrics I will assume that Max(s) and Dep(s) are possible metrical, if not phonological, constraints. 11. There are nine because there are only three couplets that fully obey Couplets are Salient (43, G3, and 3f 3), and each may occur in either couplet location. 12. Specifically, since*[’tSeImb l] includes segments [ ] and [l], which are absent in its lexical representation /’tSeImb/, it violates the paraphonological Faithfulness constraints Dep( ) and Dep(l). It also obeys no constraints that I can imagine that are not also obeyed by [’tSeImb]. Therefore, any quatrain including *[’tSeImb l] is always paraphonologically defeated by a candidate containing [’tSeImb], no matter how the constraints of the paraphonology are ranked. Substituting a real word like featherbed for chambeler does not help, since the paraphonology must construe [’fD "bd], not as the word featherbed but as a candidate surface representation for underlying /’tSeImb l/; it plainly cannot win. e
e
e
e
e
13. The derivation does not crash when a third candidate exists that wins in all components. Such cases nevertheless cause the input to be designated as unmetrical. 14. One might speculate, however, that examining the winning candidate(s) of the comparator could give the poet a hint: ‘‘I need a line ending that sounds like . . . [’tSeIm s b]; perhaps featherbed?’’ 15. The rather obscure-looking numbers for constraint violations in this column implement a scheme laid out in Prince and Smolensky (1993, section 5.1.2.1), which permits a single constraint to handle di¤erent degrees of violation (there is a multivalued scale of line saliency) in multiple locations (there are four lines in a quatrain). I have also tried the alternative approach, of setting up multiple constraints each defining a cuto¤ point on the scale, and found that it also leads to a working analysis. For details of the numerical scheme, see HM 493 and http://www.linguistics.ucla.edu/people/hayes/quattabl.htm.
148
Bruce Hayes
16. A spreadsheet containing randomly collected examples of lexical inversion from Anglo-American folk song is posted at http://www.linguistics.ucla.edu/people/hayes/ FaithfulnessInMetrics/. Of the 64 inversions whose songs use the grid of (2), 58 (¼ 90.6%) are in line-final position. The exceptions mostly fall under the categories discussed in HK, 291– 294. 17. Three cases of lexical inversion in (4/G)3(4/G)3 I have noticed are Karpeles 1932, 3B; Karpeles 1974, 60G; and Karpeles 1974, 72. 18. A constraint with this name is assumed in Hayes and MacEachern (1998), but there it takes the rather artificial form ‘‘Prefer lines of type F.’’ Constraint (32) is by contrast well supported elsewhere in metrics. 19. This passage is alarmingly reminiscent of section 6.1 above in which it is argued that OT’s principle of always outputting the best candidate can be a problem in metrics. As it turns out, it is possible to eat one’s cake and have it too. Depending on the rankings, derivations can either yield winners faute de mieux or crash with no output. 20. By ‘‘reasonable’’ I mean ‘‘without gratuitous violations of other constraints.’’ 21. And, possibly, other very long syllables as well; the issue is not crucial here. 22. Some examples (all 4343) are Karpeles 1932, 31D, 42B, 97A, 186A; Karpeles 1974, 12A, 18G, 95C, 130A, 141B. 23. A further detail is that in verse types where lexical inversion is to be metrical line finally, we must also rule out suicide candidates that replace mismatched feminine endings with nonfeminine endings by inserting a syllable, as in Fair Ellinor she was a gay lady O. This will follow if Dep(s), which rules out such candidates, likewise dominates Match Stress. This assumption also holds for section 6.7.4 below. 24. I use the word essentially because the relevant type of grammar will generate certain forms with extremely low probability. If this probability is low enough, say, one in a million, then these rare outcomes could not be distinguished empirically from performance errors and thus could not sensibly be counted as wrong predictions. 25. In 74/170 cases, consultants presented with the text of lines that contained line-final inversions in the original responded with a noninverted setting. In 790 cases, consultants were presented the text of a line that had a feminine ending but was not inverted in the original; of these, they replied with an inverted rendering only 26 times. Thus the ‘‘uninversion’’ rate was 43.5 percent, whereas the ‘‘spontaneous inversion’’ rate was only 3.3 percent. 26. A tutorial on how probabilities are derived from ranking value di¤erences may be found in Zuraw (2000).
II
Phonology and Morphology
7
The Phonology of Perceptibility Effects: The P-Map and Its Consequences for Constraint Organization
Donca Steriade
7.1
Introduction
This chapter outlines a proposed revision in the structure of Optimality-Theoretic (OT) phonologies (Prince and Smolensky 1993). The proposal is to let a distinct grammatical component, which I call the P-map (P for ‘perceptibility’), project correspondence constraints and determine their ranking. The P-map is a set of statements about relative perceptibility of di¤erent contrasts, across the di¤erent contexts where they might occur. For instance, the P-map will be the repository of the speaker’s knowledge that the [p]-[b] contrast is better perceived before V’s (e.g., in [apa] versus [aba]) than before C’s (e.g., in [apta] versus [abta]). The point of departure here is the theory of correspondence set forth in McCarthy and Prince 1995, with its distinction between Max/Dep constraints, which identify the elements of two representations that stand in correspondence, and the Ident F constraints, which require a precise featural match between correspondent elements. The general rationale for the P-map proposal is that attested phonological systems display less diversity than predicted by versions of OT in which correspondence and phonotactic constraints interact freely. In particular, the range of pairings between constraint violation and ‘‘repair strategy’’ is more limited than current versions of OT would lead one to expect. An example of this need for a tighter fit between predictions and typology involves the e¤ect that constraints on obstruent voicing have on phonological systems. Consider a common constraint like (1), an underlying string like /tæb/, which violates (1), and the range of possible responses of the grammatical system to this violation, as sketched in (2). (1) A phonotactic constraint: *[þVOICE]/ ]word Voiced obstruents do not occur at the end of the word.1
152
Donca Steriade
(2) Conceivable grammatical responses to the violation of (1) in underlying representation /tæb/2 Change in underlying representation, to satisfy (1) a. Devoicing: /tæb/ ! [tæp] b. Nasalization: /tæb/ ! [tæm] c. Lenition to glide: /tæb/ ! [tæw]
Corresponding constraint ranking *[þvoice]/__]word *[þvoice]/__]word *[þvoice]/__]word [Gconsonantal] *[þvoice]/__]word *[þvoice]/__]word *[þvoice]/__]word (segments) *[þvoice]/__]word features)
d. C-Deletion: /tæb/ ! [tæ] e. V-Insertion: /tæb/ ! [tæb ] f. Segment reversal: /tæb/ ! [bæt] e
g. Feature reversal: /tæb/ ! [dæp]
Ident [Gvoice] Ident [Gnasal] Ident Max C Dep V Linearity Linearity (for
The table in (2) should be read on the understanding that the correspondence constraint named in a given cell is the lowest ranked among the correspondence constraints that are in potential conflict with the phonotactic. Of the changes in (2), only the devoicing in (2a) is actually attested as a reaction to *[þvoice]/ ]word violations. This is not surprising if one consults one’s linguistic intuition, but it is unexpected in the context of OT, in its present form: if the ranking between the correspondence constraints in (2b–g) is free, one expects at least the range of fixes shown in (2). My claim is not that nasalization, C-deletion, and so forth are unattested processes, but that they are unattested as responses to the voicing problem posed by (1). This means that one does not encounter sound systems in which all the final voiced stops, and only those stops, turn to nasals, delete, or trigger epenthesis or metathesis. (3) indicates what systems of alternations would look like if some of these changes did occur.
tun,
ii. Word final:
tat,
top,
taÐ,
$
$
$
$ tim,
tek
tu,
tat,
ti,
top,
tu,
$
$
$
$
$
tud-a, tat-a, tib-a, top-a, tug-a, tek-a $
ii. Word final: b. Deletion of final voiced obstruents i. Before V:
$
$
(3) Unattested systems (lexically related forms linked by arrows) a. Nasalization of final voiced obstruents i. Morpheme shapes before V: tud-a, tat-a, tib-a, top-a, tag-a, tek-a
tek
The Phonology of Perceptibility Effects
153
tib , top,
$
$
$
$
$
tud , tat,
e
e
ii. Word final:
tug , tek e
$
c. Epenthesis after final voiced obstruents i. Before V: tud-a, tat-a, tib-a, top-a, tug-a, tek-a
The diagnosis for the problem encountered—the fact that devoicing is the only available cure for violations of (1)—starts with the observation that of all the inputoutput pairs displayed in (2), the one judged most similar is the pair [tæb]-[tæp] in (2a). (Evidence for the relevant hierarchy of similarity is reviewed in section 7.4.) The aim, in any departure from the underlying representation (UR), is to change it minimally to achieve compliance with the phonotactics. The modifications in (2b–g) are less minimal, because they result in greater input-output dissimilarity, than devoicing. This is why they are systematically avoided. If this is the correct diagnosis, then what is needed is a mechanism that relates rankings between correspondence constraints to perceived di¤erences in degree of similarity. I refer to knowledge of relative phonological similarity as the P-map. The function of the P-map I discuss here is that of guiding the speaker in search of the minimal input deformation that solves a phonotactic problem. A set of similarity rankings will also be needed in lexical access: the listener has to pick which lexical entry most closely resembles a frequently ambiguous auditory input.3 The grammatical reflex of the P-map is the projection and ranking of correspondence constraints. Thus, if the P-map identifies the pair [p]-[b] as more similar in the context V ] than [b]-[m] for the same context, then the P-map’s e¤ect on the grammar will be to rank higher the faithfulness condition corresponding to the more distinctive contrast [b][m], hence Ident [Gnasal]/V ] Ident [Gvoice]/V ]. This idea is outlined in (4) using the same example as an illustration. (4) P-map e¤ects on the ranking of correspondence conditions P-map comparisons
More distinctive contrast (e.g., [b]-[m] in V__]
Ranking of correspondence constraints
Higher-ranked constraint (e.g., Ident [Gnasal]/V__]
vs.
Less distinctive contrast [b]-[p] in V__])
Lower-ranked constraint Ident [Gvoice]/V__])
The reader will note the parallel with Prince and Smolensky’s (1993) treatment of phonetic scales: the grammatical reflex of the physical scale is the fixed ranking between constraints referring to points on the scale. Consider now present OT. The concept of minimal modification embodied in this theory is the candidate that optimizes satisfaction of correspondence constraints, as ranked in a given grammar. No independent principle determines the ranking of
154
Donca Steriade
potentially conflicting correspondence conditions. This means, in the context of the example in (2), that either [tæp] or [tæm] will count as minimal modifications of the input /tæb/, depending only on the unconstrained ranking between Ident [Gnasal] and Ident [Gvoice]. I assume below that the phonotactic (1) induces some input modification: the question is which. (5) Devoicing as minimal modification /tæb/
Ident [Gnasal]
Ident [Gvoice]
+ tæp tæm
* *!
(6) Nasalization as minimal modification /tæb/ tæp
Ident [Gvoice]
Ident [Gnasal]
*!
+ tæm
*
The problem with this view is that for at least some phonological properties, and perhaps for all, there exists a crosslinguistically constant notion of minimal modification; that is why a violation of (1) is only resolved by (2a) and not in other ways. This study is a contribution to our understanding of this notion. The di‰culty outlined in (2)—which I call ‘‘Too-Many-Solutions’’—arises with particular clarity in Optimality Theory. This is because OT views phonology as a problem-solving system: the problem is the conflict between phonotactic constraints and lexical forms that violate them, such as the UR /tæb/ of (2). The Too-ManySolutions conundrum arises when the system of constraints and rankings predicts too many resolutions of a given phonotactic problem. But the same di‰culty comes up in any other approach to phonology in which changes in the underlying form are seen as the sound system’s responses to phonotactic violation.4 Thus Kisseberth’s (1970) insight that conspiracies arise when the sound system aims at a specific target structure via multiple means leads to equivalent questions in the context of rulebased phonology: if phonologies aim to eliminate final voiced obstruents, why don’t they employ obstruent nasalization, deletion, metathesis? 7.2
Sources of Evidence for the P-Map
The P-map hypothesis is that speakers possess judgments of relative similarity of the form in (7) and use these to determine a partial ranking of correspondence constraints.
The Phonology of Perceptibility Effects
155
(7) The pair of strings x-y is less similar than the pair w-z (abbreviated as D(x-y) > D(w-z), where D ¼ di¤erence). If some of these judgments are crosslinguistically invariant, then the ranking of certain correspondence conditions will be invariant too. The invariant correspondence ranking will give rise to the typological laws suggested earlier. We observed that if a final voiced obstruent must be avoided, then the repair is to devoice it rather than nasalize it. The P-map conjecture is that the source of this law is a similarity ranking: D(oral C-nasal C) > D(voiced C-voiceless C). This similarity ranking induces a correspondence ranking, Ident [Gvoice] Ident [Gnasal]. If the similarity ranking is constant, the correspondence ranking is constant too—with qualifications discussed below—and this fact explains the law. Empirical tests of the P-Map hypothesis can proceed by observing invariant preferences for one repair strategy over conceivable others. Some of these observations are discussed below and in earlier work (Fleischhacker 1999, 2005; Jun 1995; Steriade 2001). What must be verified then is that each invariant preference corresponds to a constant judgment of relative similarity. Conversely, one can start from documented judgments of relative similarity and verify the P-map’s prediction that these judgments correlate with preferences for certain input modifications as against others. Both strategies are currently hampered by the scarcity of direct evidence on relative similarity. Few studies address this question and the vast majority compare di¤erent contrasts in the same position (say ba-pa versus with ba-ma5) rather than the same contrasts across positions (e.g., ba versus pa compared to ab versus ap). However, an additional hypothesis about the nature of similarity judgments expands the range of evidence bearing on the P-map. This second hypothesis is that phonological similarity is evaluated on perceptual representations (cf. Flemming 1995/20026). This means that the representations inspected in evaluating similarity are those that encode the perceptual correlates of a contrast—the auditory properties that di¤erentiate its terms. These properties vary with context. The voicing contrast, for instance, is conveyed in part by VOT di¤erences but these are unavailable when the C’s occur in final position. We can reason then that a voicing contrast like b versus p will give rise to di¤erent similarity judgments depending on context: the pair aba versus apa will register as more dissimilar than ab versus ap if one of the voicing correlates is missing word-finally. More generally, the hypothesis that phonological similarity is perceptual in nature turns evidence about cue distribution into a source of information about similarity. As a result, confusion-rate data (Miller and Nicely 1955; Wang and Bilger 1973) is potentially relevant to the P-map hypothesis: if the pair of auditory stimuli x-y give rise to a higher rate of confusion than the pair z-w, that may be because their auditory representations are less well di¤erentiated. This is at least one factor in the
156
Donca Steriade
phonological similarity judgment D(x-y) > D(z-w). Confusability and similarity are distinct matters: subjects rating similarity of auditory stimuli do not report confusion, just similarity. But if similarity is evaluated on perceptual representations, then confusability and similarity are related, because they share at least one source: some auditory representations are di¤erentiated by fewer or less salient properties than others. This makes them more similar and, in the limit, more confusable. In substantiating a claim of relative similarity one can rely on speakers’ direct judgments of similarity, or refer to confusion studies indicating that one contrast is more perceptually robust than another, or reason from the observation that in the position considered, one contrast misses an acoustic correlate and the other does not. Finally, there is the evidence of linguistic tasks that imply a judgment of identity or near-identity: rhyming and alliteration. On the former, see Zwicky 1976, Hanson 2003, as well as Steriade and Zhang 2001. On the latter, see Fleischhacker 2005 and Minkova 2002. All forms of evidence are employed below. Beyond these observations, no model of similarity computation is o¤ered here. (See Frisch, Broe, and Pierrehumbert 2004 for a model of context-free similarity evaluated on articulatory representations.) Rather, the strategy here is to observe that subjects hold certain judgments of relative similarity—whatever their source may be—and that the preference for certain repair strategies correlates with these judgments. This study focuses on only some of the predictions of the P-map hypothesis, as they a¤ect voicing neutralization, cluster resolution, epenthesis, and V-deletion. The aim is to show in each case that preferred methods of resolving phonotactic violations exist; that these preferences are not accounted for by currently available mechanisms; that each preference for a particular solution is explained by the idea that the least perceptually distinctive contrast whose modification removes the violation is always the one sacrificed; and that a solution can be obtained by ranking correspondence constraints via the P-map. 7.3
The P-Map
The P-map is a mental representation of the degree of distinctiveness of contrasts in various positions. It can be viewed as a set of statements about relative similarity between sounds or other phonological properties. The example in (8) involves a hypothetical set of similarity rankings about voicing distinctiveness in di¤erent contexts. (8) D(p-b/V
V) > D(p-b/
V) > D(p-b/V
) > D(p-b/C
C)
(8) states that p and b are less similar intervocalically than prevocalically, even more similar postvocalically, and maximally similar between C’s. The statement can be generalized to a broader class, such as all obstruent pairs that di¤er only in voicing,
The Phonology of Perceptibility Effects
157
if similarity rankings do not di¤er across class members. The P-map fragment in (8) embodies the hypothesis that voicing di¤erences are equally perceptible for all pairs of obstruents but not equally perceptible across positions, the optimal context for this being the intervocalic position (Steriade 1999a). Several properties of P-map statements like (8) are critical here. As already indicated, they must reflect the e¤ect of the syntagmatic context on perceived similarity. Voicing contrasts are not equally well perceived in all positions, and this has fundamental e¤ects on the phonology of voicing. Likewise, the distinctiveness of a contrast between any specific segment class and its absence (here Ø) is also a¤ected by context. For instance, pairs like [fIts]-[fIs]—with a t-Ø di¤erence in the V C context— register as less similar than pairs like [fIst]-[fIs]—with the same t-Ø di¤erence in the C ] context.7 A corresponding P-map statement appears in (9). (9) D(t-Ø/V
C) > D(t-Ø/C
])
Relative distinctiveness judgments can involve suprasegmental or serial position di¤erences. Thus (10) records a hypothetical similarity ranking: C1 C2 reversal under adjacency is less distinctive than the reversal of C’s separated by a V. (10) D(C1 VC2 -C2 VC1 ) > D(C1 C2 -C2 C1 ) According to (10), a pair like apsa-aspa is more similar than the pair apas-asap. The point here is not to examine the truth of or rationale for (10) but to illustrate how the terms of similarity comparisons may involve strings larger than one segment.8 While relative similarity judgments can be documented among many string pairs, one must also consider that no knowledge of relative similarity is available in other cases. Another hypothetical comparison illustrates this: Is the pair task-tass more or less similar than task-task[ ]? Suppose that the P-map fails to record a di¤erence in this case: then the P-map cannot be the source of any ranking between correspondence constraints prohibiting either type of modification (a variant of Max C in task-tass; a variant of Dep V in task-task[ ]). This does not mean that the two constraints cannot be ranked, but if crosslinguistically invariant correspondence rankings emerge only from crosslinguistically invariant similarity rankings, then the Max C versus Dep V ranking will be free to vary on a language-specific basis. The final point to emphasize about the P-map is that distinctiveness and its opposite, similarity, are properties of contrasts (Flemming 1995 [2002]): the statement ‘‘a is more perceptible than b’’ means ‘‘a is more reliably distinguished from a reference term x than b is distinguished from x.’’ It is not the sounds or the articulations a and b that are being compared for perceptibility but the contrasts a versus x and b versus x. This point is fundamental to the success of the P-map as an analytic tool and follows from the assumption that the P-map is so structured as to permit a definition of e
e
158
Donca Steriade
the concept of minimal modification. Consider what information is needed to discover the minimal modification that will render a representation like /tab/ compatible with a constraint like *[þvoice]/ ]word . Since several modifications of this input achieve compatibility with this constraint, the P-map must indicate which of these represents the minimal modification. For example, if we compare [tap] and [tab ] as potential modifications of input /tab/, the necessary comparison involves the inputoutput pairs /tab/-[tap] and /tab/-[tab ]. If there is a guide to the minimal modification, this guide must exist in the form of statements about the relative perceptibility of contrasts like these, or their generalized forms. The contrast is that between the unchanged input string and its modified output correspondent, as it occurs in the context of the modification. There is another sense of contrast and another sense of perceptibility, distinct from the one used here. Suppose that there are invariant properties that underlie sound classes—either invariant acoustic properties or articulatory gestures common to all manifestations of a given class. Then we can talk about the fact that some context allows a better recovery of these invariants. For instance, suppose that the invariant properties of [b] may be better recovered intervocalically than interconsonantally. What that means is that we can better distinguish [b] in V V from all other sounds that could have occurred there. This is the broad sense of contrast. This may be a useful notion but not for the purpose of defining the minimal modification: it does not tell us which pair—/tab/-[tap] or /tab/-[tab ]—is the most similar input-output pair. e
e
e
7.4 7.4.1
A P-Map Account of Voicing Neutralization Differences of Relative Similarity
As a preliminary to the P-map analysis of devoicing—our answer to one aspect of the Too-Many-Solutions problem—I now outline the evidence for a hierarchy of perceived similarity between the pairs in (11). Each pair corresponds to the contrast between an input string with a voiced obstruent in final position and its modified counterpart. (11) a. b. c. d. e. f.
D vs. T/V D vs. N/V D vs. G/V C vs. Ø/V Ø vs. V/C C1 VD2 ] vs.
]word ]word ]word ]word ]word D2 VC1 ]word
D ¼ voiced stop, T ¼ voiceless stop N ¼ nasal G ¼ glide or lateral
The present task is to show that, among these, the voicing contrast D vs. T/V ]word is least distinctive—that is, that its terms are perceived as more similar
The Phonology of Perceptibility Effects
159
than those of the other pairs. In fact, the voicing contrast (11a) stands out because it is the only one that lacks its primary perceptual correlate: the VOT value. This is one reason to expect the (11a) pair to be considered most similar. There are, however, no studies that compare overt similarity judgments for the relevant five pairs in (11) (ab, a-c, a-d, a-e, a-f ). This gap can be filled by combining rhyming studies, studies of foreign-accent perception, and similarity studies for CV pairs, where C quality is systematically varied. Regarding the latter, if the study of CV sequences shows that voicing pairs (e.g., [ba]-[pa]) are more similar than oral-nasal pairs (e.g., [ba]-[ma]), then we can reason that the same result will obtain a fortiori for the VC pairs, since the voicing contrast is, if anything, further attenuated in VC sequences. Voicing versus Manner Studies of imperfect rhyming provide a direct comparison between the final voicing contrast D vs. T/V ]word and the manner contrasts D vs. N/V ]word and D vs. G/V ]word . Zwicky (1976) analyzes 236 half-rhymes in which a consonantal mismatch is ignored, out of a corpus of rock-and-roll rhymed texts with a total of 700 half-rhymes. Relevant here is that next to the 18 instances of a voicing mismatch (pairs like died-light or wise-price) there are only 5 comparable cases where nasality or obstruency di¤erences are discounted (i.e., mid-sin). Hanson (2003) studies slant rhyme in the poetry of Robert Pinsky: here V’s di¤er freely in rhyming pairs, while final C’s stand under a violable requirement of identity. She notes that of the 128 imperfect slant rhymes in her corpus, 96 percent di¤er only in voicing (e.g., woes-loss). Pinsky is not an isolated case. Hanson discusses similar data in Pope and Yeats; for the latter, out of a total of 66 rhyming pairs containing a difference in the final C, 94 percent involve a voicing di¤erence. No rhymes are cited where nasality or laterality is ignored. The rhyming results are supported by the studies of similarity comparing CV sets (Walden and Montgomery 1975) or isolated C sets (van den Broecke 1976). The first of these studies identifies four dimensions of contrast: sibilant versus nonsibilant, sonorant versus obstruent, stop versus nonstop, and, to a much lesser extent, the [p]-[t]-[k] place contrast. Voicing was not a global contrast factor and the overall similarity between voicing cognates (e.g., [pa]-[ba]) emerges as much greater than that between oral/nasal or continuant/noncontinuant pairs. Van den Broecke’s study records Dutch subjects’ impressions of similarity between single isolated C’s uttered silently, and here too di¤erences based on nasality and sonority emerge as dominant. Conversely, pairs judged to possess the highest degree of similarity are pairs of similar sonority, most of them [p]-[b]-type pairs. Greenberg and Jenkins (1964) report similar results in one of their experiments, where subjects were asked to list associates of nonsense stimuli like [klæb]. For all forms that, like [klæb], could yield a lexical item through a change of the final C’s voicing, the most common responses involved voicing changes. Thus for [klæb], the most commonly mentioned forms were [klæp] 7.4.1.2
160
Donca Steriade
(23/46 responses) and hands (a clear associate of clap: 12/46 responses). Significantly, there were other potential associates that also di¤er by exactly one feature from the stimulus: for [klæb] a minority associate is [klæm] (11/46). The one feature di¤erentiating the stimulus [klæb] from the [klæm] response is nasality. Apparently, however, nasality is more significant than a di¤erence in obstruent voicing, since [klæm] was much less frequently elicited than [klæp]. This brief review indicates that voicing is, in any context, perceived as less distinctive than contrasts based on obstruency di¤erences, and moreover that this weak voicing contrast is being suppressed—by final devoicing—in one position where it is least distinctive to begin with. This supports the proposal that devoicing is preferred to nasalization, gliding, or lateralization as a means of complying with the voicing constraint (1), because devoicing induces the smallest input-output dissimilarity. Next I consider evidence on the distinctiveness of voicing as compared with the C-Ø contrast. The aim here is to suggest that dropping the C, to avoid violating (1), is a more salient departure from the input than simply devoicing it. To this end, we could note that the C-Ø contrast involves multiple dimensions of di¤erence (because C and Ø di¤er in voicing, labiality, obstruency), whereas the voicing contrast involves just one of these dimensions. This relates to theories of correspondence in which feature values stand in correspondence and their mismatch is penalized by Max and Dep constraints (Casali 1997; Lombardi 1998). In these theories, discussed below, the loss of a segment necessarily violates a superset of the constraints violated by simple featural modification. However, I claim that the perception of similarity does not reduce to counting features. We thus look for independent support for rankings like D(b-p) < D(b-Ø). Fleischhacker (1999) solicited from English speakers relative similarity judgments between a target word and a modification of it. Some modifications involved changes of final obstruent voicing, while others involved C-loss, metathesis, or V-insertion. The results relevant to us were of the type in (12). 7.4.1.3
Voicing versus the C-Ø Contrast
(12) Voicing versus C-Ø similarity di¤erences: Fleischhacker (1999) Reference term
More similar to
Than to
print
prind
prin, prit
Fleishhacker also tested possible correlations between, on the one hand, greater perceived similarity between target and modified form and, on the other hand, greater preference for one modification than another. She did this by ensuring that, in some of the sets compared, the more similar form was also phonotactically disfavored. Thus prind is judged more similar to print that prin or prit, but it is phonotactically disfavored relative to these, both because it violates (1) and because it
The Phonology of Perceptibility Effects
161
contains a complex coda. Despite the phonotactic improvement, the preference test correlates with the similarity test, suggesting that preference for a given modification is first and foremost a function of its similarity to the source, and only secondarily a matter of phonotactic well-formedness. I consider now the e¤ects of precedence, bearing in mind that the voicing constraint (1) could be satisfied by displacing the C: [tæb] ! [bæt]. Data presented in Fleishhacker 1999 allows us to indirectly compare these e¤ects to those induced by a voicing di¤erence. In the absence of a direct comparison between voicing and serial position contrasts, one can rely on the assumption that similarity is a transitive relation. Thus if C-Ø di¤erences are more distinctive than di¤erences of voicing, and if serial position di¤erences (D(C1 VC2 C2 VC1 )) are more distinctive than C-Ø di¤erences, then by transitivity, serial position is more contrastive than voicing. Fleischhacker’s (1999) study shows that precedence modifications are judged more significant than either coda or onset C-deletion. 7.4.1.4
Voicing vs. Precedence Relative to the V
(13) Fleischhacker (1999) C-Ø contrast versus precedence relative to V Reference term
More similar to
Than to
flip gulf
fip guf
filp gluf
Since C-Ø contrasts are more distinctive than voicing, we infer that contrasts involving position relative to the V are more distinctive than voicing. Feature Transfer Available similarity data does not bear on the possibility of single-feature transfer as an alternative to final devoicing—that is, /tæb/ ! [dæp] as against /tæb/ ! [tæp]. Here however it is safe to reason without data: whatever the dissimilarity degree of /tæb/ versus [tæp] might be, that of /tæb/ versus [dæp] will be greater, since two C’s modify their voicing value in the case of featural metathesis, as against only one in the case of devoicing. This case appears irrelevant to the discussion of standard correspondence theory: single-feature movement of the /tæb/ ! [dæp] sort will violate twice Ident[Gvoice], whereas mere devoicing will violate it only once. In this case the correct preference appears to be built into the existing system. However, the variant of correspondence theory that adopts Max [aF] constraints—instead of or in addition to Ident [GF] constraints (Casali 1997; Lombardi 1998)—will allow the [dæp] candidate to emerge as the minimal modification of the input /tæb/, under rankings like *[þvoice]/ ], Max [þvoice] Linearity [voice]. 7.4.1.5
162
Donca Steriade
(14) Feature reversal in a Max [aF] theory /tæb/
*[þvoice]/__]
tæp
Max [voice]
Linearity [voice]
*!
*
+ dæp
**
The aim here is not to argue against Max [aF] constraints, but to point out that without recourse to a theory of perceived similarity their mere existence causes candidates to emerge that need further weeding out. However, since standard correspondence theory accidentally avoids this issue, I consider in what follows only full segment reversal as an option that needs to be explicitly excluded. The last case discussed is the possibility of removing violations of (1) through epenthesis. The question involves the relative distinctiveness of the [tæb]-[tæp] contrast as against [tæb]-[tæb ]. I am discussing only one choice of epenthetic V, [ ], because any other V will likely represent an even more salient departure from the original. On this, see below. We can directly compare for distinctiveness the voicing and Ø-[ ] contrasts on the basis of data reported by Magen (1998), who sought to determine which features of the Spanish accent are most noticeable to English speakers. Among the more common aspects of Spanish-accented English are schwa insertion (as in [ spik] for speak and [kloz d] for closed ), deletion of final sibilants (as in stand for stands), and the modified realization of the voicing contrast: medial [z] realized as [s] and initial voiceless stops realized without aspiration and perceived as voiced. Magen asked her English subjects to rate for native quality the original, Spanish-accented utterances as well as edited versions of these originals, in which specific manifestations of the Spanish accent had been targeted and changed, so that the utterances would acquire nativelike quality in those specific respects. In this way, one can observe how English speakers rated the V-Ø di¤erence between the original and edited version of forms with epenthesis (e.g., Spanish-accented [kloz d] versus modified [klozd]) and compare this with the rating di¤erence between the original and edited version of forms with voicing changes (e.g., Spanish-accented [ris n] reason versus edited [riz n]). The relevant results are that voicing changes did not elicit statistically significant rating di¤erences; in contrast, C-deletion and epenthetic schwa significantly altered the ratings and in fact ranked as the most noticeable di¤erences observed. We reach the same conclusion about the relative salience of voicing versus V-Ø in a di¤erent way, on the basis of Fleischhacker’s (1999) study, supplemented with results of earlier work done on English and Swedish by Wingstedt and Schulman (1988). These researchers did not directly compare devoicing and epenthesis, but V-Ø vs. Voicing
e
e
e
e
e
e
e
e
7.4.1.6
The Phonology of Perceptibility Effects
163
rather epenthesis and C-deletion. Wingstedt and Schulman’s subjects rated Cdeletion outputs as preferable to the epenthesis outputs: the relevant triplets were in this case three modifications of a base form like conduct: conduc versus condu[k t] versus condut. e
(15) Preference judgments in Wingstedt and Schulman 1988: Final C-Ø versus Ø-V Reference term
Best modification
Worse
Worst
conduct
conduc
condu[k ]t
condut
e
Fleischhacker’s results allow us to compare a di¤erent version of the same question, because she inserted the V after the last C. She also distinguished similarity and preference ratings and was thus able to show that these ratings correlated. (16) Fleischhacker (1999) similarity judgments: Final C-Ø contrast versus Ø-V Reference term
More similar to
Than to
Than to
heft
hef
heft
het
e
(17) Fleischhacker (1999) preference judgments: Final C-Ø contrast versus Ø-V Reference term
Best modification
Worse
Worst
heft
hef
heft
het
e
In this context, Wingstedt and Schulman’s preference data becomes relevant to the issue of similarity. Recall that Fleischhacker had compared the e¤ects of devoicing with those of C-deletion and had verified that forms related via C-deletion ( printprin) are perceived as more dissimilar to the base relative to forms related via voicing ( print-prind ). Reasoning again from the assumption of transitivity, it follows that devoicing will be less distinctive a departure from input than V-insertion. From this I deduce that devoicing is preferred to epenthesis because it is a less salient modification of the input. As noted above, Magen’s study leads to the same conclusion. This exhausts all the alternatives to devoicing considered in (2). 7.4.2
The Analysis
The discussion of relative similarity has yielded the dissimilarity hierarchy in (18). (18) A hierarchy of distinctiveness in contrasts D(C1 VC2 -C2 VC1 ), D([ ]-[Ø]) > D(C-Ø) > D([Gson]/
]) > D([Gvoice]/
])
e
Relevant to the discussion of final devoicing is only the fact that the word-final voicing contrast emerges as less distinctive than other contrasts considered. Next I
164
Donca Steriade
show that this fact alone resolves the Too-Many-Solutions problem as applied to final devoicing. The solution anticipated earlier is that correspondence constraints are ranked as a function of the relative distinctiveness of the contrasts they refer to. Since it is the Pmap that contains information on distinctiveness, the analysis must establish a link between correspondence constraints and P-map statements. There are two aspects of this process. First, if the P-map encodes a similarity ranking between two contrasts— D(x-y) > D(z-w)—each of the contrasts must map onto a distinct correspondence constraint, Corresp (x-y) and Corresp (z-w). Otherwise relative similarity rankings will exist that fail to be reflected in the structure of the correspondence system. I formulate this requirement below. It amounts to the claim that the dimensions and degrees of similarity di¤erentiated by the system of correspondence are projected from the P-map. (19) P-map projects correspondence constraints Let D(x-y) Ki stand for the perceptual di¤erence between members of sound classes x and y in context Ki . If D(x-y) Ki > D(w-z)/ Kj , then there exist distinct sets of correspondence conditions, Corresp (x-y/ Ki ) and Corresp (w-z/ Kj ).
Corresp (x-y/ Ki ) is an abbreviated reference to constraints prohibiting the correspondence of a member of class x in one representation, to a member of class y in a related representation, when both x and y occur in context K. The basic requirement on the relation between the P-map and the correspondence system is that the more distinctive contrasts are protected by higher-ranked correspondence conditions. (20) Ranking correspondence constraints by relative distinctiveness If D(x-y)/ Ki > D(w-z)/ Kj , then any correspondence constraint referring to D(x-y)/ Ki outranks any parallel constraint referring to D(w-z)/ Kj . The term parallel constraints refers to constraints that link the same pair of representations: input-to-output and varieties of output-to-output correspondence (baseto-reduplicant, una‰xed base-to-a‰xed base, and so on). Thus, if D(Gnasal)/ K > D(Gvoice)/ K, then (20) requires that Ident [Gnasal]/ K I-O Ident [Gvoice] K I-O. However, Ident [Gnasal]/ K I-O may or may not outrank Ident [Gvoice]/ K Base-Derivative, because these two constraints do not link the same pair of representations and thus are not parallel constraints. I now clarify the reference in (20) to ‘‘any correspondence constraint referring to D(x-y)/ Ki .’’ Ident [GF] constraints refer to P-map di¤erences of the form D(x-y)/ K, where x and y are distinct values of the same feature. In contrast, Max and Dep constraints refer to P-map di¤erences of the form D(x-Ø)/ K, where x is
The Phonology of Perceptibility Effects
165
a phonological property (e.g., a segment) and K is a context where x occurs in one representation and its absence occurs in a related representation. Thus Max segment/ K could be paraphrased as: ‘‘There is no D(Ø-C) between I and O, such that I contains C in K and O contains Ø in K 0 and K corresponds to K 0 .’’ Regarding Linearity, it is possible to view the di¤erence D(AxB-ABx), where A and B are context strings and x is a segment, as the sum of the di¤erences D(x-Ø)/ A B and D(x-Ø)/AB . If so, then under the P-map hypothesis, Linearity refers to the same type of P-map di¤erences as Max and Dep, and thus Linearity reduces to conjunctions of context-sensitive Max and Dep constraints. This suggestion has great potential benefits as well as problematic aspects. The observation of phonological similarity judgments implicit in rhyming practices suggests that modifications of linear order induce much greater perceived di¤erences between strings than featural mismatches or segment loss. Rhyming pairs involving metathetic mismatches like task-fax are virtually absent from several large corpora of half-rhymes, in clear contrast to subsequence half-rhymes like desk-mess (Steriade and Zhang 2001; Steriade 2003). This observation is explained if metathesis induces a superset of the D(x-Ø) di¤erences caused by deletion or insertion: cumulative di¤erences between rhyming lines create worse, hence less frequent half-rhymes. A related benefit of this proposal is its ability to explain the extreme rarity of metathesis as a solution to phonotactic violation.9 The possible drawback is that, by the same token, if Linearity reduces to Max and Dep conjunctions and all Max and Dep constraints are context sensitive, it becomes di‰cult to characterize any systems that prefer metathesis over deletion as a solution to phonotactic violation. In what follows I continue to refer to metathetic di¤erences as D(AxB-ABx) and to Linearity as an independent class of constraints. The problem noted is left unresolved here. Returning now to the analysis of final devoicing, we observe that from the principle in (19) and the distinctiveness hierarchy in (18) it follows that each contrast distinguished by (18) gives rise to a distinct set of correspondence conditions. From (20) it follows that correspondence conditions extracted from (18) are ranked by distinctiveness as in (21). (21) Ranking of I-O correspondence constraints by the distinctiveness scale (18) Linearity (C1 VC2 vs. C2 VC1 ), Dep ( vs. Ø)/C ] Max (C vs. Ø)/V ] Ident [Gson]/V ] Ident [Gvoice]/V ] e
Recall now that the correspondence constraints in (21) are the only ones whose violation could in principle satisfy the *[þvoice]/ ] constraint, for inputs that violate it. Our starting point was the observation that each of the five constraints in (21) can, in the present version of correspondence theory, be ranked lower than the others, thus predicting at least five distinct solutions to violations of (1). The constant ranking in (21)—derived from the constant similarity relations in (18)—eliminates
166
Donca Steriade
this di‰culty. Although there are in principle six di¤erent ways of ranking *[þvoice]/ ] relative to members of the correspondence hierarchy in (21), only two sets will yield distinct e¤ects: one set contains the five rankings in which *[þvoice]/ ] Ident [Gvoice]/V ], all of which amount to final devoicing; the other set contains the ranking in which Ident [Gvoice]/V ] *[þvoice]/ ]. Only two distinct outcomes are thus predicted: violate the phonotactic or apply final devoicing. This is the result we were aiming to derive. 7.4.3
Consider the Alternatives
The next step is to consider if alternative views on correspondence and constraint interactions, unaided by the P-map, can achieve the desired result of cutting down appropriately on the number of solutions to phonotactic constraint violation. Two of the solutions listed in (2), nasalizing the voiced stop or leniting it to an approximant, will have the e¤ect of changing not one feature but possibly two: in both cases an oral stop ([son, nasal, cont]) becomes a sonorant (either [þson, þnas, cont] or [þson, nas, þcont]). One might hope to discover a formal solution to the Too-Many-Solutions problem, by noting that a one-feature modification (i.e., violating one Ident F constraint) is better than a two-feature modification (violating two Ident F constraints). It is not clear how this idea can be implemented, because rankings of the form Ident F Ident G, Ident H cannot be ruled out in principle. However, there is independent reason to believe that the cause of our problem does not reside in the count of features being modified. This can be shown by observing that in languages like Turkish (Inkelas and Orgun 1995) where stops—not fricatives—are subject to final devoicing, the active constraint must be (22). (22) The Turkish version of *[þVOICE]/ *[þvoice, son, cont]/ ]
]
This constraint can, in principle, be satisfied by turning voiced stops into fricatives to avoid devoicing. But Turkish reacts to violations of (22) exactly as Russian or Dutch react to violations of (1): by final devoicing. Underlying forms like /kitab/ are devoiced ([kitap]), not lenited (*[kitaB]10 or *[kitav]). The real generalization is that stricture contrasts are not being sacrificed when the phonotactic problem at hand is readily solved by voicing adjustments. As discussed above, changes in stricture induce greater perceived di¤erences than changes of voicing. The same point arises in connection with place phonotactics. Certain heterorganic obstruent clusters—among them tp, dp, tk, dk—are frequently disfavored or impermissible, as in Korean, ancient Greek, or classical Latin. Consider now ill-formed /dk/ inputs (e.g., Latin ad-kelera:re, surface [ak:elera:re]). Such inputs lend themselves to multiple fixes: [ak:elera:re] versus *[askelera:re], *[alkelerare], and so on.
The Phonology of Perceptibility Effects
167
One of these solutions is widely attested—gemination or place assimilation—while the others are simply unheard of. The mapping /adkelerare/ ! [ak:elera:re] entails the loss of two features (voicing and coronality): it is unclear what feature-counting alternative to the P-map will establish this mapping as preferable to the /adkelerare/ ! *[alkelera:re] mapping, in which obstruency is lost and laterality is added. The generalization here is that when the same phonotactic problem can be addressed by adjusting either place or stricture, the solution is to change place. Consider now the viability of C-deletion as a solution to final voicing violations. Recall that underlying forms like /tæb/ could—but never do—satisfy (1) by dropping the final voiced stop altogether. Here too one might imagine that a di¤erent modification of the theory of correspondence, one that substitutes Max [aF] for Ident F constraints (Casali 1997; Lombardi 1998), will explain the preference for devoicing. Thus the output of devoicing, [tæp], violates only Max [þvoice], while the output of C-deletion, [tæ], violates Max [þvoice], plus Max [labial], Max [cont], Max [nasal], and so on. On this view, the C-deleting candidate loses under any ranking of the Max constraints. But this cannot be the answer either. Consider the constraint against stop þ noncoronal stop sequences (tp, tk, dp, dk) active in ancient Greek. Most such inputs arise at the boundary between the verb root and the perfect ending -ka, and the constraint is satisfied in this case through [t/d] deletion —for example, ke-komid-ka ! [kekomika] ‘I have eaten’. Thus the Greek solution to the tp, tk problem is not to place-assimilate, as in Latin or Korean, but rather to drop the first stop altogether—that is, to violate Max [a voice], Max [coronal], Max [cont], and so on. If we look at this problem in terms of the number of features being sacrificed from the input, we cannot understand why the [d] of komidhad to drop, when it could well have been turned into [l], [r], [s], yielding well-formed *[kekomilka], *[kekomirka], or *[kekomiska]. Each of these alternatives contains fewer Max F violations than the solution actually adopted, which was to eliminate the [d] altogether. (23) Failed attempt at C-deletion in a system with Max F and no Ident F /ke-komid-ka/
Max [cont]
Max [coronal]
kekomika
*
*!
+ kekomilka
*
A theory that relies exclusively on Max [aF] cannot explain any pattern in which a segment is deleted in contexts where the phonotactic violation can be met by modifying a subset of its features. If we adopt, along with the Max [aF] constraints, Dep [aF] constraints, then the ranking Dep [þstrident], Dep [þnasal], Dep
168
Donca Steriade
[þcontinuant] Max [son], [cont] can induce [t/d] deletion. This move, however, brings back the original problem: in a system where every feature value possesses its Dep constraint, devoicing violates Dep [voice]. Then what rules out Dep [voice] Max [son], [cont]? What emerges from this discussion is that some hierarchy of features must be assumed in any approach; one must recognize that modifications of voicing, especially final voicing, matter less than modifications of obstruency. This is the first step on the way to the P-map. For the P-map analysis, the Greek [t/d] deletion process raises not the formal problem faced by Max F analyses but an empirical question: Is the contrast between unreleased t/d and Ø in prestop position judged less distinctive than that between t and s or t and n or t and l in the same position? If yes, then the we predict that, in such a context, straight deletion is more likely than fricativization or lenition to a sonorant. We lack similarity data bearing on this, but studies of assimilation and cluster simplification (Wilson 2001a; Steriade 1999a) suggest that a perceptibilitybased solution will be fruitful for this case too. 7.5 7.5.1
P-Map Effects in Cluster Simplification Size-of-Cluster Constraints
Many languages constrain agglomerations of C’s when they exceed some specified size. If the constraint responsible for size-of-cluster phenomena prohibits strings of the form C i C j / K, where K specifies a context, segmental or prosodic, then a representation violating it can achieve compliance in at least three ways: by deleting C i , by deleting C j , by modifying either of them or by adding a V, the insertion of which will yield further choices regarding site and V quality. In this section I briefly suggest that this wealth of apparent choices in dealing with size-of-cluster constraints fails to reflect phonological reality: the actual solution comes much closer to being predetermined by the composition of the string containing the violation. While the choice between V-insertion and C-deletion might remain free in resolving a size-of-cluster violation, other decisions (which C to delete; which C to modify and how; where to insert a V and which V to insert) are partly or fully predictable. They are predictable largely in terms of the relative similarity between the input and the modified output: it is the most similar input-output pair that is predominantly selected. The issue of predictability in intervocalic CC cluster simplification has been independently identified by Wilson (2001a), whose formal proposal di¤ers from mine but whose discussion raises points related to those made here. The partial predictability of epenthesis site in initial clusters is analyzed in a framework akin to the P-map by Fleishhacker (1999, 2005). Here I extend Wilson’s observations by considering briefly the choice of C’s to delete in sequences more complex than VCCV.
The Phonology of Perceptibility Effects
7.5.2
169
Similarity with Ø in Cluster Simplification
Wilson (2001a) notes that when C-deletion targets a sequence VC i C j V, the lesser perceptibility of C i leads to its loss. I consider here the choice between deletion of C i and C j in VC i C j Ck V. To simplify matters, I assume, with Wilson, that the prevocalic Ck is undeletable here. My basic empirical point is that the target of deletion is predictable from considerations of confusability, not from its prosodic position or its adjacency to the V. The P-map’s predictions for cluster simplification in VC i C j Ck V are derived from the degree of distinctiveness of two contrasts: C i vs. Ø/V C, and C j vs. Ø/C C. To be judged su‰ciently distinct from Ø in a given context, the sound must in fact be su‰ciently distinct from both of the elements adjacent to it. To see this, suppose that C j in VC i C j Ck V is confusable with C i : the percept resulting from this confusion is VC i C i Ck V or VC i :Ck V. The e¤ect of shortening-in-clusters (Haggard 1973; Klatt 1973) renders VC i :Ck V confusable, in turn, with the simplified VC i Ck V. Thus C j ’s similarity to C i leads, under C i -C j adjacency, to C j ’s similarity to Ø. The same holds if C j is similar to Ck . Likewise, postvocalic C i in VC i C j Ck V is similar to Ø, if it is too similar with either the preceding V (a confusion leading to the V:C i Ck V percept) or to the following C j . Finally, consider a sequence C j Ck V in utterance-initial position: the initial C j is confusable with Ø if it is confusable with either the absence of sound that precedes it or the Ck that follows. Mutatis mutandis, the same holds for confusion with Ø of Ck in an utterance-final VC j Ck . Similarity with Ø means then similarity with either of the adjacent elements, whether silence or sounds. From this we predict that position relative to the syllable boundary or a V will not guarantee that C i is less confusable with Ø than C j . The simplest example illustrating this is the loss of postvocalic liquids in systems where other postvocalic C’s are preserved. In such cases what identifies the element deleted is not proximity to the V or syllable position but similarity to the neighboring V. Cho (1999) presents an example of this type in his analysis of postvocalic [1] deletion in Korean VlCCV. A more complex illustration of the same point is the di¤erence between the confusability with Ø of interconsonantal stridents and stops. Consider first the case in which the sequence VC i C j Ck V contains three stops. As a stop, C i is su‰ciently distinguishable from the immediately preceding V. Moreover, since this V carries C i ’s transitional cues to place and voicing, it provides information distinguishing C i from other C’s, including C j , which might have occurred in the same position. Therefore C i is not confusable with either the V or the following C j , hence it is not confusable with Ø. The medial C j , on the other hand, is confusable with Ø. Because no V is adjacent to C j , the string VC i C j Ck V contains less information allowing the listener to di¤erentiate C j from any other stop that might have occurred in the VC i Ck V position, including from C i or Ck . If C j is confusable with either of the adjacent stops, then the string VC i C j Ck V is confusable with VC i :Ck V or VC i Ck :V and hence
170
Donca Steriade
with the shortened variant VC i Ck V. And therefore C j is more confusable with Ø, than C i . In this way we reach the conclusion that, if cluster simplification targets the C that is most confusable with Ø, then it will operate in this case at the expense of the medial C j . However, this holds only for cases in which all but position relative to the V is equal between the three C’s. Suppose that C j in our VC i C j Ck V string is a strident. Then the distinctiveness di¤erence between D(C i vs. Ø)/V C, and D(C j vs. Ø)/C C might be obliterated or reversed, because the inherent noisiness of the strident C j identifies it as distinct from any adjacent nonstridents, even in the absence of vocalic transitions. If so, the relevant correspondence constraints (Max C/V C and Max strident/C C) either remain unranked or Max strident/C C might in fact rank higher. 7.5.3
Test Cases
Several predictions follow from this. First, a language may delete interconsonantal stops but not interconsonantal stridents. This pattern occurs in Dihovo Macedonian (Groen 1977). To analyze it we need a size-of-cluster constraint, interpreted as the requirement that each C be adjacent to a V: (24) C//V Every C is adjacent to a V. The Dihovo pattern of cluster simplification corresponds to (25): (25) Stops, not stridents, are deleted between stops in VCCCV. Max [cont]/V C, Max strident/C C C//V Max [cont]/C
C
The clear e¤ect of the P-map in this case is the ranking Max stop/V C, Max strident/C C Max stop/C C. The position of the phonotactic C//V relative to the correspondence constraints is left undetermined by the P-map, which allows us to predict variation in the patterns of cluster simplification. Thus the modified hierarchy in (26), where C//V has climbed higher, requires that some cluster simplification take place even in V-stop-strident-stop-V clusters. (26) All VCCCV clusters are reduced to VCCV. C//V Max [cont]/V C, Max strident/C
C Max [cont]/C
C
Further elaboration of (26) yields two types of simplification for VC i C j Ck V sequences where C i is a stop and C j a sibilant: either VC i C j Ck V ! VC i Ck V or VC i C j Ck V ! VC j Ck V. What is, however, invariant is that if the middle C j is a stop surrounded by obstruents, it will always be deleted. Colloquial Latin illustrates the more revealing pattern: interobstruent stops are lost, whereas interobstruent [s] is preserved at the expense of the stop preceding it.
The Phonology of Perceptibility Effects
171
(27) Two types of cluster simplification in Latin (Niedermann 1953) VC i stop Ck (V) ! VC i Ck (V)
VC i s Ck (V) ! V s Ck (V)
pa:sktus ! pa:stus nokts ! noks temptare ! tentare lampterna ! lanterna k w inktus ! k w intus
sekstus ! sestus obstendo ! ostendo apsporto ! asporto sekskenti: ! seskenti pinstus ! pi:stus
This cluster-reduction pattern suggests that the strident-Ø contrast is more distinctive, even in the absence of contextual cues, than the postvocalic stop-Ø contrast. (28) Simplified cluster-reduction hierarchy for Latin11 C//V Max strident/C C Max [cont]/V C Max [cont]/C
C
(29) Cluster reduction in obstendo /obstendo/ optendo
Max strident/C__C
Max [cont]/V__C
*!
+ ostendo
*
(30) Cluster reduction in kwinktus /kw inktus/ kw iktus + kw intus
Max [cont]/V__C
Max [cont]/C__C
*! *
The Latin asymmetry between postvocalic stops and sibilants as targets of cluster simplification is encountered in several languages, among them Finnish, Catalan (Wheeler 1979), and colloquial Polish (Madejowa 1992). The alternative pattern of deletion, where every interconsonantal obstruent deletes, whether it is a stop or a sibilant, is perhaps also attested in Greek, Sanskrit (Steriade 1982), and Korean (Kim-Renaud 1974), but alternative interpretations are available for these cases. In particular the analysis of cases like Korean kaps-to ‘price-and’ ! [kapto] must take into account the fact that all fricatives are prevocalic in Korean: if [p] had deleted, the actual outcome would have to be *[katto], not *[kasto]. The tableau in (31) indicates that the Latin ranking of correspondence constraints need not be changed to derive this case.
172
Donca Steriade
(31) Cluster reduction in kapsto /kapsto/ kasto
*s/__C
Max strident/C__C
*!
Max [cont]/V__C *
katto
*
+ kapto
*
*!
The Greek and Sanskrit instances of deleted interobstruent [s] are sparsely attested and involve exclusively su‰xal [s]. It is possible then that the Latin reduction pattern represents the general case. My general claim, however, is more modest. A P-map account, as sketched here, predicts only this: insofar as a C-cluster contains one and only one C whose similarity to Ø is greater than that of the other cluster members, cluster reduction will target this one C. Similarity to Ø means similarity to an adjacent element. We have seen that an interobstruent stop—or a stop flanked by a nasal and an obstruent—can be identified as the most confusable with Ø among all components of its C-cluster. This corresponds to the observation that stops in such contexts are the systematic, invariant targets in cluster simplification. It may also turn out that the inherent salience of stridents renders the D(strident-Ø)/C C contrast more distinctive than D(stop-Ø)/V C. If so, a stronger prediction is made: the postvocalic stop will always be deleted, unless morphological factors intervene, in V-stop-s-C sequences. 7.6
Insertion and the Ranking of DEP Constraints
The P-map account of the choice of epenthetic segments derives from the hypothesis of a context-dependent hierarchy of similarity between individual segments and Ø. If a phonotactic constraint requires insertion of a segment in some context K, then the segment most confusable with Ø in K is predicted to be the choice of insertion. I outline now how this prediction follows from proposals made thus far. The class of correspondence constraints violated by insertion take the form in (32). (32) DEP (I-O) schema There is no D(Ø-x) between I and O, such that I contains Ø in K and O contains x in K 0 and K corresponds to K 0 . Like all correspondence constraints, the Dep constraints are projected from the Pmap. This means that if D(Ø-x)/ Ki > D(Ø-y)/ Kj , then corresponding to this similarity ranking there exists a ranking of correspondence constraints so that the more confusable contrast with Ø corresponds to the lower-ranked Dep constraint (cf. (20)). It follows that the outcome of phonotactically motivated insertion is to a large extent
The Phonology of Perceptibility Effects
173
predetermined. This prediction is mitigated by the e¤ect of conflicting phonotactics and thus the e¤ect of markedness on epenthesis has to be considered as well. Next I outline only one aspect of the evidence bearing on this point: the selection of epenthetic segment quality. 7.6.1
Epenthetic Glottals
The typology of epenthetic C’s has been usefully outlined by Lombardi (2002), who identifies a general pattern, insertion of [], and minor deviations from it, due either to structure preservation (in the form of constraints forbidding []), morphological constraints, or the dispreference for rhyme []. The central question is what accounts for the preference for inserted []. Lombardi assumes that the relevant factor is markedness: [] is the least marked among C’s. But what fact other than its propensity to get inserted reflects []’s extreme unmarkedness? This is a harder question: the standard evidence for markedness, the implicational universals, suggest otherwise: []’s presence in an inventory is not asymmetrically implied by the presence of all other C’s, or indeed by the presence of all other members of its stricture class. Under the P-map hypothesis, [] has, with [h], a uniquely favorable property for an epenthetic C: it does not possess an oral constriction and thus it will fail to induce coarticulatory changes on neighboring segments, unlike the orally articulated C’s. If we compare input-output pairs of the form V(input)-CV(output), the most similar ones will be V-V or V-hV or V-GV, where G is homorganic to V. Both epenthesis of [h] and epenthesis of homorganic glides represent in fact the only widely attested epenthesis patterns, along with the more common case of [] insertion. Thus, if the lack of coarticulatory V modification translates into the similarity hierarchy in (33), then the P-map hypothesis predicts the preference for [] as epenthetic segment, regardless of how it rates in markedness. (33) D(Ø-t)//V; D(Ø-k)//V; D(Ø-p)//V > D(Ø-)//V The view that markedness determines the choice of epenthetic segments runs into independent di‰culties. Several of the languages Lombardi cites, where [] occurs as an exclusively epenthetic C, must in fact be assumed to rate the markedness of [] as higher than that of all their other C’s. Thus to explain the fact that German allows only epenthetic [] we have two options. We can assume that a constraint *[V *[] Max C while at the same time assuming Max C *[p], *[k], and *[t]. But this contradicts the universal markedness ranking Lombardi assumes: *[p], *[k] *[t] *[]. The alternative is to rely on the P-map-based ranking: Max/Dep [p],[t],[k]/ V Max/Dep []/ V. The German choice of epenthetic [] follows when this correspondence ranking is embedded in the complete analysis: Max [V, Max/Dep [p],[t],[k]/ V *C *[V Max/Dep []/ V. On this view, it is correspondence, not markedness, that di¤erentiates C types. The same conclusion follows
174
Donca Steriade
from Lombardi’s analysis of Asheninca, one of the rare languages where something other than a laryngeal or a glide is inserted in hiatus. Lombardi argues that [t] is inserted in Asheninca because *[] is undominated. If we grant this, it follows that *[] *[t], since *[t], but not *[], is outranked by Max C. This too is incompatible with the claim of unmarked status for []. I conclude that there is either no constant context-free, all-purpose preference for glottal as against other stops, or, if there is a preference, it is the opposite from the one needed to predict the proper choice of epenthetic C. The choice is predicted by the P-map. 7.6.2
Epenthetic Schwa
I document next the status of [ ] as the inserted V of choice. This too follows from a hierarchy of similarity with Ø. What defines for our purposes the class of schwa vowels is not their midcentral quality (Romanian and English [ ], for instance, are not epenthetic, while Romanian [ ] is) but rather the fact that the schwalike V is significantly shorter and more variable in quality than all other V’s in an inventory. This characterization allows for some diversity in the actual quality of a language’s neutral V, while permitting us to make specific predictions about what di¤erentiates it from other V’s. A systematic di¤erence of duration between schwa and other V’s of Dutch is documented in Koopmans-van Beinum 1994, and known informally to obtain for English and French schwa. Further, Dutch schwa is also more variable in its F2 values than other Dutch V’s (Koopmans-van Beinum 1994; van Bergem 1995). Assuming then that the defining properties of schwa are shortness and variability, the preference for schwa as an epenthetic element follows form the fact that it is, in both duration and relative absence of invariant articulatory properties, the closest thing in a V system to no segment at all (i.e., to Ø). Note that this is not the same as saying that schwa has no properties. First, it is a vowel. When it does occur, speakers count an extra syllable. This is invariant. Further, schwa in Dutch is less durationally variable than other V’s (Koopmans-van Beinum 1994): it is least subject to contextual or context-free lengthening. In many languages where schwa is unstressable, as in Dutch, Indonesian, and French, this can be attributed to the fact that schwa cannot be lengthened. In that respect then it does have a second invariant property: short duration. Thus we cannot explain why schwa is preferentially inserted by assuming that it possesses no properties, or a subset of the invariant properties of other vowels. Schwa epenthesis is preferred because the P-map identifies it as the closest thing to no epenthesis at all. The preference for schwa insertion may manifest itself in a language independently of the composition of the V inventory. However, when structure preservation does not constrain its occurrence—that is, in languages possessing contrastive or nonalternating schwas—this preference for inserting schwa is absolute. The statement in (34) holds of all relevant cases I have encountered, some of which are listed below. e
v
The Phonology of Perceptibility Effects
175
(34) If a language contrasts schwa and zero in some context, or if it contains nonalternating forms with schwa, and if it resolves clusters through epenthesis, then the choice of productive epenthetic vowel is limited to schwa. (Indonesian (Adisamito 1993), Romanian (Avram 1990 and below), German (Giegerich 1987), Damascene Arabic (Bohas 1986), French (Dell 1978 and below), Meitei (Chelliah 1997), Miya (Schuh 1996), Welsh, English, Dutch (Booij 1995; Kuijpers, van Donselaar, and Cutler 1997), Berber (Kossman 1995; MacBride 1990)) The comments about the markedness of [] apply here too: schwa insertion does not stem from a context-free preference for this segment. It most clearly does not in languages where schwa is permissible only where a V is otherwise needed. In Berber, for instance, schwa—but no other V—must be prevented from occurring in open syllables (MacBride 1999); in Miya, it cannot occur after a sonorant (Schuh 1996). Phenomena of this sort require *[ ] constraints, whose high ranking is not paralleled by other *V conditions: this precludes a claim of unmarked status for schwa. What explains the V-epenthesis generalization is the existence of a hierarchy of Dep V constraints containing, at its bottom, Dep [ ]. Here too I speculate that the source of this Dep V hierarchy are speakers’ judgments of relative similarity between individual V’s and Ø. e
e
7.7
All-Purpose Segments
I turn next to a di¤erent respect in which the P-map proposal tightens the theory of correspondence. The suggestion here will be that the segments most likely to be inserted are also most likely to be deleted. The behavior of [ ] will illustrate this, but reports about specific segments being both preferentially inserted and deleted go beyond this case.12 This phenomenon is not predicted by the classic theory of correspondence, but it follows from the same principle (21) that solved the Too-ManySolutions puzzle involving final devoicing. (21) predicts that if D(x-Ø)/ K > D(y-Ø)/ K then not only will Dep (x)/ K Dep (y)/ K but also Max (x)/ K Max (y)/ K. Then y has priority for both insertion and deletion over x. More concretely, this means that if some y-Ø contrast is identified as more confusable than other segment-Ø contrasts, the insertion and deletion of y will be the preferred response to all phonotactics for which insertion or deletion of segments from y’s class represent a potential solution. The general class of situations described is this: a language avoids hiatus, hence it must delete a V when adjacent to others or insert a C between them. The same language avoids CCC clusters, hence it must insert some V in such clusters, or else delete a C. Because it appears impossible to predict the preference between C-insertion e
176
Donca Steriade
and V-deletion, or between C-deletion or V-insertion, these choices are settled on a language-specific basis. The language we are interested in eliminates clusters by Vinsertion and hiatus by V-deletion. Under these conditions, the P-map hypothesis predicts that the V deleted in hiatus is the same as the V inserted as a clusterresolution strategy. That is because the criterion that selects a V for deletion is the same as the one selecting it for insertion. This criterion is the greater similarity of the contrast between that V and Ø relative to all other V’s. I summarize this below, using [ ] as the V judged to be most confusable with Ø. e
(35) If, for any choice of V0 , D(V-Ø) > D( -Ø) Then Max V, Dep V Max ( ), Dep ( ) e e
e
e
Note that, aside from the P-map, nothing guarantees that the Max and Dep constraints corresponding to di¤erent V’s will be ranked as pairs. Nothing excludes rankings like (36), which predict that schwa is deleted but that [a] is inserted. (36) Schwa deleted and [a] inserted in system lacking P-map, where MAX/DEP constraints exist for individual segments Max (a), Dep ( ) Phono-constraints Max ( ), Dep (a) e
e
Systems exist in which hiatus is resolved at the expense of certain V’s only (Pulleyblank 1988; Casali 1997; below). As Casali argues, these systems reflect a hierarchy of distinct Max constraints, independent of the markedness scale: deletion targets the V associated with the lowest ranked Max V. This dispenses with alternatives to the P-map analysis that assume that targets of deletion/insertion are determined by markedness conditions alone, interacting with a monolithic Max V, Dep V. One can document systems where both V-insertion takes place and specific V’s are deleted, either to avoid hiatus or to shorten the word. In all such cases, it is the prediction of the P-map analysis, (35), that is upheld. The pattern emerges more clearly if we restrict our attention to productive, lexically unrestricted insertion and deletion. Thus in French, it is schwa that deletes in hiatus, regardless of its location relative to the other V. (37) French schwa deletion a. Optionally deleted in VC CV contexts; no other V deletes la pelouse [lapluz] ‘the lawn’ cf. phrase-initial pelouse [p luz] pas de role [pad‚ol] ‘no role’ cf. phrase-initial de role [d ‚ol] b. Obligatorily deleted in hiatus; no other V deletes t’entendre ‘to hear you’ cf. te remercier [t ‚m‚sje] ‘to thank you’ vivre ailleurs [viv‚ aj˛‚] ‘live elsewhere’ cf. vivre la` [viv‚ la] ‘live there’ e e
e
e
Compare ni entendre [ni þ tþ d‚ ] ‘neither to hear’ and vivra ailleurs [viv‚a aj˛‚] ‘will live elsewhere’, both of which surface with hiatus, in the absence of a deletable vowel. Schwa is also inserted, optionally, to avoid clusters of obstruents. e
The Phonology of Perceptibility Effects
177
Romanian [ ] gives rise to identical patterns: (38) Romanian schwa a. [i] optionally deleted next to a V; no other V deletes vine ‘ndat ‘comes immediately’ cf. indat ‘immediately’ vine ‘nainte ‘comes before’ cf. inainte ‘before’ b. Nondeleting V’s vine odat ‘comes once’; vine la ‘comes that one-masc.’ vine aja ‘comes that one-fem.’ c. [i] optionally inserted in obstruent clusters CCC(C); no other V is inserted opt-spre-zece [optsprezetSe] @ [opt sprezetSe] ‘18’ (‘eight-to-ten’) Comparable patterns appear in Dutch (Booij 1995; Kuijpers, van Donselaar, and Cutler 1997), Meithei (Chelliah 1997), and English (Hooper 1978). 7.8
Conclusion
I note in closing potential points of dispute relating to the P-map. At the most basic level one can dispute the premise this account shares with most modern phonology, namely that phonology is a problem-solving system, or—as Goldsmith (1993) puts it—‘‘an intelligent system.’’ If the phonotactic in (1) is not viewed as a problem to be solved, or as a standard of well-formedness independent of the lexicon’s contents, but rather as a generalization over the words that happen to be attested in one’s language, then no Too-Many-Solutions problem arises: learners, on this view, do not seek to find a solution to (1) but to learn whatever patterns happen to be instantiated by their lexicon. Similarly, one may question whether the Too-Many-Solutions problem arises in the initiation of sound change. The view presented here is that innovators may aim to improve a sound system and that they do so in the safe regions of confusability identified by the P-map. We assume, for instance, that speakers who initiate final devoicing have a choice of methods to satisfy (1)—or a choice of spontaneously occurring speech variants to promote (Lindblom et al. 1995)—and choose final devoicing because it involves the smallest departure from established speech norms. But it may be possible to look at the initiation of sound change in di¤erent terms if most naturally occurring variants to an established lexical form represent its common misperceptions. In that case, innovators have the more passive role of simply favoring the more commonly noted deviations from the norm, without reflecting on their phonotactic virtues or on their similarity to canonical forms. This possibility has been discounted here on the strength of evidence that speakers know not only what the more common deviations from the norm are but also which deviations are more similar to the norm. We have seen that knowledge of similarity is displayed in rhyming
178
Donca Steriade
practices and experiments seeking overt similarity judgments. It remains to be established, however, that the available knowledge of similarity is exploited by speakers in constructing grammars and in favoring one change over others. (See also Steriade 2001.) A di¤erent class of possible objections to the P-map involves the fact that there is, at least at first sight, a considerably greater variety of alternations than a theory of perceived sound similarity may predict. If the case for predictability of C-deletion or V-insertion appears overstated, a way of testing the P-map proposals is to focus on fully productive, not yet lexically entrenched processes. For the moment, it seems necessary only to acknowledge the existence of parochial constraints governing alternations, in addition to phonotactics and P-map generated correspondence constraints. Finally, I have focused here on aspects of perceived similarity that correspond to broad crosslinguistic generalizations, and for this reason it may appear that a claim of universality is made regarding the contents of the P-map. This is not the intention. If the perception of similarity is governed, in part, by ‘‘the contents of the universe of discourse’’ (Tversky, cited in Frisch, Broe, and Pierrehumbert 1997), then the same pairs of sounds will rate di¤erently for similarity, when embedded in di¤erent systems. The existence of such e¤ects is not denied; the development of a firstapproximation version of the P-map will hopefully make it possible to identify them. Notes I am grateful to the editors, Kristin Hanson and Sharon Inkelas, and to Bert Vaux for detailed and helpful comments on a first draft. Questions from audiences at Berkeley, Stanford, UCLA, and the 1st North American Phonology Conferences at Concordia University, Montre´al, have helped clarify some of the issues discussed here. Most of the writing was done in fall 1999, while I was in residence at the Miller Institute for Basic Science, at UC Berkeley. Many thanks to Andrew Garrett, Larry Hyman, and Sharon Inkelas for engineering my stay at the Miller. This study is dedicated to Paul Kiparsky, whose writings converted me in the late ’70s from Indo-European studies to phonology, and whose teaching and example since then have inspired my work. 1. What exactly causes final obstruents to devoice is a matter of debate; see Lombardi 1999 and Steriade 1999 for two recent proposals, among many. The precise nature of the phonotactic does not a¤ect the problem outlined in this study. 2. The correspondence constraints cited are those proposed by McCarthy and Prince 1995. The argument carries over to other views on correspondence. See section 7.5.3. 3. See Connine, Titone, Deelman, and Blasko 1997. 4. See Ito 1986, Paradis 1988, Goldsmith 1992, Calabrese 1995. 5. See Mohr and Wang 1968, Walden and Montgomery 1974, van den Broecke 1976.
The Phonology of Perceptibility Effects
179
6. Flemming’s point is that it is faithfulness that’s evaluated on perceptual representations: ‘‘The output must sound like the input.’’ The same holds for any similarity relation. 7. See Wingstedt and Schulman 1988 and Fleischhacker 1999 for relevant experimental evidence. A further indication of the same point is that partial subsequence rhymes of the form VC1 C2 -VC1 Ø (e.g., best-less) are frequent in English (Zwicky 1976) and elsewhere (Steriade 2003) while subsequence rhymes of the form VC1 C2 -VØC2 (e.g., bets-less) are extremely rare. 8. See, however, section 7.4.2. 9. See Blevins and Garrett 2004 for relevant discussion and references. 10. Readers who object that [kita”] is a non-structure-preserving change—as Turkish lacks bilabial fricatives—will recall that under the ranking Ident [Gvoice] *B, Ident [Gcont] structure preservation should be irrelevant. Indeed, voicing adjustments can be non-structurepreserving in languages like German and Catalan, where final devoicing is incomplete and does not obliterate the contrast. Thus a formal ranking generating [kita”] exists under the standard account. The question is why it is unattested. 11. I am not claiming that C//V is undominated in Latin: there exist clusters like mbr, ltr, str, and so on. But the focus here is on the fate of medial C-obstruent-obstruent sequences in which cluster simplification did occur regularly in the spoken language. To obtain the more accurate account, one must assume Max constraints that outrank C//V. For these Max constraints, the P-map’s claim is that they involve clusters whose individual members are better distinguishable from Ø than the C’s that do in fact delete. 12. See Pulleyblank 1988 for the observation that the same vowel may be both the prevalent target of deletion and the preferred inserted element in selected languages. A particularly interesting case of epenthetic/deletable C is discussed in McCarthy 1993, namely postvocalic rinsertion and deletion in New England varieties of English. Not surprisingly, postvocalic [r] in most varieties of American English is an approximant hardly distinguishable from the end of a preceding low back vowel; it may thus be the closest thing to Ø in that context.
8
Vowel Duration, Syllable Quantity, and Stress in Dutch
Carlos Gussenhoven
8.1
Introduction
A persistent conundrum in the analysis of Dutch word prosodic structure has been the fact that closed syllables require a foot head, while long vowels apparently fail to project a foot (van der Hulst 1984; Kager 1989, 261; Zonneveld et al. 1999, 499; among others). This type of selective quantity sensitivity is highly marked: in other quantity-sensitive languages, long vowels are heavy (bimoraic), and will be in a stressed syllable (type A), while in addition such languages may require closed syllables to be heavy, and be stressed (type B). The latter option was termed weight-byposition by Hayes (1989a). In (1), this well-known typology is given. (1) a. Short vowel (V) b. Long vowel (VV) c. Closed rhyme (VC)
Type A m mm m
Type B m mm mm
?Dutch m m mm
There have been three responses to the apparent quantity-weight mismatch in Dutch. First, Lahiri and Koreman (1988) proposed that weight and quantity are represented separately: while weight is counted in moras, quantity is counted in X-slots. As a result, the light long vowels of Dutch can be represented as having one mora dominating two X-slots. Second, Kager (1989, 261) suggested that weight, which in other languages is determined by the branchingness of the syllable peak (short vowel versus long vowel, diphthong, short vowel plus C) is determined by ‘‘melodic complexity,’’ or the number of segment root nodes associated with moras after the first mora in the rhyme (light monophthong versus heavy diphthong or VC). A third response is that by van Oostendorp (1995), who assumes that so-called long vowels are not in fact represented as long, but di¤er from short vowels in lacking the vocalic feature [lax]. The specification of the duration of Dutch vowels will, in his view, be provided during the phonetic implementation.
182
Carlos Gussenhoven
It is argued that none of the above responses is tenable, and that the solution lies in a reevaluation of the phonetic facts. Unlike what is generally assumed, the long tense vowels of Dutch are only longer than short lax vowels in stressed syllables— that is, in the head of the foot. This suggests that Dutch stressed syllables are bimoraic, while unstressed syllables are monomoraic. The dependence of vowel length on stress will allow us to assume, with van Oostendorp (1995), that there are no underlying moraic representations for Dutch syllables. (For long ‘‘marginal’’ vowels, see section 8.3.) Bimoricity (and restricted occurrences of trimoricity) are the result of (i) weight-by-position (the projection of moras by coda consonants) or (ii) stress-to-weight (a bimoricity requirement on stressed syllables). Thus, in the analysis presented here, quantity plays a major part in the phonological representation of Dutch vowels, unlike van Oostendorp’s (1995, 1997) analysis, which assumes mora-less surface representations as well as mora-less underlying representations. It will be shown that moraic representations of vowels are in fact part of the lexical phonology of Dutch, and that a description of Dutch word prosodic structure is impossible if the moraic structure is left unspecified. In section 8.2, some new facts about Dutch vowel duration are given, together with a brief description of the experiment that yielded them. Section 8.3 shows that vowel quantity is partly determined in the lexical phonology, and that the specification of quantity cannot be left to the phonetic implementation in the sense of van Oostendorp. Section 8.4 describes the regular stress patterns of Dutch, drawing on the analyses in Nouveau 1994 and van Oostendorp 1997, while section 8.5 shows that the success of this description crucially depends on the correct analysis of the moraic structure of Dutch words, including the part before the main stress. A conclusion is o¤ered in section 8.6. 8.2
Duration and Distribution of the Vowels of Dutch
Table 8.1 lists the vowels of Dutch that can appear in a stressed syllable nucleus (e.g., Moulton 1962; Gussenhoven 1992; Booij 1995). There is a set of lax, short vowels, a set of tense vowels of which the [high] vowels are long, a set of tense or lax, oral or Table 8.1
The vowels of Dutch that can appear in stressed syllables Short (lax)
Long or short (tense)
Long
I y
i y e ø
i y ˇ
þ
u o a
Diphthong u þ þ
i ˇy u
Vowel Duration, Syllable Quantity, and Stress in Dutch
183
nasal, long ‘‘marginal’’ vowels, which only occur in recent loans and onomatopoeic words, and finally, there are three diphthongs, as shown in table 8.1.1 The vowel [ ] (not listed) only appears in unstressed syllables. The heaviness of VC-rhymes is not obvious, but can be seen in trisyllabic words. Dutch has quantitity-sensitive trochees, built from right to left. The atypical ‘‘fact’’ that closed syllables are heavy and long-voweled open syllables light can be seen by comparing words of the type almanak ‘almanac’ with words like Gibraltar (the proper name). Both of these have a closed final syllable, but they di¤er in the structure of the penult. The closed penult attracts stress, leading to [xi’brþltþr], while the open penult is regularly skipped, causing main stress on the antepenult. The above interpretation of vowel quantity thus gives rise to the belief that long vowels, like [a] in [’þlma"nþk], are light, even though closed syllables are heavy. In reality, the duration of [e,ø,o,a] in unstressed positions is equivalent to that of short vowels. Using reiterant speech, Rietveld, Kerkho¤, and Gussenhoven (2004) investigated the duration of [a] and [i] in a large number of word prosodic contexts. While [i,y,u] are tense and have the same distribution as ‘‘long’’ tense vowels, their duration happens to be identical to that of short lax vowels: the vowels of [zIt] ‘sit’ and [zit] ‘see-3sg’ have equivalent durations, which may be only 50 percent of the duration of the vowel in [zat] ‘seed’ (Nooteboom 1972, 66). To identify the prosodic positions in which the duration contrast between short [i] and long [a] is made, nine word prosodic patterns were identified, between them covering all word prosodic positions. (E¤ects of codas on vowel duration were left out of account.) Word stress (as opposed to secondary stress), stress (as opposed to no stress), serial position (initial, nonperipheral, and final syllable in foot, and final foot in word) were each shown to have an e¤ect on the duration of each of the two vowels. A list of words illustrating these patterns is given in (2), where S ¼ word stress, s ¼ secondary stress, and w ¼ unstressed. e
(2) "rodo’dndr n "para’dis "pasifi’kasi "lokomo’tif "mini(")mali’zasi ’oli"fþnt ’oli"fþnt pi’rat pi’jano
swSw swS swwSw swwS swswSw Sws S w s w (inflected word) wS wSw
‘rhododendron’ ‘paradise’ ‘pacification’ ‘locomotive’ ‘minimalization’ ‘elephant’ ‘elephants’ ‘pirate’ ‘piano’
c
e
To exclude confounding segmental factors, the experiment made use of reiterant CV syllables, as produced by four speakers, with alternating occurrences of two out of the three consonants /k/, /s/, /m/ for the C-position, in all possible permutations.
184
Carlos Gussenhoven
Because the reiterant words were pronounced in a carrier sentence of the type Ik DOE (. . .) niet, ‘I do (. . .) not’, where the word doe was realized with a pitch accent, the data abstract away from accentual lengthening and utterance-final lengthening (Cambier-Langeveld 2000). Forty-eight (12 reiterant versions 4 speakers) realizations were obtained for each of the nine prosodic patterns. The results showed that there was never a significant di¤erence between the durations of the two vowels in positions outside the foot head, while occurrences of [a] in the foot head were significantly longer than [i]. The contexts in which a significant duration contrast was absent included the word-initial syllable immediately before the word stress, as well as the word-final syllable immediately after the word stress—for example, in the first and last syllables respectively of [pi’jano] ‘piano’. 8.3
Making the Moraic Structure Reflect the Phonetic Facts
Since phonological representations of vowel quantity are reflected in phonetic duration (Duanmu 1994; Hubbard 1995; Broselow, Chen, and Hu¤man 1997), it is proposed that the ‘‘long’’ vowels in the second column of table 8.1 are bimoraic in stressed position and monomoraic in unstressed position. Thus, from now on, a word like piano will be given as [pi’jano]. This suggests, first, that Dutch ranks a Stress-to-Weight Principle (SWP) (3) high,2 and second, since closed syllables attract stress, that it also ranks a Weight-to-Stress Principle (WSP) (4) high. (3) STRESS-TO-WEIGHT PRINCIPLE (SWP) Foot heads are (minimally) bimoraic. (4) WEIGHT-TO-STRESS PRINCIPLE (WSP) Bimoraic syllables are foot heads. WSP is not only relevant to closed syllables. As observed by Zonneveld (1993), the truly long vowels (see table 8.1, third column) do not tolerate being in an unstressed position: (Rio de) Janeiro [‰a’nro], *[’‰an"ro].3 Nor could the diphthongs (see table 8.1, fourth column) appear in unstressed penults: Khomeiny [ko’mini], *[’komi"ni]. These truly long vowels must thus be represented in the lexicon with two moras, while diphthongs are bimoraic by virtue of the fact that they contain two segments in the nucleus. These facts are consistent with Dutch being an unexceptional ‘‘type B’’ language. Broselow et al. (1997) assume a default markedness constraint SylMon (5), whereby syllables are monomoraic (cf. NoLongVowel in Kager, chapter 17, this volume). This constraint will make ‘‘long’’ vowels short in weak positions. (5) SYLMON Syllables are monomoraic.
Vowel Duration, Syllable Quantity, and Stress in Dutch
185
Tableau (6) illustrates how these constraints characterize [’bata] as the correct form of the brand name Bata. The trochee, assumed in (6), causes the first vowel to be bimoraic by SWP, while the weak syllable defaults to a monomoraic [a], since bimoraic [a] in weak position violates WSP. (6)
bata
WSP
SWP
+ a. (’bata)
*
b. (’bata) c. (’bata)
SylMon
*! *!
**
The bimoricity of syllables with short lax vowels is due to the following tautosyllabic consonant, which is obligatory in this context. Constraint LaxþC (cf. (7)), a reformulation of part of a constraint proposed by van Oostendorp (1995), requires a lax vowel to be monomoraic and be followed by a consonant in the same syllable.4 The moricity of the coda consonant is ensured by high-ranking Weight-by-Position (WbP), given in (8) (Hayes 1989). (7) LaxþC s m [þlax][þcons] (8) WEIGHT-BY-POSITION (WbP) A consonant in the coda projects a mora. As shown by van der Hulst (1985), a consonant after a short lax vowel is ambisyllabic in Dutch if it is required to be in the onset of the next syllable by Onset (9). For instance, a word like Hetty (a proper name) has an ambisyllabic [t], transcribed [t.t]. (9) ONSET A syllable has an onset. In (10), SWP SylMon, together with Onset, causes the [t] to be ambisyllabic, and the first syllable to be bimoraic as a result. Lengthening of [], as in candidate (10c), is correctly prevented by LaxþC.
186
Carlos Gussenhoven
(10)
hti
Onset
WSP
SWP
LaxþC
SylMon
+ a. (’ht.ti)
*
b. (’h.ti)
*!
*
c. (’h.ti) d. (’ht.i)
*!
*
*!
*
e. (’ht.ti)
*!
**
Our account does not so far explain why words like [lat] ‘late’ have long vowels: the bimoricity of the syllable would already be guaranteed by the coda [t]. It is suggested that the long tense vowel in these forms results from the e¤ect of a constraint that maximizes the sonority of the syllable peak. The bimoricity of tense vowels before a tautosyllabic consonant is thus unrelated to the prosodic status of the syllable. The relevant constraint, given as SonPeak (11), can be seen as part of the family of HNuc (Prince and Smolensky 1993, 134). SonPeak must be ranked below LaxþC, to prevent short lax vowels from lengthening. (11) SONPEAK m [cons] Constraints Onset, WSP, and SWP are irrelevant in tableau (12), which shows the crucial ranking LaxþC SonPeak (12) a.
lþt
Onset
WSP/SWP
LaxþC
+ i. lþt ii. lþt b.
SonPeak
SylMon
*
*
*!
**
lat i. lat + ii. lat
*!
* **
Finally, the diphthongs (column 4 of table 8.1) and the truly long (marginal) vowels will behave like long vowels: both elements of the diphthong are [cons]. If these long vowels are lax, they escape shortening by LaxþC because of high-ranking FaithMora (13), which preserves the lexical mora structure.
Vowel Duration, Syllable Quantity, and Stress in Dutch
187
(13) FAITHMORA Preserve mora structure. The rankings FaithMora LaxþC, and LaxþC SonPeak SylMon, together with Onset and WSP/SWP, account for the moraic structure of Dutch vowels, with the exception of short tense [i,y,u], which are short even in stressed position. In the next section we will see how the quantity of these vowels must be accounted for in the lexical phonology, which provides us with an argument for rejecting the (implicit) assumption of van Oostendorp (1995) that the duration of Dutch vowels is purely a result of the phonetic implementation.5 8.3.1
Short Tense Vowels
The quantity of Dutch [i,y,u] is determined by the segmental context: they are long when appearing before [r] in the same foot (Gussenhoven 1993). In (14), words in which [i] appears before [r] in the same foot are compared with words in which this vowel appears in prosodically identical words, but in which it is not followed by [r]. wier Olivier giro pierement fakir
[Vir] [’oli"fir] [’xiro] ["pir ’mnt] [’fakir] e
(14) a. b. c. d. e.
‘algae’ ‘Oliver’ ‘giro’ ‘barrel organ’ ‘fakir’
wiek kolibri kilo lineaal kieviet
[Vik] [’koli"bri] [’kilo] ["lini’al] [’kivit]
‘wing’ ‘kolibri’ ‘kilo’ ‘ruler’ ‘peewee’
When [r] is in the next foot over, the high vowels are short, as in piraat [pi’rat] ‘pirate’, corduroy [’k rdy"r j] ‘corduroy’, admiraal ["þtmi’ral] ‘admiral’. The constraint HighV-m (15) reflects the widespread tendency for high vowels to be shorter than nonhigh vowels. The relatively short distance between the tongue body and the roof of the mouth explains this e¤ect (see Laver 1994, 435). c
c
(15) HIGHV-m High vowels are monomoraic. To ensure that [i,y,u] are long before [r] in the same foot, we postulate Pre-r-mm (16). The articulatory motivation for this constraint is probably to be found in the conflict between a vocalic tongue posture (a convex dorsum and a tongue blade curling down into the lower jaw) and the tongue posture for a coronal [r], for which the front is held in a concave shape behind a tongue tip that curls up. The articulatory transition from a vocalic posture to that required for [r] will thus be more elaborate than a transition to the position for postvocalic [t,s,n,l], for which the front of the tongue may, but need not be concave. Evidently, Pre-r-mm HighV-m, for otherwise high vowels could never be bimoraic in Dutch. (16) PRE-r-mm Tense vowels are bimoraic before [r] in the same foot.
188
Carlos Gussenhoven
Tableau (17) shows how this works for short and long occurrences of high tense vowels. Candidate aiii, in which the intervocalic onset consonant is ambisyllabic, is ruled out by SonPeak, which for this reason must be ranked above WSP/SWP. That is, we cannot satisfy SWP by filling a second mora with the following [t]. (17) a.
rita +
Pre-r-mm
HighV-m
i. ’ri.ta
WSP/SWP
SylMon
*
ii. ’ri.ta
*!
iii. ’rit.ta b.
SonPeak
* *!
*
xiro +
i. ’xi.ro
*
ii. ’xi.ro
*!
iii. ’xir.ro
*!
* * *
*
I give the constraint rankings established so far, with the generalizations they account for: WSP/SWP SylMon: Stressed syllables are bimoraic, unstressed syllables monomoraic. Pre-r-mm HighV-m: High tense vowels are long before [r] in the same foot, but short otherwise. LaxþC SonPeak: Lax vowels are short, despite their occurrence in bimoraic syllables, and thus must be followed by C in the rhyme (or: do not maximize a lax vocalic peak). SonPeak SylMon: Tense long vowels are long even before a coda C (or: do not minimize moraic structure at the expense of a vocalic peak). SonPeak SWP: Do not make consonants ambisyllabic merely to have bimoricity in stressed syllables. Faithmora LaxþC: Do not shorten long lax (marginal) vowels. 8.3.2
Why Vowel Duration Is Represented in the Phonology
At this point it might be objected that the durational phenomena we have dealt with could be accounted for in the phonetic implementation. Specifically, the question arises of whether we couldn’t have a single articulatory instruction, to be carried out with reference to a phonological representation without quantity distinctions, which says: ‘‘make nonlax vowels long when stressed or followed by a tautosyllabic conso-
Vowel Duration, Syllable Quantity, and Stress in Dutch
189
nant, but not when they are high, except when [r] appears in the same foot.’’ This course of action is unavailable, however, because the quantity of high tense vowels is subject to a lexical process. Crucially, this means that moraic structure must be represented in the lexical phonology. There are five irregular past-tense forms that have short [i] before [r] in the same foot. In (18), an example is given. (18) wierp
[Virp]
‘threw’
(cf. vier-t [firt] ‘celebrateþ3sg-pres’)
As Booij (1995, 94) points out, these forms end in a cluster of [r] and a labial obstruent, while no other words do. A phonological account that exploits this observation might assume, first, that all coda consonants project a mora, except postconsonantal [t,s], and second, that there is a constraint *mmmm (19), undominated in Dutch, disallowing tetramoraic syllables. The first assumption is widely supported in work on Dutch phonology (see Booij 1995, 26). The second is evidently supported by the typology of quantity, trimoraic syllables already being rare. (19) *mmmm Syllables are maximally three moras long. As shown in (20), to represent wierp with a long vowel would mean violating undominated *mmmm. The form [firt] escapes shortening, because coronal [t] fails to project a mora. The correctness of this solution is underscored by examples like Ataturk [’ata"tyrk] ‘Ataturk’ and kirsch [kirS] ‘Kirsch’ (Paul Boersma, personal communication), which have short pre-[r] vowels, as expected under this analysis. (20) a. *
mmmm u i
rp
b.
mmm f i rt
While these facts already look pretty phonological, the clinching reason why it is not possible to translate the e¤ect of *mmmm into a phonetic implementation rule is that the short [i] of the past-tense forms survives an inflectional a‰xation process that removes the labial obstruent from the coda. The structure that arises can be compared with phonologically underived forms, as in (21), to show that long [i] appears in phonologically comparable contexts. [’Vir.p n] [’kir.k "xart] e e
(21) a. wierpen b. Kierkegaard
‘threwþplur’ ‘proper name’
This means that vowel quantity di¤erences are involved in lexical representations. The facts of (21) are consistent with a lexical phonology version of Optimality Theory (Kiparsky, forthcoming): wierp is subjected to the constraint grammar at the stem level, and its output is evaluated as the input of the constraint hierarchy at the word level, where the moraic structure specified at the stem level is preserved.
190
8.4
Carlos Gussenhoven
Dutch Stress
In this section, I consider the implications of our representation for a description of regular Dutch stress. The description essentially follows Nouveau 1994 and van Oostendorp 1997, but di¤ers from these in that it describes the foot structure of the whole word, not just of the foot with main stress. I show that regular main stress cannot be satisfactorily described if the prosodic status of the section of the word before the main stress is ignored. Crucially, the assumption must be made that, in Dutch, quantity sensitivity is restricted to the right-hand part of the word, and that before the main stress Weight-by-Position is not in e¤ect. 8.4.1
The Regular Stress Pattern
The basic facts about Dutch stress are summarized in (22) and (23) (van der Hulst 1984; Kager 1989; Nouveau 1994; van Oostendorp 1997; among others). The examples in (22) show the simplest case: main stress falls on the penult. (22) a. a.’xa.ta b. a.’mþn.da c. ’a.r n
‘Agatha’ ‘Amanda’ ‘Aaron’6
c
Penultimate stress systematically fails to appear in three situations. First, it does not appear when the word is monosyllabic, as in (23a). Second, it is systematically omitted when the word is minimally trisyllabic, has an open penult and a final closed syllable, and word stress is on the antepenult, as illustrated in (23b). Third, as shown in (23c), superheavy syllables, which can appear in word-final position only, attract main stress. (23) a. ’la ’kþt b. ’ma.ra."t n c. "þt.mi.’ral "le.di.’kþnt
‘drawer’ ‘cat’ ‘marathon’ (but: pa.’lm.bþÐ ‘Palembang’) ‘admiral’ ‘bed’
c
A treatment in OT was presented by Nouveau (1994, 184¤.) and modified by van Oostendorp (1997). The present analysis di¤ers from these earlier ones in that the representation of long vowels is bimoraic in stressed syllables, rather than monomoraic, but otherwise it essentially follows the older analysis in the characterization of regular main stress. All constraints are from Prince and Smolensky 1993 as well as McCarthy and Prince 1995 unless indicated otherwise. In section 8.3, we already took the e¤ect of RhythmTrochee (24) for granted. NonFin (25) is interpreted to ban main stress on the final syllable, as in Nouveau 1994, while Fs Right (26) will see to it that a foot with main stress is rightmost in the word.7
Vowel Duration, Syllable Quantity, and Stress in Dutch
191
(24) RHYTHMTROCHEE Feet are left-dominant. (25) NONFIN Main stress is not on the word-final syllable. (26) Fs RIGHT, or Align(Pw,Rt,Fs ,Rt) The right edge of the word coincides with the right edge of a strong foot. NoClash (27) forbids adjacent stressed syllables (i.e., adjacent foot heads). Highranking NoClash ensures that monosyllabic feet can only exist word-finally (see Gussenhoven 1993). (27) NOCLASH Foot heads are not adjacent. Finally, FootBin requires that feet are binary, either at the moraic level or at the syllable level. Violations are incurred by trisyllabic feet and monomoraic feet. In e¤ect, because high-ranking WSP/SWP already weeds out all monomoraic foot heads, FootBin’s only role in our analysis is to ban ternary feet. (28) FOOTBIN Feet are neither monomoraic nor trisyllabic. Cases like Agatha, with three open syllables, are derived straightforwardly: the winning candidate manages to obey all five relevant constraints. Candidate (29b) is ruled out by NoClash, and forces the prestress initial syllable to be unfooted—that is, directly attached to the Pword node. This is in accordance with the finding in Rietveld et al. (2004) that no quantity contrast exists in such syllables. Candidates (29c–f ) all founder on one of the other three constraints, as shown in tableau (29). (29)
axata
FootBin
NoClash
NonFin
WSP/SWP
Fs Right
+ a. a’(xa.ta) b. (a)’(xa.ta)
*!
c. ’(a.xa)ta
*!
d. ’(a.xa)(ta)
*!
e. (a.xa)’(ta) f. ’(a.xa.ta)
*! *!
The same result obtains if the penult is closed, as in [a.’mþn.da], the only di¤erence being that the equivalents of candidates (29c,d) are taken out by WSP as well as Fs Right. The ranking of WSP/SWP becomes critical in words with final closed
192
Carlos Gussenhoven
syllables. To stay with trisyllabic words, type (23b) (e.g., marathon) provides evidence that Dutch is quantity-sensitive, as shown in tableau (30). The winning candidate (30ai) violates Fs Right, which ranks below WSP/SWP: it alone incurs no WSP/ SWP violation, because both foot heads are bimoraic and the only weak syllable is monomoraic. Words like [pa.’lm.bþÐ] ‘Palembang’, which in addition have a closed penultimate syllable, incur WSP/SWP violations regardless of whether the main stress is on the penult or the antepenult (see candidates (30bi) and (30bii) in particular), and avoiding a violation in the penultimate syllable, as in candidate (30bi), is therefore pointless. The decision falls to Fs Right. Candidates (30aiv), (30biv), and (30bv) are excluded by NoClash. FtBin is omitted, but will reappear in section 8.5. (30) a.
maratn +
NoClash
WSP/SWP
i. ’(ma.ra)(tn)
Fs Right *
ii. ma ’(ra.tn)
*!
iii. ’(ma.ra)tn
*!
iv. (ma) ’(ra.tn) b.
NonFin
*!
!
palmbþÐ i. ’(pa.lm)(bþÐ)
*
+ ii. pa. ’(lm.bþÐ)
*!
*
iii. ’(pa.lm)bþÐ
*!*
iv. (pa) ’(lm.bþÐ)
*!
v. (pa) ’(lm)(bþÐ)
*!*
* *
Penultimate stress in disyllables with a closed final syllable is due to NoClash and NonFin, both of which rank above WSP/SWP. In tableau (31), [’a.r n] shows that the language prefers incurring a WSP violation to violating NonFin (see candidates (31a,b)). c
(31)
arn
NoClash
NonFin
+ a. ’(a.rn)
WSP/SWP
Fs Right
*
b. a.’(rn)
*!
c. ’(a).(rn)
*!
d. (a).’(rn)
*!
* *
Vowel Duration, Syllable Quantity, and Stress in Dutch
193
To continue with the examples in (23), monosyllabic words (see 23a) have stress because of the constraint Lex A Pwd (not given as a numbered item), which requires every morphological word to be footed. Superheavy syllables (see 23c) only occur word-finally. The relevant contextual constraint allowing this (‘‘No trimoraic syllable unless at right word edge’’) is here taken for granted. Superheavy syllables attract the word stress because trimoraic syllables are disfavored in weak positions, in the foot as well as in the word. This constraint, the Superheavy-to-Stress-Principle (32), can best be seen as belonging to the same family as WSP—that is, as a stricter version of the latter (see Prince 1990). Nouveau (1994) and van Oostendorp (1997) achieve the e¤ect of SHSP by postulating a degenerate or abstract final syllable for the final consonant, which makes these words escape the censures of NonFin; the effect is that of final consonant extrametricality. (32) SUPERHEAVY-TO-STRESS PRINCIPLE (SHSP) Trimoraic syllables are strong foot heads. Ranking SHSP above NonFin will have the desired e¤ect, as shown in tableau (33). (33)
kapital
NoClash
SHSP
+ a. (ka.pi).’(tal)
NonFin
WSP/SWP
Fs Right
*
b. ’(ka.pi).(tal)
*!
c. ka’(pi.tal)
*!
* **
In this section, the following generalizations were achieved: Lex A Pwd undominated: SHSP NonFin: NonFin WSP/SWP: NonFin Fs Right: WSP/SWP Fs Right:
8.5
Monosyllables have word stress. Final superheavy syllables have word stress. No word stress on a final closed syllable in disyllables. No word stress on a final closed syllable in trisyllables. Trisyllables with final closed syllable and open penult have antepenultimate stress (i.e., do not make a disyllabic final foot with heavy weak syllable).
Whole-Word Foot Structure
The description of Dutch stress presented in the previous section would appear to be seriously challenged once words with closed initial prestress syllables like armada [þr’mada] ‘armada’ are considered. If the attested candidate (34a) is disregarded for the moment, high-ranking NoClash would incorrectly characterize *[’þrma"da],
194
Carlos Gussenhoven
candidate (34e), as the optimal form. Candidate (34b) violates NoClash, and candidate (34c) violates NonFin. The choice between candidates (34d) and (34e) would be decided by WSP/SWP, which will not tolerate an unfooted bimoraic syllable. Importantly, to characterize the attested candidate (34a) as optimal, we must remove the foot from the initial syllable, so as to satisfy NoClash, and declare it monomoraic, so as to satisfy WSP/SWP.8 (34)
þrmada
NoClash
NonFin
WSP/SWP
Fs Right
m + a. þr.’(ma.da) b. (þr).’(ma.da) c. (þr.ma).’(da) mm d. þr.’(ma.da) e. ’(þr.ma)(da)
*! *! *! !*
The inevitable conclusion that closed word-initial prestress syllables are monomoraic in Dutch is supported by three independent arguments. The first is based on duration measurements. The best comparison we can make is with the first syllable of a word-initial weak foot. If, for example, the duration of [kþn] in cantorij ["kþnto’ri] ‘church choir’9 were to be shorter than that of [kþn] in kantoren [kþn’tor n], this would be strong evidence that kantoren has an unfooted first syllable, for if it were footed, it would either have to be equal in duration to the first syllable of cantorij, or, in view of the fact that there are widespread tendencies to shorten the foot head as more syllables occur in the foot, longer. In a production experiment with four speakers, it was consistently the case that the prestress syllable was shorter than the segmentally identical head of a disyllabic foot before the main stress (Elise Hofhuis, unpublished research). This is strong evidence that closed prestress syllables are not footed. The second argument was presented in Gussenhoven 1993 and concerns the fact that the Dutch intonational pattern known as the ‘‘chanted call,’’ which potentially produces a new pitch level on every postnuclear foot, systematically fails to produce such a pitch level on word-initial prestress syllables, whether closed or open, but may produce one on (binary) feet before the word stress. The di¤erence would be brought out if cantorij and kantoren were to be used as second members in compounds, as in nepcantorij and nepkantoren, where nep ‘fake’ would be accented. Whereas can- could begin a second pitch level, kan- could not. The third argument is distributional. Monomoricity of the word-initial prestress syllable predicts that one type of segment should be systematically excluded there, namely, the ‘‘mare
Vowel Duration, Syllable Quantity, and Stress in Dutch
195
ginal’’ long vowels (see table 8.1, third column). Rhymes consisting of strings of different segments can be accommodated, however marked multiple association of segments with a single mora may be, as shown by pneumatisch [pnœy.’ma.tis] ‘pneumatic’, which has a monomoraic diphthong in its first syllable. However, the combination of a single mora and a long vowel amounts to a contradiction in terms. Interestingly, while words like creˆmerie ["kr.m .’ri] ‘creamery’ can exist, in which [] occurs in a foot head, words like *[.’tþp.p ] are impossible (cf. e´tappe [e.’tþp.p ] ‘leg (sport)’). In conclusion, not only open, but also closed word-initial prestress syllables are monomoraic and unfooted, as argued in Gussenhoven 1993, which explains the fact that candidate (34a) is optimal. e
e
e
8.5.1
Accounting for Monomoraic Closed Syllables
If closed word-initial prestress syllables are monomoraic, Weight-by-Position (WbP)—the constraint that requires coda consonants to project a mora—must not be operative in that syllable. In fact, to the left of the main stress there is little or no evidence for the working of WbP at all (see Booij 1995, 106; Zonneveld et al. 1999, 504). Indeed, van der Hulst and Kooij (1992) proposed that main stress in Dutch results from quantity-sensitive footing from the right, but that the rest of the word is subsequently footed quantity-insensitively from the left.10 The weightlessness of syllables before the word stress can be observed in words that contain a string of three syllables prior to the main stress, of which the first is open and the second closed. If the second syllable attracted stress, the three-syllable stretch would be realized as an unfooted syllable followed by a binary foot. The words in (35) belie that expectation: the secondary stress is on the first syllable.11 (35) "aristo’cratisch ‘aristocratic’, "decompo’sitie ‘decomposition’, "enunci’atie ‘enunciation’, "evange’list ‘evangelist’, "emanci’patie ‘emancipation’, "identi’teit ‘identity’, "paterna’listisch ‘paternalistic’, "potenti’eel ‘potential’, "protestan’tisme ‘protestantism’, "tubercu’losis ‘tuberculosis’ Evidently, quantity sensitivity only obtains in the stretch from the main stress to the word end. Constraint WbP 0 (36), a version of WbP which is confined to the stretch beginning with the main stress, expresses this. (36) WbP 0 From the main stress onward, a coda consonant projects a mora. The pattern illustrated in (35) suggests that Dutch ranks high All-Ft-Left (37), a constraint that imposes a violation for every syllable by which the left edge of any foot fails to coincide with the left edge of the word. With Parse-s (38), which requires that syllables be parsed into feet, ranked above All-Ft-Left, exhaustive footing is achieved (Prince and Smolensky 1993). Tableau (39) shows this for enunciatie. The tableau dispenses with NoClash, Fs Right, SHSP, and NonFin, which all
196
Carlos Gussenhoven
relevant candidates satisfy: the point is that the low ranking of generic WbP allows the second syllable to escape the censure of WSP/SWP, which, had it been bimoraic, would have had to be a foot head, causing candidate (39d) to be optimal. As it is, the competition is decided by All-Ft-Left, which the winning candidate (39a) best satisfies; the fact that the syllable [nyn] is in a weak position and therefore violates WbP is no longer relevant to the outcome. (37) ALL-FT-LEFT, or Align(Ft,Lt;Pwd,Lt) The left edge of every foot coincides with the left edge of the Pword. (38) PARSE-s Syllables are parsed into feet. (39) en nsiasi
Foot Bin
+ a. (e.n n).si.’(a.si) b. (e.n n.si).’(a.si)
8.6
WbP 0
WSP/ SWP
Parse-s
All-FtLeft
WbP
*
***
*
***
* *
*!
c. e.n n.si.’(a.si)
**!*
***
d. e(n n.si).’(a.si)
*
***!*
Conclusion
The present account of the duration of the vowels of Dutch, the moraic structure of its syllables, and its foot structure shows that Dutch is an unexceptional ‘‘type B’’ language: truly long vowels and diphthongs (see columns 3 and 4 in table 8.1) attract stress in the same way that closed syllables do. The reason the language has been characterized as atypically requiring long vowels to be light and closed syllables to be heavy is that earlier researchers failed to accommodate the fact that the long vowels concerned only acquire bimoricity (and thus length) as a result of their being in a stressed location, as determined by the regular foot structure of the language. By attributing their length to the working of Stress-to-Weight, a phonetically realistic representation of vowel quantity has become possible. Moreover, there has been no need to represent [i,y,u], which are short even in stressed positions, as bimoraic, either underlyingly or on the surface, as earlier analyses were forced to do. They are only long, and bimoraic, when occurring before [r] in the same foot. Solutions to the special status of quantity sensitivity in Dutch provided by Lahiri and Koreman (1988) (separate representation of weight and length) and Kager (1989) (counting segments in the rhyme rather than moras) are thus rendered unnec-
Vowel Duration, Syllable Quantity, and Stress in Dutch
197
essary. My solution confirms the representation proposed by van Oostendorp (1995): underlyingly, the di¤erence between [þ] and [a], for instance, is captured by including [lax] in the representation of the former vowel, and quantity is not present in underlying representations. However, it disagrees with van Oostendorp’s analysis in requiring moraic structure to be present in the surface representation, a position enforced by the fact that quantity di¤erences are in part morphologically determined. Exceptionally long lax vowels (see column 3 in table 8.1) are accounted for by lexical specification of bimoricity, which is respected due to the high ranking of FaithMora. Interestingly, as pointed out by Sharon Inkelas, this analysis correctly predicts that there are no exceptionally short tense vowels (other than [i,y,u]): marking [a] as monomoraic will not prevent this vowel from being long in stressed positions, due to SWP/WSP. The relevance of moraic structure was further underscored by the foot structure of the word before the main stress foot, where crucially Weight-by-Position must be suspended. Failing to do so predicts incorrect main stress in words like ar’mada, which contain a closed word-initial prestress syllable. My analysis demonstrates that it may be unwise to present analyses of part of the data representing some phenomenon, like the data for main stress, and ignore other parts, like the data for secondary stress and vowel duration. Notes I thank the audiences at the universities of Groningen, Constance, Leiden, Nijmegen, and Amsterdam (Free University), where earlier versions of this chapter were presented in 1998 and 1999, for their useful comments, in particular Geert Booij, Mirco Ghini y, Haike Jacobs, Wouter Jansen, Aditi Lahiri, and Dominique Nouveau. Rene´ Kager, Aditi Lahiri, and Marc van Oostendorp commented on earlier drafts, thus allowing me to make various improvements. This chapter represents a revised version of the first half of a paper ‘‘Duration and Quantity in the Dutch Word’’ presented at the Conference on the Phonological Word in Berlin (ZAS), held October 24–26, 1997. The second half of the paper, on exceptional stress, will appear elsewhere. 1. The marginal nasal vowels (third column) are generally given without the length mark, but as Ton Broeders pointed out to me, they are in fact long. 2. Pace Prince (1990), who argues that SWP does not exist, and that the tendency is for the heads of trochees to shorten, as in Hayes 1987. Prince admits that what looks like the e¤ect of SWP may be seen in languages with dynamic stress. In fact, Dutch also has Trochaic Shortening, though it is variable and confined to disyllabic feet to the left of the main stress, as in ["po()li’tik] or substandardly ["p l ’tik] ‘politics’. Rene´ Kager points out that an e¤ect similar to SWP is obtained by HNuc, as shown in a handout to his Metrical Phonology class at the University of Utrecht of November 29, 1996: ‘‘The best main stress is the heaviest syllable.’’ As in van Oostendorp 1997 he assumes that length is coded by an underlying tenseness feature, while HNuc is held responsible for the bimoricity of main stressed vowels. His restriction to main stress equates the categorical occurrence of short vowels in unfooted syllables and in weak ec
198
Carlos Gussenhoven
syllables with syllables with secondary stress, leaving their (frequent) bimoricity unexplained. Also, no account of the quantity of [i,y,u] is given. 3. Trommelen and Zonneveld (1989) assume that [] is the surface form of underlying [i] before [r], observing that the latter combination does not occur and that all occurrences of [] before a full-voweled syllable precede [r]. However, there are counterexamples in both directions: theta [’tta] ‘theta’ and Teixeira (de Mattos) [tk’sira] (proper name). 4. Van Oostendorp also makes the complementary assumption that vowels that are not lax— that is, our ‘‘long’’ tense vowels—have no coda, in spite of the fact that they apparently do (e.g., laat ‘late’). This assumption is not made here. 5. Zonneveld et al. (1999, 500) make the notion of such a phonetic implementation rule explicit, but they reject the analysis and go on to defend the solution presented in Lahiri and Koreman 1988. 6. The biblical name is often pronounced [a’ar n], as pointed out to me by Marieke Polinder. c
7. Both Nouveau and van Oostendorp have a constraint that requires the main stress to be on the last syllable of the word, while the present analysis assumes a constraint aligning the main stressed foot with the right word edge. No crucial di¤erences in coverage follow from this. 8. Unlike what is the case in English, where a closed word-initial prestress syllable is footed if it does not represent a Latinate prefix (Chomsky and Halle 1968), in Dutch all prestress initial syllables, whether closed or open and whether or not they are prefixes, are unfooted (Gussenhoven 1993). 9. This word has a su‰xal [i], which attracts main stress; compare simplex [’sld "ri] ‘celery’, which has the regular pattern of words with final heavy and penultimate light syllables (see section 8.3). e
10. The notion of di¤erent directions for main stress assignment and secondary stress assignment already occurs in Booij 1983 with reference to Kenneth Pike’s work on Auca (though without di¤erentiation for quantity sensitivity). 11. There are words like this that, in addition to the pattern exemplified in (35), may alternatively be pronounced with secondary stress on the second, closed syllable karakteri’stiek ‘characteristic’, gerontolo’gie ‘gerontologie’, appendi’citis ‘appendicitis’, electrici’teit ‘electricity’, amonti’llado ‘amontillado’. A case for quantity sensitivity cannot easily be made on the basis of these words, all the more so since, as pointed out by Booij (1995, 106), there are words like piraterij [pi"rat ’ri], which show that a light second syllable may have secondary stress. That example also shows that the situation is more complex, since the reason for the location of the secondary stress on the second syllable in this word would seem to be that [a] is opener, more sonorous, than [i]; similarly, as pointed out by Haike Jacobs, the word caleidoscoop [ka"lid ’skop] ‘kaleidoscope’ illustrates that diphthongs can attract the foot in competition with [a] (cf. variable amonti’llado). e
e
9
Sympathy Meets Argentinian Spanish
Ellen M. Kaisse
9.1
Introduction
The fact that phonological processes are often opaque—that is, that they may appear to have applied when they should not have, or to have failed to apply when they should have—has occupied the attention of phonologists ever since Kiparsky’s (1971b, 1973e) landmark treatments of the phenomenon. As new theories of phonology emerge, they must be tested against opaque surface forms. A theory that cannot generate such forms must be judged inadequate. Opacity is a well-known problem for two-level, nonderivational theories such as standard Optimality Theory (Prince and Smolensky 1993). While a number of solutions have been proposed that attack this problem piecemeal, McCarthy (1999a) stands out as the first comprehensive attempt to treat virtually all opaque phenomena with the same single enrichment to standard OT: ‘‘sympathy.’’ Sympathy allows us to identify a form that is not underlying and use it as the basis for evaluating the faithfulness of potential outputs. McCarthy’s goal is to supplant all other mechanisms accounting for opacity, many of which are poorly understood and controversial—output-output constraints (Benua 1997), locally conjoined constraints (Smolensky 1995),1 stratal OT (Kiparsky 1998; McCarthy and Prince 1993; Orgun 1996; Rubach 2000), two-level constraints (Koskenniemi 1983; McCarthy 1996), and so forth. The strategy, which we must applaud, is to fully investigate and understand the characteristics of this one novel mechanism and push it to its full capacities while relying as much as possible on ‘‘familiar and indispensable OT constraints.’’ In addition to providing a careful explication of the characteristics of sympathy, McCarthy o¤ers a comparison between its predictions and those of standard derivational treatments of opacity using rule ordering. While the overlap in coverage, and indeed in concept, between the two theories is large, McCarthy is able to isolate one significant di¤erence—a class of cases that are ruled out by sympathy and thus predicted not to occur in natural languages, but that are easily expressed in a
200
Ellen M. Kaisse
derivational framework. He demonstrates that within sympathy theory, the following prediction holds: (1) If two notionally distinct processes . . . violate exactly the same faithfulness constraints, then they must always act together in rendering a third process opaque. (McCarthy 1999, 344) But is McCarthy correct in predicting that natural languages do not contain similar but distinct processes that respectively do and do not opacify a third process? I will argue that he is not. In this chapter, I will discuss data from Rı´o Negro Argentinian Spanish involving a sometimes-opaque interaction between a coda-based process and resyllabification into onset. I first treat the phenomenon in a multistratal derivational framework (Lexical Phonology), following the analysis in Kaisse 1999 and Harris and Kaisse 1999. Translated into monostratal OT using sympathy theory, this phenomenon is shown to meet the description of two processes violating the same faithfulness constraint, yet not to comply with McCarthy’s prediction. I will conclude that sympathy cannot, indeed, handle all cases of opacity. Because this example is emblematic of cases where intraword and interword phenomena diverge in their behavior, the best solution, I will argue, is to return to the derivational account based on Lexical Phonology’s division between lexical and postlexical strata, or perhaps to adopt a derivational OT model (Kiparsky 1999; McCarthy and Prince 1993; Rubach 2000) that allows us to maintain a division between lexical and postlexical evaluation. In general, cases violating the prediction in (1) may not be terribly hard to find (Levi 2000, for instance, contains another example, of a quite di¤erent character) and throw doubt on the ability of sympathy to adequately treat all cases of opacity within Optimality Theory. This chapter thus joins Odden’s (2000) paper in arguing that sympathy, despite its power, cannot be the sole solution to problems of opacity in OT. 9.2
S-Aspiration in Rı´o Negro Argentinian Spanish
Like many dialects of Spanish, Argentinian varieties realize /s/ as [h] in the coda of a syllable.2 The phenomenon, often referred to as ‘‘s-aspiration,’’ is illustrated below within morphemes and between a stem and su‰x, environments where virtually all dialects with aspiration show the same behavior. (2) ka.s-a gu.sa.n-o pa.s-ar dje.ses
‘house’ ‘worm’ ‘to pass’ ‘tens’
kah.ka.r-a guh.t-o pah.tel djeh.-mo
‘rind’ ‘pleasure’ ‘cake’ ‘tithe’
The rule for dialects like that of Rı´o Negro is given in (3), and is shown more formally in (4) to be a debuccalization after nonconsonantal segments (glides and
Sympathy Meets Argentinian Spanish
201
vowels), expressed as the spreading of [consonantal] to rhymal ‘‘/s/.’’ The loss of place can be seen as the automatic consequence of the loss of consonantality. (3) s ! h/ [cons] /
]s
(4)
The data presented in (2) is valid for the vast majority of dialects that exhibit the s-aspiration phenomenon. Preconsonantal /s/ is aspirated within a morpheme (kahkara) or before a consonant-initial su‰x (djeh-mo); within a morpheme or between a stem and a vowel-initial su‰x, prevocalic /s/ is not aspirated, since it is not in coda (gusan-o, kas-a). Things get more interesting when we look at other morphological contexts, where dialects diverge in their behavior. In this chapter, I concentrate on a little-studied dialect spoken in the northern Patagonian provinces of Neuque´n and Rı´o Negro, about 800 miles to the southwest of the capital, Buenos Aires. I will refer to this dialect as Rı´o Negro Argentinian, or RN for short, since the majority of my informants are from that province or right across the river in the city and province of Neuque´n. However, I suspect that this is the dialect of most of Patagonia. It has long been noted that s-aspiration in Spanish interacts in various ways with other phenomena in the phonology, among them resyllabification. Guitart (1979) and Harris (1983, 1993) have pointed out, for instance, that in some Caribbean dialects, word-final /s/ is realized as [h], even if it is not in a coda at all on the surface, but forms the onset of a following vowel-initial word: (5) tje.ne. heh.pa.sjo
tienes espacio
‘you have space’
The same is true for RN: word-final /s/ is realized as [h] not only before pause3 and preconsonantally, where its appearance is transparently due to its position in coda, but also before a word-initial vowel, where its appearance is opaque: (6) bwe.noh. di.ah loh. la.bjoh
‘good days’ ‘the lips’
bwe.no. haj.reh lo.h o.xoh
‘Buenos Aires’ ‘the eyes’
The prevocalic and prepausal aspiration is the source of much merriment among porten˜os—speakers of the prestigious dialect of Buenos Aires—among whom RN is commonly mocked as the loh ojoh (i.e., [lohoxoh]) dialect, a reference to the way
202
Ellen M. Kaisse
speakers from Rı´o Negro pronounce the phrase meaning ‘the eyes’. Porten˜os have transparent aspiration—they do not aspirate a prevocalic /s/ that ends up in an onset—and they do not aspirate before pause, so ‘the eyes’ emerges as [lo.so.xos] when uttered in isolation. In undergoing aspiration, prevocalic word-final /s/ in RN thus exemplifies the type of opacity in which a process appears to have occurred when its structural description is not present. The environment for the emergence of the first [h] in [bwe.no.haj.reh] or [lo.ho.xoh] is not present on the surface. Thus, aspiration is nonsurface-apparent, or overapplies, in McCarthy’s (1999) terminology. Within a derivational framework, such cases are handled via counterbleeding orders. Aspiration occurs at the word level, then resyllabification between words applies, removing [h] from its position in coda and providing the following syllable with an onset. (7) a. Word level
b. Phrase level
Initial syllabification Resyllabification within words Aspiration Resyllabification between words
/los oxos/ los. o.xos ——— loh. o.xoh lo. ho.xoh
The transparent Buenos Aires derivation, where aspiration is postlexical and does not occur before a pause, would proceed as in (8). (8) a. Word level b. Phrase level
Initial syllabification Resyllabification Resyllabification between words Aspiration
/los oxos/ los. o.xos ——— lo.so.xos ———
The important additional wrinkle added by RN (and not by other, better-known dialects) is this: between a prefix and the base4 to which it attaches, aspiration applies transparently. That is, if the base begins with a consonant, the s-final prefix des is realized as [h], but before a vowel, it is realized as [s]. (9) dehkargar desarmar dehkremada desinteresada
‘discharge’ ‘disarm’ ‘defatted, skim’ ‘disinterested’
Ample evidence that prefixes are syllabified on a separate cycle from their bases, then (speaking derivationally) resyllabified, is presented in Harris and Kaisse 1999. As described there, Argentinian Spanish has a process of glide consonantalization that realizes a high front vowel as [ˆ] syllable-initially. In the (a) forms below, there is no prefix, and the Cþglide sequence, whether underlying or derived from diphthongiza-
Sympathy Meets Argentinian Spanish
203
tion, is syllabified together at the beginning of the second syllable. The glide is not syllable-initial, and therefore does not consonantalize. However, in the (b) forms, the base is syllabified on its own. Its initial glide (again be it from diphthongization or underlying) emerges as a fricative and a preceding /s/ aspirates, while a preceding nasal agrees in place of articulation. (10) a. /deserto/ /boniato/ b. /des-elo/ /kon-iuxe/
! ! ! !
[de.sjer.to] [bo.nja.to] [deh.ˆe.lo] [ko.ˆu.xe]
‘desert’ ‘sweet potato’ ‘defrost’ ‘spouse’
The transparency of word-internal aspiration versus the aspiration between words is easily accounted for in a derivational theory using rule ordering and strata: (11) a. Word level
b. Phrase level
Syllabification Resyllabification between prefix and base Consonantalization Aspiration Resyllabification between words
A typical derivation would proceed as in (12): (12) a. Word level
b. Phrase level
Initial syllabification Resyllabification within words Aspiration Resyllabification between words
/desþarmar/ des.ar.mar de.sar.mar ——— ———
/los oxos/ los. o.xos ——— loh.o.xoh lo.ho.xoh
Before closing this section, we should address the question of whether /s/ is the underlying form for all or some [h] in RN, or whether the aspiration rule has been lexicalized, yielding /h/ in underlying representations. There are several reasons to opt for the analysis with all [h] coming from underlying /s/. First, unlike aspirating dialects spoken in the Caribbean, Argentinian Spanish does not have a phone [h] in any contexts other than those of s-aspiration. (The segment spelled ‘‘j’’ is pronounced [x], a voiceless velar fricative.) Thus, /h/ would be an ill-distributed phoneme, occurring in underlying representations only in preconsonantal position and as the last segment in a few morphemes. Second, most [h]’s alternate with [s] in Rı´o Negro. Prefixal ones appear as [s] before vowels; word-final ones are subject to variable realization as [s] before pause, as discussed in note 3. Root-final /s/, it is true, is usually realized exclusively as [s], since most su‰xes begin with vowels, and most roots require su‰xes, but the occasional consonant-initial su‰x also induces alternation.5 Furthermore, roots that do not require su‰xes when they appear as singular nouns, such as /lus/ ‘light’ and /rais/ ‘root’, alternate between a word-final [h] and [s] before a plural or
204
Ellen M. Kaisse
other vowel-initial a‰x ([luh] ‘light (n)’, [luseh] ‘lights’ (n.pl), [lusir] ‘to light up’). Even a mass noun like /ras/ ‘level’, which virtually never occurs with a su‰x and thus might appear to be a decent candidate for a root with an underlying /h/ that is realized consistently as [h], has a verbal derivative /a-ras-ar/ that shows a surface [s] throughout its paradigms. I conclude that the simplest analysis does not include /h/ in the inventory of any Argentinian dialect. I should also reassure the reader that the word los in the example los ojos is representative of all s-final words in RN, whether they are monosyllabic function words like los, or polysyllabic content words like the plural americanos ‘‘Americans’’ or the singular, s-final place name Caracas. All aspirate their final /s/ even if the next word begins with a vowel. 9.3
Treatment within Sympathy Theory
We have seen how the apparent overapplication of aspiration in RN is handled in a derivational framework with strata. Essentially, we have a ‘‘sandwich,’’ where wordinternal resyllabification precedes word-internal aspiration (aspiration being a lexical rule in RN), but where interword resyllabification follows and opacifies the lexical rule. The solution is unremarkable and entirely workable. Now we must ask how a phenomenon like this would be handled in Optimality Theory. We already know that opacity is generally problematic in OT, and that sympathy theory has been introduced as a method—hopefully the method—with which opacity will be treated. In this section, I show how the transparent portion of the syllabification-aspiration interaction would be handled in OT, why the opaque portion is problematic, and why the architecture of this example makes an insightful solution within sympathy theory unavailable as well. I begin by following McCarthy’s (1999) general method for recasting a phonological process in OT. The existence of a phonologically motivated process whereby an underlying form is not realized perfectly faithfully on the surface is captured in OT by ranking a markedness constraint above a faithfulness constraint. We have three processes to recast: resyllabification between prefix and base, resyllabification between words, and aspiration. Resyllabification processes are typically handled by ranking anchor constraints (a type of faithfulness constraint) lower than constraints favoring unmarked syllable structure. Anchor constraints enforce a matchup between the edges of morphological constituents, present in underlying representations, and the edges of phonological/ prosodic constituents in the output. Resyllabification blurs the boundaries between morphemes or words by placing a consonant belonging to the end of one morpheme at the beginning of a syllable containing the next morpheme so that the syllable and morpheme boundaries do not line up:
Sympathy Meets Argentinian Spanish
(13) a. CV.C]þ[V b. CV.C]a[V
205
Resyllabification between prefix and base Resyllabification between word and word
To start out, I will make the simplifying assumption that the morphological category that a prefix attaches to is always a root, though that root may, in addition, initiate a stem and/or a word. (I sketch a slightly less simplified morphological analysis at the end of this section.) One possible faithfulness constraint that is violated when a prefix-final consonant is resyllabified is (14): (14) ANCHOR (ROOT, s, LEFT) The beginning of a root corresponds with the beginning of a syllable. The markedness constraint that outranks (14) and forces its violation is Onset: (15) ONSET Syllables must have onsets. We do not need to add any more machinery to also translate our second process, resyllabification between words. It also results from mismatching syllable and morpheme boundaries so that a root (this time, a word-initial root) does not begin its own syllable. (Again, I temporarily make the simplifying assumption that all words, like all bases of prefixation, begin with roots.) From the fact that consonants do not always end up syllabified with all and only the material from their own morpheme, I induce the ranking in (16): (16) Onset Anchor Aspiration is roughly translated as follows:6 (17) *s]coda [s] may not appear in the coda of a syllable. This markedness constraint dominates a faithfulness constraint requiring that the place of an input consonant (coronal, in this case) be realized faithfully on the surface. (18) IDENT (PLACE) The place of an output consonant is identical to that of the corresponding input consonant. Our debuccalization process violates Ident (Place). Since it occurs nonetheless, we conclude (19) *s]coda Ident (Place) We are now able to construct the tableau needed to account for the transparent case, [desarmar] ‘to disarm’.
206
Ellen M. Kaisse
(20) Generation of transparent form /desþarmar/
Onset
*s]coda
*!
*
des.ar.mar
Anchor (Root, s, Left)
+ de.sar.mar
Ident (Place)
*
deh.ar.mar
*!
*
de.har.mar
*
*!
However, the opaque, interword result cannot be generated. In (21), following McCarthy 1999, the + symbol calls out a problematic candidate that actually wins the evaluation, while + marks the attested output we are trying to derive. In a successful analysis, the + and + candidates coincide and we only need the + symbol. I have simplified the tableau to only look at Ident violations for the first s, in los: (21) Attempt to generate opaque form without sympathy /los oxos/
Onset
*s]coda
los. o.xoh
*!
*
+ lo.so.xoh loh.o.xoh + lo.ho.xoh
Anchor (Root, s, Left)
Ident (Place)
* *!
* *
*!
The reason the opaque form [lo.ho.xoh] cannot win, as McCarthy explains clearly, is that it has a superset of the violations of the transparent [lo.so.xoh]. Both resyllabifying the /s/ and turning in into [h] yields a gratuitous faithfulness violation. To deal with opacity involving overapplication within sympathy theory, we find a member of the candidate set in which aspiration, the opaque, nonsurface apparent process, has applied but where it has not overapplied. That is, we look for a candidate where the environment for aspiration is present. This is the form [loh.o.xoh], which has the [h] from /s/ in coda position. This so-called sympathetic or flower candidate7 is often but not always identical to the form found at the intermediate stage of the serial derivation. If only faithfulness constraints can serve as selectors (McCarthy 1999), then the sympathetic candidate must more closely resemble the input form than does the actual, opaque output. It has failed to undergo the last process in the serial
Sympathy Meets Argentinian Spanish
207
derivation—the one that resulted in opacity. In the case at hand, the sympathetic candidate, [loh.o.xoh], is more faithful than the output because it does not violate the anchor constraint—it did not undergo phrasal resyllabification,8 and so the initial o of oxoh begins both the root and the syllable. This is not the case in the actual output, [lo.ho.xoh]. The faithfulness constraint that picks out the sympathy candidate and eliminates the opaque output temporarily is called the selector, and is marked in tableaux with the symbol ¶. The sympathy candidate is the most harmonic candidate that does not violate the selector. In the case at hand, the selector is Anchor (Root, s, Left). Now how do we choose the actual, opaque output over the transparent candidate? This is done via a high-ranking sympathy constraint enforcing partial identity with the sympathetic candidate. The sympathy constraint and sympathetic candidate are marked in tableaux with the flower symbol {. The sympathy constraint in our case is {Ident (Place), a high-ranked clone of the Ident (Place) that enforced resemblance to the input. {Ident (Place) requires that the output form have [h] by resemblance to the sympathetic candidate, rather than [s] by resemblance to the input. We are now ready to construct the tableau. A check mark calls out those candidates that obey the selector and may thus be the sympathy candidate. (22) Tableau with sympathy: Opaque form correctly chosen as optimal *s]coda
{Ident (Place)
¶Anchor (Root, s, Left)
Faithfulness
los. o.xoh
*
*
B
Transparent form
lo.so.xoh
Sympathy
{ loh.o.xoh
Opaque form
+ lo.ho.xoh
*!
*!
*
Ident (Place)
Onset *!
/los oxos/
B
*
*
*
However, we now cannot derive the transparent prefixþbase form we had no trouble with before the introduction of the selector, ¶Anchor (Root, s, Left).
208
Ellen M. Kaisse
(23) Tableau with sympathy: Transparent form now cannot be chosen *s]coda
{Ident (Place)
¶Anchor (Root, s, Left)
Faithfulness
des. ar.mar
*
*
B
Transparent form
+ de.sar.mar
Sympathy
{ deh.ar.mar
Opaque form
+ de.har.mar
*!
*!
*
Ident (Place)
Onset *!
/desþarmar/
B
*
*
*
We now see why McCarthy’s prediction in (1) holds true. If two processes violate the same faithfulness constraint, and that constraint is the selector for the sympathetic candidate, forms violating the selector can never be chosen as the sympathetic candidate themselves, and thus are forced to resemble the sympathy candidate in whatever way the sympathy constraint requires. Here two notionally distinct processes are intraword and interword resyllabification; the faithfulness constraint they both violate is Anchor (Root, s, Left); and the third process they act together in rendering opaque is aspiration. Both resyllabification of prefixes and resyllabification between words violate Anchor (Root, s, Left). Both are eliminated from the set of sympathetic candidates and both must therefore resemble the sympathy candidate, which has [h] in coda. This is correct for [lo.ho.xoh] but not for *[de.har.mar]. The tableau in (23) thus illustrates an unfortunate property of analyses using sympathy, dubbed ‘‘chaotic interaction’’ by Idsardi (1997). The addition of a sympathy constraint and a selector constraint to account for opaque forms can have unexpected results downstream, making transparent forms that were once easy to generate become nonoptimal because they are not sympathetic to the flower candidate. Of course, in this case, the result is not unexpected at all. McCarthy explicitly predicted that a situation like this would not be treatable within sympathy theory. Before moving on to consider alternative analyses of the Rı´o Negro case within OT, let’s partially remedy the oversimplification of Spanish morphology perpetrated at the beginning of this section. We said that the category to which des- attaches is a root, and also that all words begin with roots. That gave us an easy way to unify the anchor constraint that both forms of resyllabification violate. However, des- actually
Sympathy Meets Argentinian Spanish
209
attaches only to independently occurring words, so it is not obvious that the category to its right, its base, should be called a root. It should more properly be called a word. A form like [de.sar.mar] is in violation of Anchor (Word, s, Left) then, armar being a word. And turning to the interword cases, we should recognize that not every word begins with a root. Many words begin with prefixes: des-armar ‘disarm’, intro-duccio´n ‘introduction’, en-chilada ‘stu¤ed tortilla seasoned with chilies’. Now if a word begins with a vowel-initial prefix, interword resyllabification will not violate Anchor-Root. The phrase las enchiladas, pronounced [la.hen.tSi.la.dah] in Rı´o Negro, has a mismatch between the beginning of the word enchilada and the beginning of a syllable, but that mismatch involves the anchoring of the word, not of its root, chil-. Thus to be accurate, the faithfulness constraint that both intraword and interword resyllabification violate is Anchor (Word, s, Left).9 But let us not get too involved in the details of how best to formulate the anchor constraints. As the next section will show, there are many versions of these constraints and of the sympathy tableaux that I could have constructed—indeed there are far too many. 9.4
Ways Out
The most obvious way to save the sympathy analysis is to split the relevant faithfulness condition into two. Then our two processes, resyllabification within words and resyllabification between words, will not violate the same faithfulness constraint and will escape the prediction in (1). To achieve the sought-after result, let us refer to the right edge of the morpheme containing the /s/ rather than to the left edge of the following morpheme: (24) a. ANCHOR (PREFIX, s, RIGHT) The right edge of a prefix coincides with the right edge of a syllable. b. ANCHOR (WORD, s, RIGHT) The right edge of a word corresponds with the right edge of a syllable. The tableaux in (25) show how the correct forms can now be generated choosing Anchor-Word (24b) as the selector. The correct form, [lo.ho.xoh], has a word/ syllable edge mismatch and thus is eliminated as the sympathy candidate by the selector; it therefore must match the actual sympathy candidate, [loh.o.xoh], in having the apparently gratuitous [h]. However, [de.sar.mar] has no violations of the selector. It is itself, then, the most harmonic candidate that does not violate the selector, and emerges as both the sympathy candidate and the optimal output. (I mark it as both the sympathy and the transparent candidate—it is both the optimal form and, vacuously, the sympathy candidate.) It matches itself with respect to the sympathy constraint, {Ident (Place), and thus emerges with [s].
210
Ellen M. Kaisse
(25) a. Anchor (Word, s, Right) as selector: Interword {Ident (Place)
¶Anchor (Word, s, Right)
los. o.xoh
*
*
B
Transparent form
lo.so.xoh
Sympathy
{ loh.o.xoh
Opaque form
+ lo.ho.xoh
*!
*
*!
Ident (Place)
*s]coda
Faithfulness
Anchor (Prefix, s, Right)
Onset *!
/los oxos/
B
*
*
*
b. Anchor (Word, s, Right) as selector: Prefix þ stem
Transparent form
+ de.sar.mar
Sympathy
{ deh.ar.mar
Opaque form
de.har.mar
B B
*!
Ident (Place)
*
Anchor (Prefix, s, Right)
des.ar.mar
¶Anchor (Word, s, Right)
*s]coda
Faithfulness
{Ident (Place)
Onset *!
/desþarmar/
*
B
*!
B
* * *
*
Mechanically, the solution works. But the underlying problem with this line of attack is that it attempts to replicate the insights of a lexical/postlexical distinction in the absence of an elaborated theory of phonology-morphology interactions in OT. What kind of anchor constraints can there be? We do not know. One can imagine other anchor constraints I might have devised, but again, we do not know whether they would accurately reflect true generalizations about the interaction of morphology and phonology. Perhaps worse, as Bill Idsardi has pointed out to me, there is
Sympathy Meets Argentinian Spanish
211
no way to limit which anchor constraints can be chosen as selectors. In the case at hand, what if we chose Anchor-Prefix instead of Anchor-Word as the selector? This would allow us to generate a language in which word-internal aspiration (or the process of our choice) is opaque but aspiration between words is transparent—a situation that does not occur in natural languages, to my knowledge. The tableau in (26) shows how easy it would be to generate the unnatural result—happily nonreplicable in Lexical Phonology—where [de.har.mar] and [lo.so.xoh] emerge from the evaluation. (26) Anchor (Prefix, s, Right) as selector
Transparent form
+ lo.so.xoh
Sympathy
{ loh.o.xoh
Opaque form
B *
*!
*
lo.ho.xoh
Ident (Place)
*
¶Anchor (Prefix, s, Right)
los.o.xoh
Anchor (Word, s, Right)
*s]coda
Faithfulness
{Ident (Place)
Onset *!
/los oxos/
*!
*
B B
*
B
*
/desþarmar/ Faithfulness
des.ar.mar
Transparent form
de.sar.mar
Sympathy
{ deh.ar.mar
Opaque form
+ de.har.mar
*!
*!
*
*
B
*!
* B
*
*
*
We must conclude that sympathy cannot insightfully handle the case of Rı´o Negro aspiration. What of other mechanisms in the armamentarium of OT? One could handle this case with an output-output condition, requiring that a word ending in /s/ show [h] before a vowel so as to match the [h] that the word has in isolation. However, McCarthy (1999) rightly points out the output-output conditions only work when
212
Ellen M. Kaisse
there is a base form on which to model the opaque form. There happens to be one in this case, but often there is not. Output-output conditions therefore cannot function as the sole means of capturing opaque phenomena in OT, and so are no better than sympathy if our goal is to unify the treatment of opacity with a single, wellunderstood addition to the theory. In general, output-output conditions are not really well understood, because, like earlier theories dealing with analogy, they trade in the ill-understood question of what can serve as a base and under what circumstances bases influence other forms. We might stop at this point and insist that the Harris and Kaisse (1999) treatment of aspiration in RN is the one to stick with. It has the virtue of being explicit, and of meshing correctly with a wealth of other facts about the phonology of Argentinian Spanish, such as glide-consonant alternations, stress assignment, and syllabification. However, there is one variant on OT that can handle this particular case with similar insight to the derivational treatment given by Harris and Kaisse. That is Derivational Optimality Theory (Rubach 2000) and the philosophically similar Stratal OT (Orgun 1996; Kiparsky 1999). Both these theories reject sympathy as a method for dealing with opacity. Instead they select output candidates via a series of evaluations, each evaluation corresponding, more or less, to a stratum of the Lexical Phonology model.10 In the case at hand, we will first derive [de.sar.mar], [loh], and [o.xoh] as outputs of the word-level evaluation; we will then take these outputs as the new inputs to the phrase level. (27) Word-level derivational OT tableau a. /desþarmar/ des.ar.mar
Onset
*s]coda
*!
*
+ de.sar.mar deh.ar.mar
Ident (Place)
* *!
*
de.har.mar b.
Anchor (Root, s, Left)
*
*!
/los/ los
*!
+ loh
*
/oxos/ o.xos
*
+ o.xoh
*
*! *
Sympathy Meets Argentinian Spanish
213
Nothing further happens to [de.sar.mar] at the phrasal level, and it emerges as the optimal form. However, at the phrasal level, the outputs [loh] and [o.xoh] are strung together and it becomes possible to fulfill the onset constraint by violating Anchor (Root). The opacity emerges naturally from the tableau, without sympathy or other additional devices, because the input form for los already has /h/—there is no gratuitous faithfulness violation, as there was in the parallel OT derivation. (28) Phrase-level derivational OT tableau /loh.o.xoh/ loh.o.xoh
Onset
lo.so.xoh
Anchor (Root, s, Left)
Ident (Place)
*!
+ lo.ho.xoh los.o.xoh
*s]coda
* *!
* *
*!
Successful as this translation to derivational OT is, we must not forget McCarthy’s (1999) observation: not all cases of opacity result from well-motivated stratal segregation. There are numerous cases in the derivational literature that at least appear to involve opaque interactions within a single stratum. Therefore, the nascent theory of derivational OT has a burden to bear—it must show that all genuine cases of opacity result from stratal organization. Should this prove impossible—and my guess is that it will—we are back to the conclusion that a unified treatment for opacity is not currently available within Optimality Theory. Notes An earlier version of this chapter was presented at the First North American Phonology Conference. This chapter fleshes out a brief suggestion made in Harris and Kaisse 1999 concerning the di‰culties of dealing with the Rı´o Negro dialect of Argentinian Spanish within OT in general and within sympathy theory in particular. Thanks to Sharon Hargus, Jim Harris, Bill Idsardi, Susannah Levi, and other members of the audience at NAPHCI. The following abbreviations occur throughout the text: pl ¼ place; cor ¼ coronal; n ¼ noun; n.pl ¼ plural noun. 1. Actually, McCarthy reserves locally conjoined constraints as the treatment of choice for one kind of opacity—counterfeeding on focus, as originally proposed by Kirchner (1996). This type of opacity results from ‘‘chain shifts,’’ where y becomes z, and x becomes y, but new y’s do not go on to become z’s. The reason for this exception to the general sympathy treatment is not worked out in either McCarthy 1999 or 2003. 2. This description oversimplifies the phenomenon, which is described at greater length in Kaisse 1996, 1999, as well as in Harris and Kaisse 1999 and references therein. Among other details not directly relevant here, the phenomenon is variable in many dialects, can involve
214
Ellen M. Kaisse
near or total loss of any aspiration, may require a preceding nonconsonantal segment, and so forth. 3. Realization as [h] before pause is variable in RN—sometimes word-final /s/ emerges as [s] in that context. This variation is almost certainly due to the normative e¤ect of the prestige dialect, that of Buenos Aires, where /s/ is always [s] before pause. 4. The base to which a prefix attaches may be a root, a stem, or a word. des- attaches to words. The prefix /kon-/ illustrated in (10) is attaching to a stem; iuxe does not appear in isolation. 5. For instance, the morpheme /dies/ ‘ten’ appears in [djeh.þmar] ‘decimate’, [djeh.þmo] ‘tithe’, [djeh] ‘ten’, and [dje.sþes] ‘tens’. 6. I am merely aiming for a quick, schematic view of this process. A more insightful translation of aspiration would involve a member of the family of Coda Conditions—constraints on what may appear in coda and, in particular, the inability of codas to license place on their own (Coda Condition—*Place). Indeed, as discussed in Kaisse 1996, the placeless h often takes on the place of a following velar consonant ([pax.kwal] for ‘Pascual’), indicating that a shared dorsal place node is permitted in the coda. 7. The term ‘‘sympathetic’’ candidate is confusing, in my opinion, because it refers to the flower candidate that the output, opaque form sympathizes with, not to the output form that exhibits sympathy by coming to resemble the flower candidate. One might more logically refer to the flower candidate as the pathetic form—the one that compels sympathy. 8. In cases involving anchoring, resemblance to the input is at one remove. The input does not contain syllabification per se, but it does contain discrete morphological elements. The sympathetic candidate resembles the input by better matching syllable edges to input morpheme edges. 9. Though not every prefix attaches to words, des- is virtually the only s-final prefix in Spanish, and it is a word-level prefix. (The rare and unproductive prefix dis- does not attach to any vowel-initial roots, as far as I can determine.) 10. The stratal OT material by Kiparsky available to me at this writing is more concerned with the stem/word distinction than the lexical/postlexical (word/phrase) distinction. The work of Rubach is thus a more direct source for the proposal advanced here. The ranking of constraints can vary between the strata. However, reranking does not appear to be necessary in this example.
10
Vowel Length, Cyclicity, and Output-Output Correspondence
Cemil Orhan Orgun
10.1
Introduction
For about three decades, cyclicity and level ordering enjoyed a privileged theoretical position in Generative Phonology in accounting for synchronic alternations reflecting morphological relatedness between forms (Chomsky, Halle, et al. 1956; Chomsky and Halle 1968; Kean 1974; Allen 1978; Kiparsky 1982b; Mohanan 1982; etc.). Recently, however, alternatives to cyclicity based on the notion of paradigm uniformity (the idea that morphologically related forms are grammatically constrained to be phonologically similar) have come to the forefront of phonological research. Particularly popular among these is the Output-Output Correspondence approach within Optimality Theory (Benua 1997; Burzio 1994; Kenstowicz 1995; McCarthy 1996; Steriade 1997). A number of researchers, however (Dolbey 1996, Dolbey and Orgun 1996, Booij 1996, Kiparsky, forthcoming), have argued that traditional interleaving (cyclicity or level ordering) is empirically or theoretically superior to paradigm uniformity as a means for accounting for synchronic alternations. In this chapter, I present data from English and Turkish vowel length alternations that seem to favor the traditional interleaving approach (cyclicity and level ordering) over Output-Output Correspondence. The English and Turkish data that I discuss show that interleaving e¤ects do not necessarily result from the interaction between related word forms, as paradigm uniformity posits. Rather, individual morphological constructions may use either a lexical stem or a word as the base to which the given morphological process is to be applied. Even when a given root is free (can surface as a word without further morphology), a morphological construction may use its stem form rather than its surface (word) form. Such cases crucially fail to exhibit output correspondence e¤ects.
216
10.2 10.2.1
Cemil Orhan Orgun
English Vowel Length Alternations Data
In addition to well-known morphophonemic alternations involving length as well as quality, such as Closed Syllable Shortening (wide [aI] versus width [I]) and Trisyllabic Shortening (opaque [eI] versus opacity [æ]), English vowels undergo allophonic variation in length alone. Introductory textbooks often portray this alternation in deceptively simple terms, for example, as ‘‘the general rule in English whereby a vowel is longer if the following consonant is a voiced sound’’ (Kenstowicz and Kisseberth 1979, 299). I present a fuller set of facts (based on work with several consultants, all native speakers of American English from California), which prove to have significant implications for the phonology-morphology interface. The basic facts in American English are that vowels are long before tautosyllabic voiced consonants and in final position, short otherwise. (1) a. Long high hide
[a:I] [a:I]
b. Short hybrid height
[aI] [aI]
Vowel length interacts with morphological structure in interesting ways. The relevant phenomena will be illustrated in this chapter for vowels in open syllables. However, an examination of compounds with vowel-final first members shows that the vowel in question is short in some compounds and long in others. In my work with native speaker consultants, I have found that the exact set of compounds with short vowels varies from speaker to speaker. However, all of my speakers have some compounds with long vowels and some with short. The particular examples presented in this chapter reflect the speech of one consultant. (2) a. Long dry mop pie pan lie detector
[a:I] [a:I] [a:I]
b. Short high chair high school skyscraper
[aI] [aI] [aI]
Why are long vowels found in nonfinal position in the first members of the compounds in (2a)? Alternatively, if the final vowel of the first member of a compound is in ‘‘final position’’ for the purposes of vowel length, why are short vowels found in (2b)? 10.2.2
A Lexical Phonological Account
In the framework of Lexical Phonology (Pesetsky 1979; Kiparsky 1982b,d; Mohanan 1982), an immediate solution is available. The lexicon is divided into at least two strata, corresponding to the traditional Levels 1 and 2. Vowel lengthening
Vowel Length, Cyclicity, and Output-Output Correspondence
217
applies on Level 2 but not Level 1. For the purposes of this chapter, we may assume Level 1 to be the ‘‘stem’’ level and Level 2 to be the ‘‘word’’ level, although a full account of English phonology and morphology would be much more complicated. Compounds may be formed at either level. (3) Level 1 Level 2
Phonology
Morphology
No vowel lengthening Vowel lengthening
‘‘Stem’’ compounding ‘‘Word’’ compounding
Typical derivations show how the full range of vowel length alternations receive a satisfactory account in this system: (4) Level 1 Level 2 Surface
high
high chair
dry
dry mop
high [a:I] — high [a:I] — high [a:I]
high chair [aI] — — — high chair [aI]
— — dry [a:I] — dry [a:I]
— — dry [a:I] dry mop [a:I] dry mop [a:I]
Note that we need to assume, following Inkelas 1989, a prea‰xal cycle on Level 2. Whether we have a prea‰xal (‘‘root’’) cycle on Level 1 or not is immaterial to the discussion here, as Level 1 has no vowel lengthening. I assume one for the sake of consistency. While the theoretical details of this account may or may not be acceptable to a given linguist, there is an obvious insight that this approach reflects. Words but not stems need to end in long vowels in American English. Stems as well as words may be compounded. When two stems compound, the final vowel of the first stem is not subject to lengthening because it is not a word. When two words are compounded, the final vowel of the first word is of course word final. Accordingly, it lengthens. The crucial insight here is that morphological constructions may call for a stem form or a word form. 10.2.3
Output-Output Correspondence
The Output-Output Correspondence approach to almost-allophonic alternations thwarted by morphological structure assumes that markedness constraints in the relevant dimension outrank competing faithfulness constraints (Benua 1996 for Sundanese; Kenstowicz 1995 for Italian). In English, this would mean ranking constraints that govern the distribution of long and short vowels below constraints requiring any
218
Cemil Orhan Orgun
underlying vowel length to be preserved. This way, regardless of whether vowels are underlyingly long or short, we are guaranteed a long vowel in high (word-final lengthening) and hide (lengthening before a voiced coda) and a short one in hybrid and height (neither word-final lengthening nor lengthening before a voiced coda are applicable). In order to account for the long vowel in pie pan, we posit an OutputOutput Correspondence constraint that requires vowel length to be the same in morphologically related words. This constraint outranks the markedness constraints governing the distribution of vowel length. (5) Output-Output faithfulness to vowel length a.
Base: none Input: hide
OO-Faith
hide
V-Length
IO-Faith
*!
(*)
+ hi:de b.
(*)
Base: pie, pan Input: pie pan pie pan + pie: pan
*!
(*) *
(*)
According to a common assumption in Optimality Theory called Richness of the Base (Prince and Smolensky 1993), there are no direct constraints on underlying forms. Thus, forms such as pie and hide may have underlyingly long or short vowels. Depending on which is chosen, there will be di¤erent IO-Faith violation patterns (hence the parenthesized asterisks in the tableau). However, since IO-Faith is never the ruling constraint, this does not matter. Since underlying length contrasts do not have an e¤ect on the surface forms, there will not be a vowel length contrast in English. However, OO-Faith constraints outrank well-formedness constraints. Therefore, we find unexpectedly long vowels in some morphologically derived forms when the form’s base contains a long vowel, as in (5b). But then what about compounds with short vowels? Why do we have a short vowel in high chair? Is high not a base here? Why not? Would one have to abandon the claim that there is a uniform family of Output-Output Correspondence constraints and make up special constraints for each morphological construction? Level 1 (stem) compounds would then correspond to their bases via Level 1 correspondence constraints, and Level 2 (word) compounds would correspond to their bases via Level 2 correspondence constraints. This move, of course, would amount to abandoning the enterprise: we could no longer say that cyclic phonological e¤ects are
Vowel Length, Cyclicity, and Output-Output Correspondence
219
due to general paradigm uniformity constraints. In fact, Benua (1996) has remarked that Output-Output Correspondence is superior to phonology-morphology interleaving precisely because it uses one family of general correspondence constraints rather than construction- (or stratum-) specific constraints. But if the facts make using one family of constraints untenable, that argument is lost. Thus positing constructionspecific constraints is simply a thinly veiled reconstruction of Lexical Phonology. One might as well admit that Lexical Phonology is the right approach. It is clear that Output-Output constraints do not help. Do they have to hurt though? If one could make up an even higher-ranking constraint to circumvent the undesirable e¤ects of Output-Output Correspondence, one would simply be admitting that Output-Output Correspondence does not account for English vowel length. This would not be too high a price to pay. After all, why should Output-Output Correspondence account for everything? 10.2.4
A Prosodic Account
The basis of the Lexical Phonological account was a distinction between stems and words implemented by means of level ordering. By translating this morphological account into prosodic terms, we can account for the alternations by surface wellformedness constraints alone. We assume that ‘‘stem compounds’’ consist of a single prosodic word (w), while ‘‘word compounds’’ contain two. (6) Bare forms Stem compound Word compound
[hide]w , [pie]w [high chair]w [[pie]w [pan]w ]w
With these representations, everything follows from surface prosodic structure. Interleaving and Output-Output Correspondence are simply irrelevant. However, this account in turn su¤ers from a serious defect. How are the appropriate prosodic structures assigned to these forms? In a framework that includes interleaving, this problem is not present. According to Inkelas 1989, prosodic structure assignment takes place cyclically. Thus, pie pan, pie, and pan undergo a cycle on which they are assigned word status. On the next cycle, they are compounded and the whole form is assigned (recursive) word status. Short vowel compounds such as high chair must undergo compounding at the stem level (Level 1). As a result, the individual members high and chair are not prosodic words and therefore not subject to vowel lengthening. The compound high chair is assigned word structure on Level 2 as a whole, where vowel lengthening does not apply to high since only prosodic word-final vowels are subject to lengthening. Of course, if prosodic structure assignment is done cyclically, there is little point in insisting that vowel length is determined noncyclically. Proponents of Output-Output Correspondence would therefore need to find a principled way of assigning the
220
Cemil Orhan Orgun
appropriate prosodic structures to these English forms noncyclically, perhaps by using Alignment (McCarthy and Prince 1993a). 10.3
Judgment Day
While it so far seems that the prosodic account could work if principled means of assigning the right structure could be found and the argument against Output-Output Correspondence could vanish, there are two kinds of evidence in favor of the interleaving account. The first is morphological and language internal; the second is phonological and crosslinguistic. 10.3.1
Stem versus Word Compounds
It has long been known that there is a compounding construction in English that uses irregular plural nouns (Marchand 1969). For nouns with regular plurals, the ‘‘singular’’ (that is, bare stem) form is used instead. (7) Singular mouse rat
Plural mice rats
Compound mice-infested rat-infested
There are also compounds using singular forms of regular or irregular nouns (e.g., mouse trap); these are irrelevant to our purposes. Less widely acknowledged is the fact that some compounds do use regular plural forms. (8) Singular form in compound sport jacket pants pocket
Regular plural form in compound sports pages pant press1
What is the di¤erence between these two kinds of compounds? A natural answer is that the kind in (7) is a stem-compounding construction, while the kind in (8) is word compounding. Irregular plurals are lexically listed stems, while regular plurals are inflected words (with a word-level su‰x). The theoretical tools used in this account are identical to those used in the Lexical Phonology account of vowel length alternations. Thus, Lexical Phonology unifies two apparently separate phenomena: regular versus irregular plurals in compounds, on the one hand, and long versus short vowels in su‰xed forms and compounds, on the other. To the extent that such unification of apparently disparate phenomena is an indicator of theoretical success, the Lexical Phonology account is supported. 10.3.2
Turkish Vowel Length
As Inkelas and Orgun (1995) show, Turkish has vowel length alternations that are quite similar to those that I have presented for English, although the environment
Vowel Length, Cyclicity, and Output-Output Correspondence
221
for vowel lengthening is more restricted. Due to a minimal size condition of two moras, vowels of CV roots lengthen. Long vowels are found in una‰xed as well as a‰xed forms of these roots: (9) Turkish orthography do doyu
IPA (with morpheme breaks) do: do:-ju
Gloss ‘musical note C’ ‘C-accusative’
However, these roots surface with short vowels in some compounds: (10) do bemol do mino¨r
do-bemol j do-minør
‘C flat’ ‘C sharp’
Within Lexical Phonology, these facts receive a treatment parallel to that of their English counterpart. Two levels are needed; we may call them stem and word. Shortvowel compounds are formed at the stem level.2 Vowels lengthen at the word level, where a‰xes are also added. As in English, we need to assume a prea‰xal cycle. (11) Surface (orthography in parentheses) Underlying Stem level Word level
do: (do)
do:ju (doyu)
dobemol j (do bemol )
do do do:
do do do:-ju
do, bemol j do-bemol j —
Although Turkish seems to be similar to English, there turns out to be a crucial twist. It can easily be shown that the prosodic approach that could come to the rescue of Output-Output Correspondence in English is not available in Turkish. Let us first see what a prosodic analysis of Turkish vowel lengthening would look like. On a parallel analysis, we could assume that short-vowel compounds contain one prosodic word, while a‰xed forms contain an embedded prosodic word: (12) [do:]w [[do:]w ju]w [dobemol j ]w Given the generalization that vowel lengthening applies in order to bring prosodic words to the required minimum size of two moras, it is crucial that the accusative su‰x attach to prosodic words. If we found evidence that the base of a‰xation is not a prosodic word, the motivation for vowel lengthening would be lost, and we would have to conclude that the prosodic analysis does not work. Since the prosodic analysis is crucial to saving Output-Output Correspondence, we would have to conclude that Output-Output Correspondence faces an insurmountable challenge. Lexical Phonology would emerge as the clear winner.
222
Cemil Orhan Orgun
In Turkish, coda clusters must obey Sonority Sequencing. Underlying clusters that do not obey Sonority Sequencing are repaired by vowel epenthesis. If a vowel-initial su‰x follows a root ending in such a consonant cluster, such as /resm/ ‘picture’, epenthesis does not apply, as the second consonant in the cluster is able to syllabify as an onset. (13) resim resmi
resim resm-i
‘picture’ ‘picture-accusative’
These forms show that the root cannot be a prosodic word in su‰xed forms: (14) *[[resm]w i]w Since [resm] is not a possible prosodic word, the embedded prosodic word representation is ill formed. The only escape would seem to be to claim that prosodic words are redefined according to syllable boundaries: (15) [[res]w mi]w Introducing a slight phonology-morphology mismatch of this kind would not in itself be an unreasonable theoretical move. Unfortunately, however, it would have the undesirable consequence of predicting vowel length where none is found. In particular, what happens when a su‰x is added to a CVC root, such as /kin/ ‘hatred’? The prosodic word account predicts vowel lengthening in this case. Since we concluded that prosodic words are redefined according to syllable boundaries, we must have the structure in (16): (16) [[ki]w ni]w In fact, no vowel lengthening takes place in CVC roots followed by vowel-initial a‰xes: (17) kin kini
kin kin-i
‘hatred’ ‘hatred-accusative’
However, the representation in (16), corresponding to the data in (17), is ill formed! If vowel lengthening takes place in (12) because of prosodic minimality, it must take place in (16). The prosodic word in (16) is clearly monomoraic. In addition, the representations in (16) and (17) make the unfortunate prediction that the first syllable in these forms will bear stress, as every prosodic word bears stress in Turkish. In fact, stress is final in these forms. Prosodic restructuring would present a similar problem in English.3 If syllables are redefined in a compound like rice paper, a complex onset [sp] will be formed, and the [p] of paper should therefore be unaspirated. This, however, is not the case; the [p] is fully aspirated.
Vowel Length, Cyclicity, and Output-Output Correspondence
223
The inevitable conclusion is that the domain for vowel lengthening is a morphological stem, not a prosodic constituent of any sort. Therefore, the proper account of the vowel lengthening data must be based on interleaved phonology-morphology interaction in the style of Lexical Phonology. 10.4
Conclusions
One of the major observations in phonology-morphology interaction is that bound roots are never subject to cyclic phonological e¤ects. Output-Output Correspondence captures this by observing that bound roots are by definition not pronounced as independent words and therefore cannot stand in correspondence. The flip side of this is that free morphemes are possible independent words and therefore stand in correspondence with related forms and give rise to cyclic phonological e¤ects. In this chapter, I have shown that being a free morpheme does not guarantee cyclic e¤ects. The grammar is free to choose the root form or the word form for such a morpheme. Even within a language, this choice might vary from construction to construction. Thus, English Level 1 a‰xation and compounding call for the root high [haI], while Level 2 a‰xation and compounding call for the word high [ha:I]. Output-Output Correspondence has no additional insight to o¤er here; the best we can do is try (somewhat unsuccessfully) to circumvent problems caused by the framework by using other mechanisms, such as ad hoc assignment of prosodic structure. Lexical Phonology not only handles the facts with ease, it also relates them to independent morphological facts having to do with presence or absence of inflectional morphology in compounds. Notes Many thanks to Bob Ladd, who first brought the English vowel length phenomenon to my attention. Without his observation of the short vowel in high school, this chapter would have never come into existence. I also thank Sharon Inkelas, Carlos Gussenhoven, Andrew Dolbey, and the attendees of a UC Berkeley colloquium and a Stanford phonology workshop for valuable comments. 1. Attested in a mail-order catalog. 2. Inkelas and Orgun 1995 in fact shows the need for at least four, while Inkelas and Orgun 1998 argues for five. 3. I thank Carlos Gussenhoven for bringing this fact to my attention.
11
Level Ordering in Nuuchahnulth
John Stonham
11.1
Introduction
One of the primary contributions of Lexical Phonology and Morphology (henceforth LPM) to linguistic theory is the use of stratal separation to explain behavior shared by only a subset of the elements of the lexicon within a language. The ordering of strata allows linguists to apply phonological rules to the result of combining one group of morphemes with another while excluding other nonparticipating groups of morphemes. Perhaps the best-known example is that of the application of the English stress rules at only the first stratum, discussed, inter alios, by Kiparsky (1982b, 1982d, 1985) and Mohanan (1982, 1986), but for every language one might expect to find particular phenomena indicative of such stratal membership. For any given language, the number of strata will depend on factors such as the complexity and interaction of the phonology and morphology, but a guiding principle should be that the minimum number of levels necessary for explaining the facts is to be preferred.1 Assuming this principle and the fundamental tenets of LPM, we will now proceed to the structure of the word in the Nuuchahnulth language of Vancouver Island, Canada, examining various phenomena that provide clear evidence of the need to separate morphemes into strata in the grammar. There are a number of areas of Nuuchahnulth grammar where there appear to be distinct di¤erences in the application of rules. Some of these include
Glottalization e¤ects Delabialization t´ ! 2 deletion Stress assignment Bound root/free form distinctions Su‰x combining forms
We will consider each of these cases in turn, paying particular attention to the possible arguments for di¤erent domains of application for phonological and
226
John Stonham
morphological rules in Nuuchahnulth and, in the end, arguing for a two-strata model of the morphology. 11.2
Derivation/Aspect versus Inflection
There are strong arguments for dividing Nuuchahnulth morphology into two sets, as first suggested in Sapir and Swadesh (1939) (henceforth S&S): (i) basically derivational and aspectual a‰xes, labeled ‘‘formative’’; and (ii) ‘‘incremental’’ a‰xes, as described by S&S.3 In fact, no great weight should be placed on the significance of these basically descriptive terms, but the segregation of morphemes into di¤erent groups is the crucial issue, as will be demonstrated in what follows. It should also be noted that, while in large part the analysis of morphemes that is presented here will be the same as that of S&S, there will also be some di¤erences based on more recent reanalysis of the data. The distinction between formative and incremental morphemes is relatively robust in Nuuchahnulth. The distribution is marked most clearly in the 500 or more su‰xes of the language, where typically the order of derivational su‰xes is relatively free, being based on the semantics of the construction, but su‰xes occurring after the derivational and aspectual ones are rigidly ordered, following the order presented in table 11.1.4 Additional criteria for the division include the category-changing nature and lexicality of derivational su‰xes and the paradigmaticity of incremental morphemes. As Table 11.1
Incremental su‰xes Nonparadigmatic
Paradigmatic
1
2
3
4
5
6
7
8
9
10
11
m’in¡ pl
is dim
q¡ mw
’aqt´ intent
’ap caus
’at´ now
’at pass
mit past
cnd
qt
infer
’at inal(ienable) uk/ak poss(essive) ’aa¡ irr(ealis)
ind(icative) rel(ative) sub(ordinate) dub(itative) rel(ative) dub(itative) abs(olutive) inter(rogative) purp(osive) indef(inite) rel(ative) pres(ent) imp(erative) fut(ure) imp(erative)
Level Ordering in Nuuchahnulth
227
for linear order, derivational su‰xes invariably appear closer to the root. As an example of the productivity of a‰xation in Nuuchahnulth, note the following.5 (1) aaat´qim´¡timji´m’in¡aqt´itsu dup-dup-at´-qim´-¡ta-ma´-’i´ k -m’in¡-aqt´-m‡itsu rep-suf-two-units-on foot[r]-moving-in house-pl-intent-2p.ind ‘You carry two dollars on your feet.’ As can be seen from this example, a‰xation is very productive in Nuuchahnulth. Reduplication, which may perform such tasks as marking certain types of lexical suffixes and indicating distributivity and repetitive aspect, is also common, as demonstrated by this example. The innermost copy is a concomitant of the su‰x marked [r] (reduplication), and the second indicates the repetitive aspect of the form. There is no prefixation or compounding (but see Stonham 1999a). One complication arises with respect to aspectual markers, which typically appear between derivation and inflection, as demonstrated by the following examples. (2) a. tSutSuk w ainmasnakSiat´ma tSutSuk w ainma-’as-na‡k w -Sit´ k -at´-ma‡ b. hask w a´Sit´itwin hask w a´-Sit´ k -mit-w‡in
‘they had people go about the village inviting’ invite-outside-have-mom-now-3s.ind ‘it was nearly . . . ’ nearly-mom-past-3s.qt
However, further evidence demonstrates that it is possible for such aspect markers to occur intertwined with the derivational su‰xes, although never with inflection, as in the examples in (3). (3) a. ¡isSit´mapt ¡is-Sit´-mapt b. ´utSnakSit´maiqstuat´a¡ ´utSna‡k w -Sit´-maiqstut´ k -’at´ -ma‡¡
‘species of plant’ bleed-mom- . . . plant ‘I want to get married’ marry-mom-want to (mom)-now1s.ind
This evidence supports S&S’s decision to combine derivation and aspect into a single class of ‘‘formative’’ morphemes and will be used here to support the decision to combine these elements on a single stratum. A further issue is the status of certain su‰xes presented in table 11.1. Su‰xes marking plural nouns, -m’in¡, and diminutives, -is, may arguably be considered derivational, although we will see that they fit better into the category of stratum 2 a‰xes in any case. This highlights the common observation that a distinction based solely upon the derivation/inflection division often runs into di‰culties that are not encountered by a simple stratal distinction (e.g., Hargus 1989; Inkelas and Orgun 1995).
228
11.3
John Stonham
Arguments for Stratal Separation
In what follows, various phenomena will be presented that serve to clarify the issue of stratal segregation. Some of these cases are phonological in nature, while others may more accurately be described as morphological. Some indicate a need for a separate level of (quasi)inflectional morphology, while others provide evidence for a combined level of derivation and aspect. All go to support the hypothesis that Nuuchahnulth morphology must be divided into two distinct strata. 11.3.1
Glottalization
Glottalization involves the interaction of a specific subset of su‰xes with a preceding segment. The outcome will vary depending on the segment involved but typically results in the change of stops to glottalized stops, fricatives to laryngealized glides, and sonorants to their laryngealized counterparts. The chart in (4) illustrates the various changes that may occur.6 (4) Plain p t ts t´ tS k k w q qw s´S ´w xw ¯w ¡w mnwj x¯¡
!
Glottalized p’ t’ ts’ t´’ tS’ k’ k’ w j’ w’ m’ n’ w’ j’ —
Basically, the glottalization is ‘‘triggered’’ by some characteristic of the su‰x that S&S represent by a glottalizing mark /’/ at the beginning of the appropriate su‰x. Such su‰xes are reasonably common in the language and instances of the process are frequent. They all begin with a vowel, although not all vowel-initial su‰xes trigger glottalization. How one should appropriately distinguish those su‰xes that do from those that do not cause this process is perhaps an issue for further debate elsewhere. However one does it, it will obviously be necessary to make some distinction between the two categories of su‰xes, and for the purposes of this chapter the convention used in S&S will be employed. Examples of su‰xes that trigger glottalization (5a) and those that do not (5b) are provided below. (5) a. -’um -’it´ -’at´ b. -im -it´ -at´
‘ . . . on rocks’ ‘invite . . . ; go for . . . ’ now ‘ . . . thing’ mom caus ‘ . . . receptacle (in Deer’s speech)’
Level Ordering in Nuuchahnulth
229
Examples of how these su‰xes combine with preceding elements are provided below. The su‰xes in (6a) are all glottalizing su‰xes, while those in (6b) are not. (6) a.
i. wiSk’um wiSk-’um ii. hita¡t’it´Siat´ hita-¡t-’it´-Sit´ k -’at´ iii. qa¡nak’at´7 qa¡-na‡k w k -’at´ b. i. titinkum8 dup-ti-n’uk w -im ii. ajimkit´qas aja-mik w -it´ k -qa‡s iii. k w an’ux w at´ k w an’ux w -at´
‘angry on the rocks’ angry- . . . on rocks ‘they started out of the woods’ loc-exit woods-go for[l]-mom-now ‘he was dead now’ dead-have . . . -now ‘handwiper’ suf-wipe-in hand[r]- . . . thing ‘may I be a getter of many . . . ’ many-getter of . . . -mom caus-1s.sub ‘chamber-pot (in Deer’s speech)’ urinate (female)-receptacle
As can be seen from these examples, the result of the combination of a glottalizing su‰x with a preceding stop/a¤ricate-final base is the creation of a glottalized version of the consonant. Where there is no available appropriate version, as in the case of fricatives, which have no glottalized counterparts in Nuuchahnulth, the nearest possible consonant is provided, namely, for labialized fricatives, a laryngealized labial glide, /w’/, and for coronal fricatives, a laryngealized coronal glide, /j’/. The data follow the pattern below. (7) a. t´iw’inap t´ix w -’in k -’ap b. atuj’isap’anitin a:tuS-’i‡s k -’ap-’at-mit-min c. mat´itsuaj’in mat´-wits-jua´-’in d. Saw’aat´ Sax w -’aa k -’at´
‘he laughed at him’ laugh-sound of . . . -caus ‘they gave us deer meat to eat’ deer-eat-caus-pass-past-1p.ind ‘had been seen wearing it around his head’ tied-around head[l]-perceive-treated as ‘they fled’ flee-do for . . . sake-now
When the base ends in a consonant that is not a possible candidate for glottalization, such as /x ¯ ¡/, then no change occurs to the fricative and // is inserted. (8) a. cici¡aqt´mapt dup-ci¡-’aqt´-mapt b. tSijanuxat¡ tSijanux-’at¡
‘crab-apple wood’ pl-sour-inside- . . . plant ‘the Chiyaanuh Tribe’ Chiyaanuh- . . . tribe
Where the base ends in a vowel, the result is similar to that above, with the insertion of a glottal stop between the two vowels preventing coalescence.
230
John Stonham
(9) a. ink’ w aap ink w -’a‡a k -’ap b. taktSiinim tak-tSi k -’i‡nim
‘he made a fire’ fire-in hearth-caus ‘let us each do . . . ’ each-along with . . . -1p>3obj.fut
The most interesting aspect of this process with respect to stratal separation arises with regard to the fricatives. When a fricative constitutes part of an incremental morpheme, the concomitant glottalization does not occur. At this later level of morphology, there will be no change to produce laryngealized sonorants. S&S observe that ‘‘while glottalizing formative su‰xes . . . change voiceless fricatives to glottalized semivowels, glottalizing incremental su‰xes a¤ect only non-glottalized stops’’ (p. 236). Thus, the rule applies in the same fashion to stops and a¤ricates at either level, but with fricatives there is a distinct di¤erence. Rather than posit systematic exceptions to the application of the rule, we may capture the regularities of the process by the introduction of rule di¤erences associated with stratal membership: in short, the phonological rule of glottalization occurs at both strata with somewhat di¤erent environments (see Kiparsky 1985). As a demonstration of this distinction, examine the results of the application of the su‰x -’at´ ‘now’, a clear example of an incremental su‰x (position 6 in table 11.1), compared with the su‰x -’a‡a ‘on the rocks’, an obviously derivational morpheme. (10) a. i. hiniasat´ hina-ias k -’at´ ii. ujua´at´ u-jua´ k -’at´ b. i. nawaj’aa nawa‡s-’a‡a ii. hij’aaq¡aqt´itq hi´-’a‡a k -q¡-’aqt´-i‡tq
‘leaving the house’ loc-go outside-now ‘she noticed . . . ’ ref-perceive-now ‘while sitting idly chatting on the rocks’ sit idly chatting-on rocks ‘to the rocky point where they were to . . . ’ loc-on rocks-being-intent-3s.rel
From this evidence we can clearly see the distinction between the treatment of fricatives within the so-called formative component and in the incremental component. This suggests the need for two strata within the morphology of the language, with two alternative formulations of the environment for the rule of glottalization, depending on the stratum where the rule takes place. 11.3.2
Delabialization
One of the conditioning factors for the loss of labialization of rounded velars and uvulars is the boundary between the two strata, at which point delabialization occurs. This suggests a type of ‘‘exit feature’’ that applies at the end of stratum 1.
Level Ordering in Nuuchahnulth
231
This would be possible only if there were a stratal distinction in order to provide the means to apply the rule. In Nuuchahnulth there is an underlying distinction between round and unround velar and uvular obstruents as shown below.9 (11) ka ki´ qi t´’a¯ ak
‘measure’ ‘lift up canoe’ ‘shift position’ ‘vertically flat’ ‘dog-salmon’
k w a k w i´ q wi t´ax w ak w
‘move backward’ ‘blow spray’ ‘that which’ ‘carved’ ‘whittle’
When rounded consonants appear at the end of the first stratum, they will be delabialized, necessitating a distinction between levels in the grammar. This is illustrated by the examples in (12). (12) a. qa¡nak’at´ qa¡-na‡k w k -’at´ b. hisik’atwin his-i‡k w k -’at-w‡in
‘he was dead now’ dead-have . . . -now ‘they passed by’ loc-go along . . . -pass-3s.qt
Compare the examples above with those in (13) in which the labialization remains intact. (13) a. ¡ajutSink’ w aa ¡aju-ink w -’a‡a b. upa´nak w inj’ap’at upa´-na‡k w -inj’u k -’ap-’at c. jasik’ w as jas-i‡k w -’as
‘Ten-together-on-Rocks (man’s name)’ ten-together-on rocks ‘after giving me my club’ club-have . . . -left behind-caus-pass ‘someone going along . . . ’ there-go along . . . -outside
In all of the cases with labialization, the immediately following su‰x is derivational (13), while in those in (12) it is incremental, indicating that the rule applies at the end of stratum 1. Clearly, delabialization may be used as a test of stratal separation, providing further evidence of the bistratal nature of Nuuchahnulth morphology. 11.3.3
t´ ?
The application of this rule, which merges a lateral a¤ricate with a following glottalizing su‰x resulting in a glottal stop rather than the anticipated glottalized lateral a¤ricate /t´’/, is again confined to the boundary of the two strata, further supporting a di¤erence between the two levels of Nuuchahnulth morphology. The /t´/ in these cases is typically part of a morpheme associated with the momentaneous, inceptive, or graduative aspects and is therefore attached prior to the
232
John Stonham
attachment of the triggering morphemes /-’ap/, /-’at´/ and /-’at/, and /-’a¡/, which are found in positions 5, 6, and 7 in table 11.1. The examples in (14) demonstrate the application of this rule. (14) a. ¡at’uk w iap ¡a-at’ut´-Sit´ k -’ap b. naj’aqnakSiat´ naj’aqnak-Sit´ k -’at´ c. ´u¡tqapiatqa ´u¡-tqa-’ipit´ k -’at-qa‡ d. ak’uatwin ak’ut´ k -’at-w‡in e. ¡isi´itSia¡at´quwin ¡i-si´a-itSit´ k -’a¡-’at´-qu-w‡in
‘making it (burn) bright’ complete-on fire-mom-caus ‘she gave birth’ give birth-mom-now ‘it was leaking in the house’ flow-underneath-in house-pass-sub ‘it was borrowed’ borrow (mom)-pass-3s.qt ‘they could not get through’ unable to-do . . . -inc-irr-now-cnd3s.qt
Furthermore, the rule only applies in the environment of these glottalizing su‰xes, not anywhere else within the level of inflection, as demonstrated in (15). (15) a. tS’itasSit´m’in¡at´ tS’itas-Sit´ k -m’in¡-’at´ b. wiktaqSit´isqu wik-taqSit´ k -is-qu c. haukSit´aqt´’at´min hauk-Sit´ k -aqt´-’at´-min
‘they felt cold’ feel cold-mom-pl-now ‘as if it had never been’ not-before . . . ing-dim-cnd ‘we shall eat’ eat-mom-intent-now-1p.ind
If the su‰x is another glottalizing inflectional su‰x, then the result will be a glottalized lateral a¤ricate, [t´], as in the following cases. (16) a. hiSimj’up’at´’itS hiSimj’awup k -’at´-’i‡tS b. ts’iqSiat´’i ts’iq-Sit´ k -’at´-’i‡ c. hawiat´’inim hawit´ k -’at´-’i‡nim
‘get together now!’ assemble-now-2p.imp ‘now sing your sacred chants’ sing secret chant-mom-now-2s.imp ‘you can now stop asking’ finish[m]-now-1p>3o.fut
It is interesting to note that this process does not apply to those su‰xes that trigger it, specifically -’at´, as demonstrated by the following examples. (17) a. waat´’at wa k -’at´-’at b. m’aak w aap’at´’at m’a-ak w a‡ k -’ap-’at´-’at
‘it was said’ say-now-pass ‘he got chewed to pieces’ bite-in pieces-caus-now-pass
Level Ordering in Nuuchahnulth
233
Since the su‰xes triggering this rule are clearly within the domain of stratum 2, occupying positions 5, 6, and 7 in table 11.1, and the a¤ected elements are found within the derivational/aspectual categories, this process occurs precisely at the boundary of stratum 1 and stratum 2, again suggesting some sort of ‘‘exit feature’’ as suggested for the rule of delabialization above. 11.3.4
Deletion
The deletion of // occurs when a base ending with a fricative comes in contact with a su‰x beginning with a glottal stop //. Such su‰xes should not be confused with those that trigger glottalization, represented by an initial /-’/, as discussed in section 11.3.1. In such situations, there are two possible outcomes that depend on the category of su‰x involved. If the su‰x is a stratum 1 element, then the glottal stop will disappear. If, on the other hand, the su‰x occurs at stratum 2, the glottal stop will remain following the fricative. Thus, this rule too is subject to domain distinctions. Take, for instance, the case of the su‰x /-aqsu´/ ‘at the mouth’. In the following examples one can see the outcome of this derivational su‰x combining with basefinal obstruents and fricatives. In the case of stops (18a), the // remains, whereas in the case of fricatives (18b), it disappears. i. k w isitaqsu´ k w ist-it-aqsu´ ii. apaqsu´ ap-aqsu´ b. i. ts’a¡aqsu´ ts’a¡-aqsu´ ii. ¡asaqsu´ ¡as-aqsu´ iii. hi´aqsuas¡at´t´a hi´-aqsu´-’as k -q¡-’at´-t´a
(18) a.
‘other side of the mouth’ di¤erent-at side-at mouth ‘mouth’ loc-at mouth ‘Tough-mouthed (man’s name)’ tough-at mouth ‘loud voice’ loud-at mouth ‘they were outside the door again’ loc-at opening-outside-being-nowagain
Compare this with the results of an inflectional su‰x, /-aqt´/, intent, in similar environments and we can see that the results are the same in both cases: stops (19a) and fricative-final bases (19b). (19) a.
i. ts’awakaqt´qa ts’awa‡ k -ak-aqt´-qa‡ ii. hinipaqt´qa hina-ij’ip k -aqt´-qa‡ iii. mup’itaqt´in mu-p’it k -aqt´-min
‘one would . . . ’ one-nom-intent-sub ‘that they find . . . ’ loc-obtain . . . -intent-sub ‘we would sing four times’ four- . . . times-intent-1p.ind
234
John Stonham
b.
i. tsiasaqt´qa tsias k -aqt´-qa‡ ii. tS’itSm’uq w i´aqt´qa tS’itSm’u‡-i´ k -aqt´-qa‡ iii. hi´¡aqt´itq hi´ k -q¡-aqt´-i‡tq
‘he was going to talk marriage’ woo-intent-sub ‘he got ready to exact sacrifices’ scapegoat-make . . . -intent-sub ‘the place he was . . . -ing’ loc-being-intent-3s.rel
As can be seen by these data, there is no // deletion at the later level. This rule of deletion operates only at the earliest level of the morphology, further supporting the notion of stratal separation and a stratum 1 phonological rule of // deletion. 11.3.5
The Domain of Nuuchahnulth Stress
In Nuuchahnulth, the domain of primary stress is the first two syllables of the word (i.e., the first foot). Weight, in the form of vowel quantity distinctions or the presence of a nasal in the coda, is a determining factor in stress assignment, but closed syllables containing a short vowel followed by one or more nonnasal consonants count as light for the purposes of stress assignment, as demonstrated in (20e). (20) a. b. c. d. e. f.
tSu´Suk w it´ ku´¡sinqin´ap n’a´csat´ ¡ı´j’i¡ qa¡na´k’at´ haja´akSiat´
‘he begins to suspect’ ‘he always causes a hole in the side’ ‘he sees now’ ‘he is after blood’ ‘someone died now’ ‘he did not know now’
As can be seen from these data, the stress always appears on either the first or second vowel of the word, regardless of the presence of long vowels farther on in the word. It is also the leftmost of two heavy syllables that is assigned the stress in the case where there are two in the first foot (e.g., 20b, 20c, 20d). The question arises as to what happens if the word contains no long vowels. The following data present the results. (21) a. t’a´n’anak b. na´j’aqak c. witSup’at´’isak
‘have a child’ ‘baby’ ‘come and cause me now to sleep!’
Here, stress is assigned to the leftmost vowel. Note that this result obtains even in the case where there is a longer vowel later on in the word. (22) a. ta´nakmitSiat´ b. hu´atsatSit´aqt´ma c. q w a´jats’iktaqim´
‘they turned now into mosquitoes’ ‘he will come back’ ‘wolf band’
Level Ordering in Nuuchahnulth
235
Much work remains to be done in working out the intricacies of Nuuchahnulth stress assignment and the interaction of the various factors that determine it, but one interesting observation is that stress assignment appears to be limited to the realm of derivation and aspect, that is, at stratum 1. In all of the previous examples, stress has appeared on either the root or on a derivational su‰x immediately following it, so it is worth asking the question, what happens when an incremental su‰x is in competition for stress assignment? Naturally, this is not a common occurrence given the prolific nature of Nuuchahnulth a‰xation, but it occasionally arises that a free root with a short vowel may be followed directly by one of those su‰xes described in table 11.1. Examples in (23) illustrate this situation. (23) a. u´¡qa u¡ k -qa‡ b. wı´kqu wik k -qu c. u´¡ma u¡ k -ma‡
‘it was’ be-sub ‘it might not . . . ’ not-3.cnd ‘it is . . . ’ be-3.ind
These examples support the proposal that stress assignment in Nuuchahnulth occurs before the attachment of incremental a‰xes, that is, at stratum 1. Since stress is already assigned, the weight of the vowels of a stratum 2 su‰x has no bearing on the outcome of stress assignment. It should, however, be noted that a definitive treatment of Nuuchahnulth stress awaits completion and there are some seemingly contradictory examples in the original data gathered by Sapir (Stonham 1999b, 60–61; Sapir field notes). Thus, while the claim that stress assignment occurs at an early stratum holds some weight, the evidence must still be viewed with some healthy skepticism. 11.3.6
Root/Free Form Distinctions
A further distinction encountered in the data is the morphological consideration of root allomorphy. In Nuuchahnulth, there is a high degree of root allomorphy that distinguishes between stratum 1 and stratum 2 a‰xation. In the following table the first column provides the free form that occurs with stratum 2 su‰xes and on its own. The second column indicates the basic word class of the root, and the third column provides a rough gloss. The final column shows the form of the root that is found in combination with derivational and aspectual suffixes. The examples in (24d) are of borrowed words, mainly from English. (24)
Free Form a. tsapin tsinwa
Class n n
Meaning ‘sawbill duck’ ‘mussel species’
Bound Form tsapitq tsitq
236
John Stonham
tS’aq’uniS hax w inma¯ k’i´anus t´uk w ana b. mak’it q w a¡in tik w in c. pua´ t´au d. tSikinis miSin tipin pinis
n n n n v v v adj adj n n n n
‘edible berry species’ ‘wren’ ‘furseal’ ‘Wolf Ritual’ ‘play with shells, dolls’ ‘fish with prong-spear’ ‘bake in sand, ashes’ ‘sleepy’ ‘other’ ‘chicken’ ‘boarding school’ ‘table’ ‘apple’
tS’aq’utq hax w itq k’i´atq t´uk w atq mak’itq q w a¡itq tik w atq puitq t´au tSikitq miSitq tipatq pitq
The choice of bound versus free form is not a property of any particular word class or a historical remnant, and it may even be applied to borrowed words.10 The bound allomorph is found to exist only at the level of derivational/aspectual morphology and never at the later level, suggesting that here again a distinction exists between the two domains. Take, for instance, the following examples where the first in each pair involves the free form and the second the bound form. (25) a. i. tsapin tsapin ii. tsapitqin’akj’ak tsapin-i‡n’ak w -j’ak b. i. t´uk w anat´ t´uk w ana k -’at´ ii. t´uk w atquwis t´uk w ana-u´ w -’is c. i. t´atmaptuki t´atmapt k -uk-i‡ ii. t´atmaqan’u´i t´atmapt-a‡n’u´ k -i‡
‘sawbill duck’ sawbill duck ‘(I) have a sawbill dance’ sawbill duck-imitate . . . in dance-device ‘I had given a Wolf Ritual’ Wolf Ritual-now ‘Wolf Ritual Beach’ Wolf Ritual- . . . place-on beach ‘his yew tree’ yew-poss-def ‘the one of yew’ yew-along . . . [l]-def
These are not the only instances of bound roots however. Bound roots may occur in a variety of shapes, including examples such as the following, where the first instance in each pair (a) is the free form and the second (b) is the bound form. Note that in some cases the only di¤erence between the forms is vowel quantity. (26) tS ’itSm’u / tS ’itSm’uq ‘scapegoat’ a. i. tS’itSm’uaqt´’ap ‘he was forced to make sacrifices’ tS’itSm’u k -aqt´-’ap scapegoat-intent-caus
Level Ordering in Nuuchahnulth
ii. tS’itSm’ui tS’itSm’u k -i‡ b. i. tS’itSm’uqtsk w i tS’itSm’u-tsk w i‡ ii. tS’itSm’uqi´ tS’itSm’u-i´ (27) ha mut / hamut ‘bone’ a. i. hamuti hamut k -i‡ ii. hamutuk t´’unim hamut k -uk t´’unim b. i. hamuttsk w i i¡tup hamut -tsk w i‡ i¡tup ii. hamut¡tin tS’itu´i hamut -¡tin tS’itu´ -i‡ (28) ma¡t’i / ma¡t’iq ‘house’ a. i. ma¡t’imiti ma¡t’i k -mit-i‡ ii. ma¡t’im’in¡i ma¡t’i k -m’in¡-i‡ b. i. ma¡t’iqapu´isi ma¡t’i-apu´ k -is-i‡ ii. ma¡t’iqtsk w ak ma¡t’i-tsk w i‡ k -’ak
237
‘the scapegoat’ scapegoat-def ‘what is sacrificed’ scapegoat-remains of . . . ‘make a scapegoat’ scapegoat-make . . . ‘the bones’ bone-def ‘an elk’s bone’ bone-poss elk ‘remains of whale bones’ bone-remains of . . . whale ‘the warclubs were made of bone’ bone-made of . . . warclub-def ‘the former house’ house-past-def ‘the houses’ house-pl-def ‘the little pretended house’ house-imitating . . . -dim-def ‘the remains of his house’ house-remains of . . . -poss
As can be seen from the examples above, the (a) forms involve free roots combining with stratum 2 su‰xes and the (b) forms involve bound roots combining with stratum 1 derivational and aspectual su‰xes. There is a variety of forms that bound roots may take, diverging from the free forms to a greater or lesser extent, but the su‰xes found to occur with bound forms are consistent throughout the paradigm, suggesting some close a‰liation. Such su‰xes include derivational and aspectual suffixes but no members of the previously presented incremental category. What best accounts for this a‰liation is the principle of a common stratal membership, bound root allomorphs being restricted to stratum 1, while free forms occur at stratum 2. 11.3.7
Suffix Combining Forms
Akin to the bound roots of the previous section are a set of su‰x forms that occur in a special shape only in combination with derivational and aspectual su‰xes. Since
238
John Stonham
‘‘bound su‰x’’ would be somewhat incongruous, the term ‘‘combining form,’’ first introduced by S&S, will be employed to refer to this class of a‰xes. (29) -sj’aqsti / -sj’aqstiq ‘main or leading . . . ’ a. i. uksj’aqsti ‘the leaders were . . . ’ u-sj’aqsti ref-leading . . . ii. jaqsj’aqstiitqa´ ‘the one who was the head’ jaq w -sj’aqsti k -i‡tq-a´ rel-leading . . . -3s.rel-pl b. i. uksj’aqstiqistasi ‘I was chief of the crew’ u-sj’aqsti-ista k -si‡ ref-leading- . . . persons in canoe1s.abs ii. jaqsj’aqstiqij’ipitqa´ ‘the most valuable thing they had got’ jaq w -sj’aqsti-ij’ip k -i‡tq-a´ rel-leading-obtain-3.rel-pl (30) -api / -apiq ‘up in the air’ a. i. kitsapii kitsuk-api k -i‡ ii. t´’ikapim’in¡ t´’ik-api k -m’in¡ b. i. ts’isapiqSiat´ ts’is-api-Sit´ k -a’t´ ii. sutS’aqim´ma´apiqij’ip sutS’a-qim´-ma´-api-ij’ip
‘a fallen tree that was slanting upward’ log-in air[l]-def ‘they were placed stretching upward’ have hands in position-in air[l]-pl ‘send a telegram’ rope, line-in air[l]-mom-now ‘they got five (birds) in the air’ five-units-move about-in air[l]-obtain
The (a) examples above show what happens when the su‰x completes the stratum 1 stem, whether or not further a‰xation at later levels occurs, while the (b) examples demonstrate the results within stratum 1, where a ‘‘combining’’ form is necessary for further derivational su‰xation. Again, the su‰xes that concatenate with these combining forms are the same ones that occur with the bound roots discussed in the previous section. As such, it seems obvious that there is a clear distinction made in the grammar between stratum 1 combining forms of certain su‰xes and a stem-final form, which enters stratum 2 and is subject to stratum 2 a‰xation. 11.4
Conclusions
In summary, there are a number of arguments for separating Nuuchahnulth morphology into two strata. A variety of phonological and morphological processes provide evidence for making such a distinction in the language. On this view, phonological rules such as stress assignment and glottal stop deletion would be situated at stratum 1. Furthermore, stratum 1 would be the domain of both bound roots and su‰x combining forms. Only free roots and complete stems would
Level Ordering in Nuuchahnulth
239
be allowed to enter stratum 2. At the boundary between strata would be the phonological rules of delabialization and the rule of t´ ! . There would be two versions of the glottalization rule: one at stratum 1 a¤ecting all eligible consonants, and another at stratum 2 restricted to just the stops and a¤ricates. Taking this approach to Nuuchahnulth morphology allows us to capture the special properties of the morphology in a clear and straightforward fashion. Certain phenomena are associated with stratum 1, and others are associated with stratum 2. The modules of the lexical morphology of Nuuchahnulth can be said to operate along the following lines. (31) Stratum 1
Stratum 2
Morphology
Phonology
Bound roots Su‰x combining forms Derivational su‰xes A‰x-triggered reduplication11 Free roots and stems Distributive reduplication Plural reduplication
Stress assignment Deletion Glottalization 1 Delabialization t´ ! Glottalization 2
It should be noted here that phonological rules applying at stratum 1 do not necessarily reappear at stratum 2, apparently violating Kiparsky’s (1985) hypothesis that there is only a single set of phonological rules that may apply throughout the lexicon, possibly under distinct conditions. Nuuchahnulth would appear to constitute a counterexample to this proposal, although perhaps not the first one to be brought to light (see, e.g., Buckley 1992 on Kashaya and Inkelas and Orgun 1995 on Turkish). A number of issues remain to be addressed, including the status of clitics in the overall picture. There are, in fact, a number of clitics or clitic-like objects in Nuuchahnulth that invariably occur outside of the inflectional su‰xes in (1). In addition there are a number of postlexical phonological rules, including aspiration, diphthongization, and echo-vowel placement. These issues will eventually have to be addressed in a full treatment of Nuuchahnulth morphology. For the time being, a two-stratum model appears to be capable of dealing with the various issues encountered in Nuuchahnulth morphology and phonology. Notes I would like to thank Umberto Ansaldo, Sharon Inkelas, Stephen Matthews, Joel Nevis, Winnie Yiu, and an anonymous reviewer for comments on an earlier draft of this chapter. Data for this chapter come from a project on the Nuuchahnulth textual material gathered by Edward
240
John Stonham
Sapir and others from the early part of the twentieth century. This research was supported, in part, by a Hong Kong Universities Grants Council award (RGC No. HKU 7183/97H). The abbreviations used in table 11.1 and throughout the chapter are as follows: [l] ¼ lengthened vowel; [m] ¼ momentaneous aspect; [r] ¼ reduplication; 1s ¼ first person singular; 2s ¼ second person singular; 1p ¼ first person plural; 1p>3obj ¼ first person plural acting on a third person object; 2p ¼ second person plural; 3 ¼ third person; abs ¼ absolutive mood; being ¼ existential; caus ¼ causative; cnd ¼ conditional mood; def ¼ definite; dim ¼ diminutive; dub ¼ dubitative mood; dup ¼ reduplication; fut ¼ future; imp ¼ imperative; inal ¼ inalienable possession; inc ¼ inceptive; ind ¼ indicative; indef ¼ indefinite; infer ¼ inferential; intent ¼ intentive; inter ¼ interrogative; irr ¼ irrealis; loc ¼ locative; mom ¼ momentaneous aspect; mw ¼ meanwhile; nom ¼ nominalizer; now ¼ contemporaneous; pass ¼ passive; past ¼ past tense; pl ¼ plural; poss ¼ possessive; pres ¼ present tense; purp ¼ purposive mood; qt ¼ quotative mood; ref ¼ referential base; rel ¼ relative; rep ¼ repetitive aspect; sub ¼ subordinate mood; suf ¼ su‰x; n ¼ noun; v ¼ verb; adj ¼ adjective. k indicates the boundary between derivation (before k) and inflection (after k): 1. See Stonham (1997) for discussion of this issue. 2. Examples will be represented in a phonemicized form of the IPA. The length marks // and /‡/ represent the phonological distinction of long versus variable length vowels. Morphemic analysis that is provided in certain cases employs the free form as UR and marks the stratum boundary by means of the symbol hki. 3. I will avoid use of the terms ‘‘derivational’’ versus ‘‘inflectional’’ to distinguish the strata, as the arguments would be circular, and will use the traditional terms introduced by S&S instead. 4. Henceforth the glosses of so-called incremental su‰xes will be represented in small caps, those of derivation in lower case, and those of aspect with an initial uppercase letter. It should be noted that this is a tentative division in some ways and that there are still some a‰xes whose membership remains unclear. 5. No attempt will be made here to explain the various phonological rules that interact to produce striking di¤erences between the underlying and surface forms of words. The reader is referred to Stonham (1999b) for further details. 6. Since /h/ never occurs inside the word, it does not factor into this characterization. 7. This form undergoes the further rule of delabialization discussed further on. 8. [u] results from the coalescence of /w/ with /i/. 9. It should be noted that a rule of labialization a¤ecting velars and uvulars when preceded by a round vowel obscures the facts in certain situations. 10. The exact mechanism at work here has yet to be determined, but it seems to be related to the presence of a nasal consonant somewhere near the end of the free form. 11. See Stonham (1990, 1994) for discussion of this phenomenon.
12
Inside Access: The Prosodic Role of Internal Morphological Constituency
Patricia A. Shaw
12.1
Introduction
This chapter introduces into the theoretical literature a complex variety of stress facts ,, , , from the central Coast Salish language h nq min m (Musqueam), focusing on issues fundamental to understanding the prosody-morphology interface. The main claim of the chapter, framed within the restrictive constraints of the Parallelist Hypothesis in Optimality Theory (OT; Prince and Smolensky 1993), is that constraints on prosody may access internally embedded morphological constituency. This hypothesis is crucial to the analysis of what would otherwise be opaque stress phenomena. The argument focuses on recognition at the Output level of the left edge of the innermost morphological category, the morphological root. Among the major contributions of Kiparsky’s broad-based impact on linguistic theory are many fundamental insights into the role that internal morphology can play in the course of phonological derivation. The complexities of this interface are patently at the core of the framework of Lexical Morphology and Phonology (hereafter LMP). Kiparsky’s evolving theoretical perspectives (Kiparsky 1982a, 1982d, 1982e, 1983b, 1985) on issues such as Strict Cyclicity, Subjacency, Level Ordering, lexical versus postlexical properties, and so on, have consistently focused on identifying appropriate constraints on phonological reference to internal morphological structure. These several postulates cohered into a derivational model of the phonology-morphology interface that hypothesized that once morphological derivation had progressed from an inner morphologically defined cycle or level to a subsequent stage of morphological layering, then the internal morphological structure was no longer accessible or ‘‘visible’’ to phonological processes at the later stage of derivation (‘‘Bracket Erasure’’). Subsequent theoretical perspectives—for example, OT and Declarative Phonology (Scobbie 1991)—have challenged explicitly derivational models of phonology, necessitating a reevaluation of the empirical support for those constraints on the morphology-phonology interface that were conceived from a generative derivational e
e
e
242
Patricia A. Shaw
perspective. Within OT, the diverse roles of cyclicity in a grammar have been transfigured in various ways; see for example Kager and Zonneveld 1999 as well as Hermans and van Oostendorp 1999 for overviews. Two principal constructs are of particular relevance to the issues focused on here. One is the foundational Parallelist Hypothesis (Prince and Smolensky 1993), which holds that there is no serial derivation. Rather, parallel constraint satisfaction operates on a direct pairing of Input and Output representations. Mitigating the potential e¤ect of the Parallelist Hypothesis to e¤ectively obscure all internal structure within the earlier purview of cyclic or level-ordered access is the subtheory of constraints on Generalized Alignment (McCarthy and Prince 1993a), defined as a family of constraints that align edges of morphological categories (henceforth MCat) and/or prosodic categories (henceforth PCat). Although the Parallelist Hypothesis expressly prohibits access to any level of representation akin to an intermediate stage of a derivation, Generalized Alignment counteracts its restrictiveness, allowing more liberal access to internal morphological constituency than was provided even in LMP, which adhered to the general constraints of Strict Cyclicity, Subjacency, and Bracket Erasure. Further increasing the descriptive power of OT is Output-Output correspondence, which permits morphologically related words to influence each other’s phonology (e.g., Benua 1997). The challenge from the OT perspective is how to constrain ‘‘inside access’’ in principled ways to just those domains for which there is clear empirical justification. Kiparsky (2000c, 351–352) argues that if the Parallelist Hypothesis is interpreted in terms of ‘‘the intrinsic morphological and prosodic constituency of words and phrases, as characterized by the Stem, Word, and Postlexical levels’’ of LMP, then cyclicity e¤ects can be captured in terms of Input-Output (I-O) faithfulness, rather than requiring recourse to significantly more powerful Output-Output or Paradigm Uniformity constraints. The basic tenet of Kiparsky’s OT-based LMP model holds that ‘‘the way processes interact depends [on] their functional organization in the grammar into domains, which define a hierarchy of phonological levels’’ (Kiparsky 2000c, 356). The principal focus of the present study is the issue of what kinds of access to in,, , , ternal morphological constituency are required by the stress system of h nq min m, and what kinds of constraints can be defined to delimit such access appropriately. It is argued on the basis of this prosodic system that OT, under the Parallelist Hypothesis, requires ‘‘inside access,’’ specifically, reference to an internal domain defined by the left edge of the innermost morphological constituent of the word, the morphological root (hereafter MRoot). Contributing to a growing body of recent literature (e.g., Booij 1997c; Czaykowska-Higgins 1998; Downing 2000; Hyman and Katamba 1999; Inkelas and Zoll 2000; Kiparsky 2000c; and so on) that argues for the necessity of recognizing internal morphological constituency in the analysis of various pro,, , , sodic processes, the h nq min m case provides clear empirical evidence that the e
e
e
e
e
e
Inside Access
243
MRoot must not be rendered invisible by the presence of outer layers of prefixation. Properties of stress, vowel reduction, epenthesis, infixation, and reduplicative shape variance all converge in identifying this internal morphological edge as phonologically significant. Of further theoretical import, given the apparent crosslinguistic rarity of quantity-sensitive trochaic stress systems and various controversies about their properties (Hayes 1995; Kager 1995a, 1995b), this chapter makes a substantial contribution to the detailed documentation and analysis of a moraic trochee system. Of descriptive import is the fact that the language under discussion here is critically endangered.1 ,, The chapter is organized as follows. General phonological properties of h nq , , min m relevant to the analysis are presented in section 12.2, along with detailed ,, , , argumentation for a treatment of h nq min m schwa as featureless, weightless, and nonlexical. Section 12.3 presents an overview of the general stress patterns of the language, motivating its analysis as a moraic trochee system. In section 12.4, foot construction and foot alignment are shown to be sensitive to the internal domain defined by the left edge of the MRoot. Stress patterns of morphological word (MWord) prefixes (section 12.4.1) are contrasted with those of morphological stem (MStem) prefixes (section 12.4.2); both support stress foot alignment to the left edge of the MRoot. Section 12.5 summarizes the argumentation and conclusions. -e
e
e
e
e
e
,, , , General Phonological Properties of h nq min m e
e
e
12.2
,, , , As is characteristic of Salish languages, the consonant inventory of h nq min m2 (also referred to as Musqueam, traditionally spoken by peoples living near the mouth of the north arm of the Fraser River in British Columbia) is extensive:3 e
m , m
t t’ n , n
c , ¨’ c ´ s l l’
(cˇ)
(k)
[sˇ]
x y , y
k, w kw xw w , w
q , q w
e
t’ y
y
e
(1) p , p
qw , qw ww h
In contrast, the vowel inventory is relatively small, consisting basically of three distinct full vowels /i e a/ plus schwa [ ]. The limited occurrence of /u/ gives it marginal status as a fourth full vowel.4 Only the full vowels can carry length.5 The model in (2) represents the basic tripartite parsing of the morphological struc,, , , ture of h nq min m words into MWord, MStem, and MRoot domains:6 e
e
e
e
(2) [MWord Nonredup prefixes [MStem Redup prefixes [MRoot Lexical root . . . Analyses of Salish stress have traditionally countenanced a high degree of lexical stress specification. The present analysis argues that much of this apparent specificity
244
Patricia A. Shaw
is instead derivable from prosodic constraints interacting with morphological domain properties: only the left edges of the three domains in (2) are focused on, since stress feet are left-aligned. The question of what edge stress feet are aligned to is not transparent. The MWord domain is defined by a set of prefixes that are argued to be outside of the stress domain; all nonreduplicative prefixes belong to this outer set. The MStem domain consists of various reduplicative prefixes. Although stress may indeed surface within the MStem domain, the major claim, in section 12.4, is that it is the innermost MRoot that demarcates optimal foot alignment. Conflict with a higherranked prosodic constraint renders this alignment generalization opaque. Significantly, the salience of the internal MRoot edge is not eclipsed by prefixation within the two outer levels, contrary to what a theory assuming Bracket Erasure would predict. 12.2.1
Properties of Schwa
Any discussion of Salish phonology, including stress, requires an understanding of the behavior of schwa ([ ]). Extensive argumentation is presented in Shaw et al. ,, , , 1999 for an analysis of schwa in h nq min m as featureless, weightless, and nonlexical. These are independent hypotheses; all three hold in the epenthetic schwa behavior documented here.7 A summary of some of the principal kinds of empirical evidence supporting these hypotheses follows. Supporting the first claim, that schwa is featureless, is the fact that schwa takes on coloration from adjacent consonants ranging in realization throughout both the vertical and horizontal dimensions of the vowel space—for instance, [i] before palatovelar /x/ as in /cel x/ [clic¸] ‘hand’; [V] before tautosyllabic rounded velars /k w /, /x w / as in /x w lm x w / [x w lmVx w ] ‘person of First Nations ancestry’; [þ] before a tautosyllabic glottal stop as in /sm y y ´/ [sm y yþ´] ‘belonging to a deer’ (e.g., tracks). Further, a pervasive, low-level process of translaryngeal harmony entails full assimilation of schwa to the quality of another vowel across an intervening //, e.g. [st x w ´] @ [stx w ´] ‘children’, [t ı´] @ [ti ı´] ‘this one’. It is certainly the case that the full vowels are also phonetically variable and influenced by adjacent consonants. For example, /e/ is generally realized as [], as in [mn] ‘father’ and, often, as a high [æ:] under length, , as in /e:ny / [æ:ny ] ‘1 s. ind pro’; it raises and fronts to [e] before /y/, as in [sk w ey] ‘impossible’. To take another example, phonemic /i/ is , , characteristically lowered to a high o¤glided [e i ] by a preceding uvular, as in /qic y/ , i, [qe ciy] ‘Katzie’ (place name). However, the phonetic range of variation for full vowels is much more narrowly restricted in the vowel space than is the range for schwa. Coarticulatory e¤ects from adjacent segments add phonetic detail to, but do not override, the lexically distinctive core feature specification of each full vowel. The hypothesis that schwa has no distinctive features accounts for surrounding consonants having complete influence over its realization and for its range extending throughout the vowel space. e
e
e
e
e
e
e e
e
e
e
e
e e e
e
e
e
e
Inside Access
245
Motivating the hypothesis that schwa is weightless—that is, nonmoraic—are several aspects of its stress behavior. As will be exemplified in detail in section 12.3, ,, , , stress in h nq min m consistently di¤erentiates the full vowels /i e a (u)/ from schwa, preferentially stressing a full vowel. The hypothesis that stress is weight-sensitive, with schwa being weightless, o¤ers a principled account of these patterns. Further, under the standard moraic hypothesis that vowel length is represented by two as opposed to one mora, the complementary hypothesis that schwa is nonmoraic explains ,, , , its systematic failure to lengthen in h nq min m in contexts where other vowels (i.e., the moraic full vowels) do lengthen. The third hypothesis is that schwa is basically nonlexical. The lexical-versusepenthetic status of schwa is one of the most complex and controversial issues in Salish. The role of schwa in the present analysis is that of an unmarked default vowel fulfilling a headedness requirement for a syllabic nucleus in various contexts defined by universal markedness constraints on prosodic parsing and well-formedness. Unfortunately space limitations preclude a motivated account of constraints governing schwa epenthesis (but see Shaw 2002); consequently, Input representations will generally include schwa where it is realized in the Output and when it is not directly relevant to the specific argumentation in focus. e
e
e
e
e
e
12.2.2
ProperHeadedness
The hierarchical set of prosodic well-formedness constraints hypothesized to govern headedness (ProperHeadedness, or PropHead) within the Prosodic Hierarchy is given in (3a–c) (see Ola 1995; Ito and Mester 1992). The constraint in (3c), also known as sNuc or Nuc (Prince and Smolensky 1993), ensures that syllables have a nuclear head (3c). All potentially violable, the constraints are in an entailment relation such that the satisfaction of PropHead at each successively higher level of prosodic structure is directly dependent on the satisfaction of PropHead in one of the constituents it immediately dominates; this is expressed in (3d). Epenthesizing a schwa to satisfy sNuc violates the lower-ranked correspondence constraint DepNuc, specified in (3e). In the full range of data being considered here, all syllables satisfy Onset (3f ).8 (3) Prosodic hierarchy: ProperHeadedness a. PROPHEAD-PrWd A Prosodic Word (PrWd) is headed by a Foot. b. PROPHEAD-Ft A Foot (Ft) is headed by a Syllable. c. PROPHEAD-s A Syllable (s) is headed by a Nucleus [¼sNuc]. d. PropHead-PrWd I PropHead-Ft I PropHead-s
246
Patricia A. Shaw
Correspondence e. DEPNUC Every Nuc element in the Output has a corresponding Nuc element in the Input. Syllable structure f. ONSET Syllables must have an Onset. Given this general overview, let us proceed to the detailed investigation of the conditions governing stress. ,, , , Stress in h nq min m e
e
12.3
e
The traditional view of stress in Salish languages is of a highly lexicalized system of competing strengths between strong, weak, and variable roots and a‰xes (see especially Czaykowska-Higgins 1993 and Czaykowska-Higgins and Kinkade 1998, 15– 16). The present approach diverges significantly from these treatments, drawing instead on the interpretation in Shaw et al. 1999, where it is argued that many of the perceived lexical properties reduce to a major phonological distinction between schwa as opposed to full vowels, combined with a sensitivity of stress placement to internal morphological structure. Assuming schwa to be nonmoraic, the complexities ,, , , of h nq min m stress receive a coherent analysis as a quantity-sensitive trochaic stress system, with the left edge of the stress domain being the MRoot. The lexical residue is reduced primarily to a subset of su‰xes that are lexically endowed with certain distinctive lexical or subcategorizational properties. ,, , , The general phonological properties of word-level stress in h nq min m are outlined in section 12.3.1 and formalized in OT terms in section 12.3.2. Section 12.4 investigates the influence of morphological prefixation domains on stress behavior. e
e
e
e
e
e
12.3.1
General Properties of Stress
All lexical categories surface with stress (primary stress is marked with an acute accent over the vowel, secondary stress with a grave accent), even if monosyllabic as in (4), and regardless of whether the vowel is full—that is, [i, e, a] and marginally [u] as in (4a), or schwa as in (4b). (The s-/sˇ- prefix in these and the following data is a nominalizer prefix.)9 , (4) a. yı´ ‘big’ s-ca´tw ‘halibut’ ´cı´ws ‘tired’ ‘harbor seal’ pu´s ‘cat’ (loan) qa´ ‘water’ e´sˇx w , b. ´y ‘good’ s-m ´k w ‘ball’ m ´l’q w ‘uvula’ , , s-p ´x w ‘tripe’ q w ´s ‘enter the water’ p ´q ‘white’ e e
e e
e
e
Inside Access
247
In polysyllabic words, stress is determined with reference to the left edge and, as in Proto-Salish (Thompson 1979, 721), is sensitive to the distinction between full vowels and schwa. The data in (5a) exemplify words where the full vowel precedes a schwa: the initial full vowel receives stress. If, as in (5b), an initial schwa is followed by a full vowel in the subsequent syllable, then the full vowel attracts stress: t’a´m n pu´l’ s , h ma´ k pu´ e e
‘to walk’ ‘sasquatch’ ‘living parent’s sibling’ ‘beaver’
e e
(5) a. ı´m x , se´sq c b. sˇ-x w mnı´k w , s-q le´w
‘wall’ ‘cats’ (loan) ‘pigeon’ ‘coat’ (loan)
e e e e
However, if both vowels are schwa, as in (6), then initial stress surfaces, even though it results in stressing a schwa. This establishes stress as basically trochaic, optimally aligned with the left edge. All data thus far are MRoot-initial or prefixed only by the nominalizer s-/sˇ-. A more precise characterization of ‘‘left edge’’ is given in section 12.4. (6) a. b. c. d. e. f.
q ´l m ´ ´m x w k w ´n-n- x w , m ´n s-n ´x w ´ w ´w s
(7)
Root , nı´w , pe´t’, y xa´k w
‘eye’ ‘(to) rain’ hold-n.tr-3su ¼ ‘manage to get, hold’ ‘child’ ‘canoe’ ‘tree frog’
e e e e e e e e e e ee
,, , , Sequences of adjacent short full vowels are virtually nonexistent in h nq min m. Where such sequences would arise through morphological concatenation, one or the other full vowel reduces to schwa,10 attesting to the active role played by grammatical constraints in e¤ecting a strong-weak rhythmic pattern. For example, the reduplicated Progressive forms in (7) show that vowel identity in Base-Reduplicant correspondence is sacrificed in order to optimize a left-headed (V ) foot rather than a (V V) foot. e
e
e
e
Progressive-Root (-T.TR) , , nı´-n w- t (*nı´-niw- t) , , , , pe´-p t’, y (*pe´-pet’, y ) xa´-x k w - t (*xa´-xak w - t) e
e
e
e
e e e
a. b. c.
Gloss ‘advise/be advising someone’ ‘sew/be sewing’ ‘bathe/bathing someone’11
In accordance with the claim that schwa is nonmoraic (see section 12.2.1), the re,, , , sult of such pervasive schwa-reduction e¤ects in h nq min m is a stress system actively favoring the ‘‘uneven’’ trochee.12 Further confirmation of a basic trochaic system is seen in patterns of secondary stress on polysyllabic forms. The words in (8) illustrate that footing is iterative: syllables throughout the word are grouped into bisyllabic trochaic feet, foot parsing being e
e
e
248
Patricia A. Shaw
indicated by parentheses. The head syllable of the rightmost foot in the word receives the greatest degree of prominence—that is, primary stress. The morphological composition of each form is shown on the left; its trochaic foot parse is on the right. (8) a. /¨ [-l-]qt¼ l’¼ew n/ (¨’ `lqt )(l’e´w n) [pl]-long¼connective¼side ¼ ‘long arms’ ( `´t )(na´y n) b. / ´t n¼ay n/ eat¼margin/edge ¼ ‘eat along the way’ (WS, 290) , , c. /cew- y-amx- s/ (ce´w )(ya´mx s) help-t.tr-1sobj-3trsu ¼ ‘she/he helps me’ d. /it t- lm n/ (ı`t )(t ´lm n) sleep-desiderative ¼ ‘to want to sleep’ e
e
e
e
e
e
e e
e
e
e
e
e
e e
e
e
e e
e
e e
Because the forms in (8) are exhaustively footed into bisyllabic trochees, they are not revealing with respect to edge-alignment constraints. That foot alignment is optimally to the left edge is established by two di¤erent types of data. First, trisyllabic forms with exclusively schwa nuclei are footed at the left edge, as shown by the forms in (9), all of which are MRoot-initial: * (n ´x w y t) e
e
( ´n x w ) y t (9) a. / n x w -y t/ stop-refl ¼ ‘stop oneself ’ , , (s ´wq ) n p b. /s wq¼ n p/ seek¼land ¼ ‘look for land’ , , (q ´m ) n ´p c. /q m n¼ ´p/ maple¼tree ¼ ‘maple’ (´ ´q l) l x w d. /´ q l-n- x w / Root-n.tr-3agr ¼ ‘know’
e
e
e e
e
, *s w (q ´n p)
e
, *q (m ´n ´p)
e e
e e
e
e
*´ (q ´ll x w ) e e
e
e
e
e
e
e e
e e
e e
e
e e
e
e
e e
e e
In the (unasterisked) foot parsings in (9), feet are aligned to the left edge, with stray unparsed syllables at the right margin. The parsings marked with an asterisk, with right-edge alignment, manifestly stress the wrong syllable. A second argument for left-edge alignment comes from the distribution of degenerate feet: monosyllabic feet occur only at the right edge. While, as seen in (9) and (10a), a rightmost odd-parity syllable headed by a schwa nucleus is not parsed into a degenerate foot, a final odd-parity syllable headed by a full vowel is footed, as illustrated in (10b). i. /tey¼ w ´/ (te´y ) w ´ race¼canoe/vessel ¼ ,‘racing canoe’ , w (k w e´c ) t s ii. /k ec- t- s/ see-t.tr-3trsu ¼ ‘she/he saw him/her’ , , iii. /k w in¼ q n/ (k w ´ı.n ) q n how.many¼containers ¼ ‘how many containers?’ e
e
e e
e
e
e e
e e
e e
(10) a.
Inside Access
249
iv. /pax l q w / (pa´.x ) l q w ‘yellow cedar’ (WS, 203) , , (ı`.t )(-te´wtx w ) b. i. /it t¼ewtx w / sleep¼building ¼ ‘hotel’ , , ii. /k w in¼ winx w / (k w `.n )(wı´:nx w ) how.many¼years ¼ ‘how many years’ , , , , (q `.q )(ya´s) iii. /q q¼ yas/ be.taut¼circular ¼ ‘barrel, washtub’ , , iv. /´qec s¼mat/ (´qe`.c s)(ma´t) five¼bundles/kinds ‘five bundles of a kind’ e e
ee
e
e
e e
e
e e
e
e
e
e
Left-edge foot alignment correctly accounts for the right-edge asymmetry in the distribution of degenerate feet. Although left-edge alignment is optimal, the sensitivity of stress to the schwa– versus–full vowel distinction militates against strict left-edge alignment when an initial schwa syllable is followed by one headed by a full vowel. (11a) illustrates a bisyllabic form, trisyllabic forms are given in (11b), and (11c) illustrates quadrisyllabic forms. The quadrisyllabic forms in particular strongly support the analysis of stress developed thus far. The first syllable, headed by schwa, is unfooted; the trochaic foot includes the second and third syllables, leaving the final syllable, headed by schwa, unfooted as well, as documented in (9). , , h (ma´) (¼ (5b)) (11) a. /h ma/ ‘pigeon’ b. i. /x w nit m/ x w (nı´.t m) ‘white person’ c (l’e´.q ´) ii. /c l’eq- ´/ fossilized.root-past ¼ ‘yesterday’ iii. /q w¼()ic n/ q w (ı´.c n) warm¼back ¼ ‘Cowichan’ ,, ,, iv. /ck w x¼el’ / ck w (xe´.l’ ) twenty¼people ¼ ‘twenty people’ n (wa´q w .c s) t n c. i. /n w¼aq w ¼c s-t n/ enter¼head¼hand-instr ¼ ‘thimble’ ii. /y sel’ ¼ q n/ y (se´.l’ ) q n two¼containers ¼ ‘two containers’ e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e e e
e
The proposed analysis models a fundamental asymmetry in the behavior of full vowels as opposed to schwa. The footing of full vowels is exhaustive;13 syllables headed by a full vowel are systematically parsed as a trochaic foot head, and a final full-vowel syllable at the right edge is parsed as a degenerate (monosyllabic) foot (10b). In contrast, the footing of schwas is much more restrictive. Aside from
250
Patricia A. Shaw
monosyllabic words compelled by culminativity (see section 12.3.2.3), a foot may be headed by schwa, but only under the following two conditions: (i) the foot is bisyllabic, and (ii) the syllable in weak metrical position is not headed by a full vowel. As a consequence, parsing of syllables headed by schwa to the foot level is not necessarily exhaustive. As seen most dramatically in (11c), more than one single syllable headed by schwa can remain unfooted in a word. However, even in a word all of whose syllables are schwa vowels, some degree of footing is obligatory; in all forms seen thus far it is the case that no two adjacent syllables may remain unparsed.14 12.3.2
An Optimality-Theoretic Interpretation
The distinction between full vowels and schwa, clearly central to an account of the ,, , , h nq min m stress system, is modeled in (12) as a moraic distinction. Each full vowel has underlying moraic structure, while schwa is entirely devoid of both moraic and featural specification. Completely unmarked, epenthetic schwa realizes a prosodically required syllable Nucleus position, abbreviated as Nuc, thus functioning as the default segment. e
e
e
(12) Nucleus
a. Full vowel Nuc
b. Schwa Nuc
c. Schwa reduction Nuc ¼ m
Moraic weight
m
Root node
([f ])
([f ])
Features
Although weight (moraicity) and syllable headedness are optimally correlated, as reflected by universal constraints such as Wgt-to-Str and PeakProminence (Prince and Smolensky 1993), the hypothesis here (following Shaw 1992, 1993, and so on; Jiang-King 1996; Ola 1995; Blake 2001; Howe 2000) is that they have independent prosodic status. Giving Nuc formal status independent of moraic weight predicts the well-known fact that weight may be nonnuclear (e.g., where a coda consonant contributes to a heavy syllable), as well as the complementary fact that a Nuc consisting of schwa may be weightless (Anderson 1982; Kager 1990; Kinkade 1998; among others). If schwa is represented, as in (12), as simply a nuclear node, its phonetically variable realization is predictable through feature spreading from its paradigmatic and syntagmatic context. As diagrammed in (12c), the familiar process of ‘‘schwa reduction,’’ whereby a ‘‘full’’ vowel in weak metrical position reduces to a light default vowel, is formalized as deletion of all content below the Nuc, consistent with the basic hypothesis that schwa is nonmoraic and lacks distinctive featural (represented as [f ]) content. Vowel deletion, which subsumes schwa deletion, entails deletion from above the Nuc node, including all content dependent therefrom.
Inside Access
251
The claim that schwa is nonmoraic plays a significant role in accounting for the stress behavior of schwa in contrast to full vowels: its weightlessness makes it less optimal as a prosodic Foot head under constraints such as Wgt-to-Str (see (13c)) that value a correlation between weight and prominence. Given these structural assumptions regarding the appropriate representation of full vowels as opposed to schwa, the general stress patterns of data sets (9) through (11) derive straightforwardly from the constraints in (13). The basic parsing into trochaic feet results from the high ranking of (13a), FootForm ¼ Trochee (FtForm). Although the data thus far have shown only that feet are built from the left edge, section 12.4 will show that it is in fact the left edge of the MRoot, rather than the MStem or the MWord, to which foot alignment is sensitive. Align-(Foot, L, MRoot, L) (Al-L(Ft, MRt)) constrains the left (L) edge of a foot (Ft) to correspond with the left edge of a lexical MRoot (MRt) (13b). The preferential stressing of a full vowel over schwa derives from the Weight-to-Stress (Wgt-to-Str) constraint in (13c), assuming the proposal in (12) that a full vowel is moraic (i.e., has weight), whereas epenthetic schwa is not. Wgtto-Str is ranked above Al-L(Ft, MRt). Hence, if the initial syllable is headed by schwa, footing does not start until the second syllable.15 However, as was seen in (6) and (9), if the second syllable does not provide a more optimal head, Al-L(Ft, MRt) is not compromised. 12.3.2.1
A Moraic Trochee System
(13) a. FOOTFORM¼Trochee (FTFORM) Feet are left-headed: ( s s) or ( s). b. ALIGN(FOOT, L, MROOT, L) (AL-L(FT, MRT)) Align the left edge of a foot with the left edge of the MRoot. c. WEIGHT-TO-STRESS (WGT-TO-STR) If a syllable nucleus has weight, then it is prominent within a foot. d. Ranking FtForm, Wgt-to-Str Al-L(Ft, MRt) The interaction of these constraints to e¤ect the appropriate parsings of the data considered thus far is illustrated in the tableaux in (14),16 in which a left bracket demarcates the left edge of the MRoot. The crucial ranking identified in (13d) of FootForm and Wgt-to-Str above Al-L(Ft, MRt) is established by the evaluation of competing candidates in the tableau for [x w nı´t m] ‘white person’. Consider, first, candidate (14ai). The foot head is the full vowel [i], which satisfies Wgt-to-Str. Moreover, the left edge of the foot (x w nı´) aligns directly with the left edge of the morphological root, thereby satisfying Al-L(Ft, Rt). The problem with candidate (14ai) is that the foot is a right-headed iamb, not a left-headed trochee. The fact that candidate (14aiii), not (14ai), is the correct output requires FtForm to be ranked above Al-L(Ft, MRt). Candidate (14aii) conforms to FtForm and Al-L(Ft, e
e
e
252
Patricia A. Shaw
MRt), but at the expense of fatally violating Wgt-to-Str, thus motivating the crucial ranking of Wgt-to-Str above Al-L(Ft, MRt). The second tableau, in (14b), for [ ´n x w y t] ‘stop oneself ’, shows that unless the second syllable has weight—that is, a full vowel—then Al-L(Ft, MRt) will not be compromised. e
e e
(14) a. x w nı´t m ‘white person’ (¼ (11bi)) e
[x w nit m
FtForm
i. (x w nı´) t m
*!
e
e
e
e
e
ii. (x w ´ni) t m
Wgt-to-Str
Al-L(Ft, MRt)
*!
e
e
+ iii. x w (nı´t m)
*
e
e
b. ´n x w y t ‘stop oneself ’ (cf. (9)) e
e e
[ n x w -y t e
e e
i. ( n ´x w ) y t e
Wgt-to-Str
Al-L(Ft, MRt)
*!
y t) e
e e
e
ii. (n
´x w
FtForm
*!
e
+ iii. ( ´n x w ) y t e
e e
The Role of Binarity The generally alternating pattern of stress in su‰ciently long words (8) is governed by two constraints. Maximal parsing of syllables into feet is e¤ected by Ft-Bin-s, which requires feet to be disyllabic (15a), and by Parse-s-2, which is violated by any sequence of two unfooted syllables (15b). Degree of stress, in words with more than one foot, is handled by the constraint Rightmost (15c), which captures the generalization that the head syllable of the rightmost foot in the word gets primary stress, while the heads of all other feet surface with secondary stress. (15d) shows crucial rankings among these constraints as well as constraints in (3) and (13). 12.3.2.2
(15) a. FT-BIN-s Feet are binary at the syllabic level. b. PARSE-s-2 Adjacent syllables cannot both remain unparsed.17 c. RIGHTMOST The rightmost foot in a word has greatest prominence. d. Ranking Wgt-to-Str, Parse-s-2 Al-L(Ft, MRt) FtBin-s PropHead (3) FtBin-s, DepNuc (3) Given the criterial role of Wgt-to-Str in optimizing prominence on full vowels, it is important to consider how it is that words with no full vowel get footed at all. This
Inside Access
253
is shown in the tableau in (16), which shows how Parse-s-2 and Al-L(Ft, MRt) can function crucially in such cases. (16) ´n x w y t ‘stop oneself ’ (cf. (9)) e
e e
[ n x w -y t e
Parse-s-2
e e
a. n x w y t
*!
b. ( ´) n x w y t
*!
Al-L(Ft, MRt)
*
e
e e
e
e
e
y t)
*!
d. ( `n x w )(y ´t)
*!*
e
e
e
c. (n
´x w
Ft-Bin-s
*
e
e e
+ e. ( ´n x w ) y t e
e e
f. ( ´n x w y t)
*!
e
e e
Candidates (16a) and (16b) both fatally violate Parse-s-2: this constraint allows a single syllable to remain stray, but requires two adjacent syllables to be parsed into a foot. (16c) and (16e) have only one stray syllable each, satisfying Parse-s-2; (16c) is less optimal because its foot does not align with the left edge of the MRoot. Interestingly, the fully footed candidate in (16d) is even worse than (16c) because its final (degenerate) foot incurs two alignment violations, since it is two syllables away from the Root edge. The left-aligned exhaustive parse in (16f ) violates Ft-Bin-s. It is thus candidate (16e), with a final stray syllable, that ends up being optimal since it satisfies both Parse-s-2 and Al-L(Ft, MRt). The tableaux in (17) illustrate cases where Ft-Bin-s makes it optimal to leave an initial schwa syllable stray.18 (17) a. x w nı´t m ‘white person’ (cf. (11bii)) e
e
[x w nit m e
e
Al-L(Ft, MRt)
i. x w (nı´t m)
*
ii. (x w `)(nı´t m)
*
e
e
e
+
Wgt-to-Str
e
, b. h ma´ ‘pigeon’ (¼ (5b), (11a)) , Wgt-to-Str [h ma , i. (h ´ma) *! , ii. (h `)(ma´) , + iii. h (ma´)
Ft-Bin-s
*!
e
e
Al-L(Ft, MRt)
Ft-Bin-s
e
*
**!
e
*
*
e
254
Patricia A. Shaw
With respect to the role of Ft-Bin-s, recall the distinction in the data of (10) where a single final syllable headed by schwa is not footed, whereas a single final syllable headed by a full vowel is footed. The tableau in (18a) shows the e¤ect of the constraint ranking in (15e) on a stress-initial odd-parity word where the final syllable is headed by schwa. Under the postulated constraint rankings, there is no optimal way to parse the final syllable into a foot, which is exactly the correct result. In contrast, however, a final odd-parity syllable headed by a full vowel is parsed into a degenerate foot. As shown in the tableau in (18b), the very same ranking accounts for cases like this. In successful candidate (18bi), the fact that the full vowel in the final syllable is footed (te´wtx w ), thus bearing stress, satisfies the highly ranked Wgt-to-Str constraint. The degenerate foot does violate lower-ranked Ft-Bin-s and Al-L(Ft, MRt), since the final degenerate foot is misaligned from the L-edge by two syllables, but it still fares better than candidate (18bi), which violates Wgt-to-Str. , (18) a. k w e´c t s ‘she saw him’ (¼ (10aii)) , Wgt-to-Str Al-L(Ft, MRt) Ft-Bin-s [k w ec- t- s , *! * i. (k w e`)(c ´t s) , w` *!* * ii. (k ec )(t ´s) , + iii. (k w e´c )t s ee
e e
ee
e e e e
b. ı`t te´wtx w ‘hotel’ (¼ (10bi)) e
[it t¼ewtx w e
i. (ı´t ) tewtx w e
+ ii. (ı`t )(te´wtx w )
Wgt-to-Str
Al-L(Ft, MRt)
Ft-Bin-s
**
*
*!
e
A direct comparison of these two tableaux shows that the right-edge asymmetry in the distribution of degenerate feet that was observed in the data of (10) falls out straightforwardly from a single ranking of the constraints Wgt-to-Str Al-L, FtBin-s in the proposed analysis. The other constraint relevant to whether stray syllables surface is Parse-s-2. Evidence that it too is crucially ranked above Align-L(Ft, MRt) is established through consideration of even-parity words like (19a). Since schwa is weightless, the nonparsing of a syllable headed by schwa does not violate Wgt-to-Str. One unparsed schwa syllable is tolerated, as in all the examples in (11) and (17), but two adjacent unparsed schwa syllables are not: Parse-s-2 ensures that they will be parsed into a bisyllabic foot. Rightmost (15c) (not shown in the tableaux here due to space) accords it primary stress.
Inside Access
255
The tableau in (19b) illustrates the prosodic parsing of a word with an initial sequence of schwa-headed syllables. In each candidate form, the second foot consistently violates Al-L(Ft, MRt) by two syllables. Candidate (i) shows that Ft-Bin-s will rule out any stray schwa syllables if they can be parsed into a binary foot. Candidate (ii), with one foot and two adjacent stray syllables, fatally violates Parse-s-2. The successful candidate (iii) satisfies both Parse-s-2 and Ft-Bin-s. (19) a. ı`t t ´lm n ‘to want to sleep’ (¼ (8d)) e
ee
[it t¼ lm n e
Wgt-to-Str
Parse-s-2
e
e
i. (ı´t )t l.m n
Al-L(Ft, MRt)
*!
e
e e
+ ii. (ı`t )(t ´lm n)
**
e
e e
b. `´t na´y n ‘eat along the way’ (¼ (8b)) e
e e
[ ´t n¼ay n e
Parse-s-2
e e
i. ( `´)t (na´)y n e
e
e
ii. ´.t (na´y n)
*!
e
e
e
+ iii. ( `´t )(na´y n)
Al-L(Ft, MRt)
Ft-Bin-s
**
*!*
** **
e
e e
As seen in section 12.4, prefixed forms may violate Parse-s-2; however, all the forms evaluated thus far are MRoot-initial, and obey Parse-s-2. Tableau (20) illustrates an odd-parity five-syllable word beginning with two schwa syllables and ending with a schwa syllable. In accordance with what has been seen so far, schwa cannot head a degenerate (nonbinary) foot, and thus the final syllable remains unfooted. The established ranking of Parse-s-2 Al-L(Ft, MRt) predicts this outcome. ,, , , ,, , , (20) h `nq mı´n mq n ‘the h nq min m way of speaking’ (cf. q n ‘throat’) ,, , , h nq min m¼q n Parse-s-2 Al-L(Ft, MRt) Ft-Bin-s ,, , , **,*!*** * a. (h `nq )(mı`n m)(q ´n) , , ,, **,*!** * b. (h `nq )(mı`)(n ´mq n) , , , , *! ** c. h n.q (mı´.n m)q n ,, , , + d. (h nq )(mı´n m)q n ** e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e e
e
e
e
e
e
e
e e e e
Candidates (20a) and (20b) represent the two most plausible exhaustive parsings of the five-syllable string; both satisfy high-ranked Wgt-to-Str (not shown in (20)). Comparing these with the successful, nonexhaustively parsed candidate in (20d), we
256
Patricia A. Shaw
see that Al-L(Ft, MRt) prevents the final schwa syllable from being footed; the respective accumulation of additional misalignment violations renders (20a) and (20b) much less optimal than candidate (20d). In contrast, candidate (20c), the most plausible nonexhaustive parsing that respects both Wgt-to-Str and Ft-Bin-s, ties with (20d) in number of alignment violations but fatally violates Parse-s-2; treatment of the initial sequence of two schwa syllables is the critical factor favoring (20d) over (20c). 12.3.2.3 Culminativity Given the argumentation above that a single schwa syllable can remain unparsed, as in (20), (17a,b), and (16), it is important to consider what ,, , , accounts for the stressing of such syllables in monosyllabic words such as /pq/ [p ´q] ‘white’ (4b). First, note that in monosyllabic words with a full vowel, like [e´sˇx w ] ‘harbor seal’ (4a), the high-ranking (in fact inviolable) constraint Wgt-to-Str forces violation of Ft-Bin-s.
e
(21) e´sˇx w ‘harbor seal’ (¼ (4a)) [esx w a. esˇx w + b. (e´sˇx w )
Wgt-to-Str
Ft-Bin-s
*! *
A deeper generalization, however, is that all lexical forms in the language require stress on at least one syllable. This is captured by the hypothesis that the constraints on proper headedness (PropHead, in (3)) are consistently satisfied. The ranking in (22) of PropHead constraints above DepNuc (3) ensures that words without a lexical full vowel will surface with schwa as a default syllable head. Similarly, the systematic stressing (footing) of monosyllabic words, regardless of nuclear weight, is captured by ranking PropHead constraints above Ft-Bin-s, thus ensuring that each prosodic word has at least one foot, even if foot binarity cannot be satisfied. , , (22) p ´q ‘white’ (¼ (4b)) ,, [pq PropHead-PrWd, -Ft PropHead-s Ft-Bin-s DepNuc ,, a. pq *! * , , *! * b. p q , , * * + c. (p ´q) e
e
e
To review the basic claims of the stress analysis developed thus far, the reader is referred to the constraints and rankings summarized in (13) and (15). Based on these
Inside Access
257
generalizations, we turn to an investigation of the interplay of morphological domains with left-edge stress alignment. 12.4
Morphological Domains of Stress
In the following subsections, various sets of apparent exceptions to the basic generalizations governing stress are explored, and it is argued that each set receives a principled explanation under the basic hypothesis that it is the left edge of the innermost morphological domain, the MRoot domain, which defines optimal stress foot alignment. The analysis is structured with reference to the MWord, MStem, and MRoot domains, presented in (2). First, in section 12.4.1, the outer layer of MWord-level prefixes is examined, and shown to be consistently outside of the stress domain. The complex interaction of stress with MStem-level progressive and plural reduplicative prefixes, as well as a plural infix, is analyzed in section 12.4.2, with the several independent patterns each converging to the same conclusion: constraints governing stress make crucial reference to the internal MRoot-domain left edge. 12.4.1
MWord-Domain Prefixes
No MWord-domain prefix19 has a lexical full vowel: only schwa surfaces in this domain, and all such schwas are epenthetic, thus permitting syllabification of an adjacent resonant. The data in (23) illustrate the consistent lack of prosodic prominence on these prefixes. The stress pattern in the data of (23a) is expected from the analysis given in section 12.3.2. However, the stress patterns in (23b–d) are not. (23) a. t m-lı´l’ season-salmonberry ¼ ‘salmonberry time’ , x w n-a´m t early-be.at.home ¼ ‘get home early, soon’ , b. t m-w ´y¨’ season-cold ¼ ‘winter’ , w -p ´k w suddenly-billow.forth ¼ ‘suddenly billow forth’ (WS, 283) c. x w n-h ´ye20 early-leave ¼ ‘to leave early, soon’ t w-t ´k y somewhat.like-turkey (loan) ¼ ‘like a turkey’ (WS, 255) w -s-k w ´yw-y t established-s-move-refl ¼ ‘he moved’ , d. w -m ´k w - l p established-all-2p ¼ ‘you all’ (WS, 254) e
e
e
e
e
e
e e
e
e
e e
e
e
ee
e
e
e
e
258
Patricia A. Shaw
t l- ´n c 21 from-where ¼ ‘from where?’ e e e
e
In (23a) it is to be expected that Wgt-to-Str would initiate trochaic foot construction with the second-syllable full vowel, leaving the initial schwa syllable as stray. But, in (23b–d) where there is a sequence of syllables headed by schwa, stress would be expected on the basis of phonological criteria alone to fall on the leftmost schwa. Specifically, given initial stress , on words like [w ´w s] ‘tree frog’ , and similar data of (6), one would expect *[w ´p k w ] rather than the attested [w p ´k w ] in (23b). Similarly, in the trisyllabic forms in (23c), one would expect initial stress, as previously attested by the data of (9) in words like [ ´n x w y t] ‘stop oneself ’. The morphological parsings of these contrastive data sets are compared in (24), where the left MWord- and MRoot-domain edges are identified: e e
e e
e e
MRoot [w ´w , s [p ´k w [ ´n x w y t [k w ´yw-y t e e
MWord [ [w [ [w -s-
w ´w s, w p ´k w ´n x w y t w sk w ´ywy t
e e e e e e
e e
e e e e e e e e e e
e e
a. b. c. d.
e
(24)
‘tree frog’ (16) ‘suddenly billow forth’ (23b) ‘stop oneself ’ (9) ‘he moved (himself )’ (23c)
The morphological composition of these words clearly provides a systematic explanation for the observed di¤erence in stress patterns. The unifying generalization for the data represented by (23) and (24) is that stress falls on the initial syllable of the MRoot. Data such as these provide strong empirical evidence for Al-L(Ft, MRt) (13b). Stress feet are aligned with the left edge of the internal morphological MRoot, rather than with the MWord edge. Further corroboration of MRoot alignment is provided by data like (23d), which have a sequence of four schwas. If footing originated at the leftmost MWord margin, one would expect exhaustive parsing into two feet as in (25a). The actual parse, in (25b), best satisfies the three previously motivated constraints—Parse-s-2 AlL(Ft, MRt) Ft-Bin-s—as shown in tableau (25). MRoot [
n c
Parse-s-2
[(t `l [ ) (n ´c ) e
[t l [( ´n ) c
c.
[t l [( ´) n c
Ft-Bin-s
e
e
+ b.
Al-L(Ft, MRt) *!
e e
a.
l-
e
MWord [t
e e e
(25)
*!
*
e e
e e
e
e e
The fact that it is optimal in (25) to leave both the initial and the final syllables unparsed follows directly from the analysis and constraint ranking developed in sec-
Inside Access
259
tion 12.3.2, most specifically from the claim in (13b) that feet are aligned with respect to the internal MRoot domain. The core of the above argument is based on the di¤erential behavior of stress on words with no full vowels, only schwas. In some data—for example, (6) and (9)— stress is initial; in other data (23b,c,d), stress is pen-initial. In the absence of any full vowel to bring Wgt-to-Str into play, the criterial role of the MRoot edge emerges as the crucial factor behind the di¤erence between the two data sets. The next argument is based on data in which stress falls on the third syllable from the beginning of the word. Representative examples are given in (26a,b). In each case word-level prefixes attach to an MRoot with a schwa in the initial syllable. Note that (26c) ¼ (10biii), (26d) ¼ (8b). , , a. t ns y, e´m , b. y x w k w ı´q n
MWord MRoot , , [t n[s, ye´m , [y -x w - [k w ¼ı´q n
, , c. q `q ya´s
[
, , [q `q¼ ya´s
d. `´t na´y n
[
[ `´t n¼a´y n
e
e
e
e
e e
e
e
e e
e
e
(26)
‘from the upper class’ (WS, 279) along-loc-[ascend¼belly ¼ ‘belly up’ (WS, 536) taut¼circular ¼ ‘barrel’ (WS, 313)22 ‘to eat along the way’
e e e e e
e
e e
What is initially striking here is the contrast between the data in (26a,b) where there is an initial sequence of two unparsed schwa syllables, and previously analyzed data, repeated as (26c,d), which motivate a high-ranking role for Parse-s-2 (15b). This contrast reduces directly to the principal claim of this section: MWord-domain prefixes are outside of the purview of stress constraints. Not only are word-level prefixes excluded from the domain of foot alignment, as was seen in (23–25), but they are also excluded from the domain of Parse-s-2. , , Consider the example in (26a), [t ns ye´m], where the only stress is on the third syllable. If foot alignment could not access the left edge of the internal MRoot, we , , would expect secondary stress on the initial syllable, *[t `ns ye´m], just as in examples , , like [q `q ya´s] (26c). However, as shown in the tableau in (27), the interplay of foot alignment to the MRoot edge with constraints on weight and binarity, correctly accounts for the stress pattern in words of this type. e e
e e
e e
e
e
e e e
+ d.
e
c.
, , yem
, , [(t `n [s )(ye´m) , , [t n [(s ´ yem) , , [t n [(s `)(ye´m) , , [t n [s (ye´m) e
b.
MRoot [s
Wgt-to-Str
e
a.
n-
e
MWord [t
e
(27)
Al-L(Ft, MRt)
Ft-Bin-s
*!*
*
*
**!
*
*
*!
260
Patricia A. Shaw
Word-initial stress, as in candidate (27a), is nonoptimal. The only constraint that outranks Al-L(Ft, MRt) here and could therefore compel foot alignment to extend leftward from the MRoot edge would be Wgt-to-Str, but it is not applicable to the MWord-domain prefixes because it lacks a full vowel. Formally, (27a) is disqualified by the double Al violation: the left edge of the first foot (t `n.s ) is misaligned from , , the MRoot by one syllable on the left, and the second foot (ye´m) is misaligned by one syllable on the right. Though the candidate in (27b) does achieve perfect foot alignment with the left edge of the MRoot, it is not optimal overall, because its failure to assign prominence to the final moraic vowel [e] violates the higher-ranked Wgt-to-Str. Parsing both syllables of the root into independent monosyllabic feet, as in (27c), avoids violating Wgt-to-Str, but incurs an extra violation of Ft-Bin-s, fatal in comparison to (27d). Note that even the optimal candidate (27d) has one infraction of Al-L(Ft, MRt). The interesting consequence of aligning feet to the MRoot is that the optimal output (27d) sustains a sequence of two syllables with unstressed schwas. This raises an important question regarding the correlation of well-formedness constraints on prosodic parsing, such as Parse-s-2, with domains defined by alignment constraints. The claim advanced here is that the domains of Parse-s-2 and Al-L(Ft, MRt) directly coincide. Consider first ranking relations, under the strong hypothesis that ranking relations in a particular OT grammar are fixed. It was demonstrated, on the basis of the tableau in (19a), that Parse-s-2 is crucially ranked above Al-L(Ft, MRt). The issue that the tableau in (28), based on (27), foregrounds is this: treating (28b) ¼ (27d) as a violation of Parse-s-2 would result in a ranking paradox, since it would result in the correct output, (28b), not being evaluated as optimal. e
e
(28) Wgt-to -Str, Parse-s-2 Al-L(Ft, MRt) FtBin-s (cf. (15d)) , , Parse-s-2 Al-L(Ft, MRt) MWord [t n- MRoot [s yem , , ** L a. [(t `n [s )(ye´m) , , *! * b. [t n [s (ye´m) e
e
e
e
e
e
If, however, the domain of Parse-s-2 is formally recognized as coextensive with the domain of stress foot alignment, then MWord-domain prefixes like t n- simply would not count in the evaluation of violations to Parse-s-2, in the same way that they would not contribute to an Al-L(Ft, MRt) violation. This is shown in (29). e
Inside Access
, , t ns ye´m
MWord [t n" N/A N/A e
e e
Parse-s-2/MRt Al-L(Ft, MRt)
MRoot , , [s (ye´m) ‘from the upper class’ " nonviolation (because only one s) violation (by one s) e
(29)
261
A second prediction of establishing the MRoot as the domain of Al-L(Ft, MRt) and Parse-s-2 is that foot construction and stress will be insensitive to the number of of syllables preceding the MRoot domain. Consider, for example, a sequence of two schwa-headed syllables preceding the MRoot in the MWord domain: both will remain unparsed. As a consequence, there should be no stress on either syllable. The form and tableau in (30) confirm this prediction: e
e
, [w -y - [ey e e
, a. [(w `y ) [(e´y) , + b. [ w y [(e´y)
, ¼ [w y e´y] ‘. . . it is (you who will) continue along’ (WS, 520)
Parse-s-2
e e
(30) MWord MRoot , [w - y [e´y [established-along-[continue
Al-L(Ft, MRt) *!*
e e
e e
The candidate in (30a) exhaustively parses syllables into feet. The undesirable consequence, however, is that the leftmost foot incurs two Al-L violations, given its twosyllable distance from the MRoot domain edge. In contrast, the candidate in (30b) leaves both the MWord prefixes unfooted. Under the hypothesis that the highly ranked Parse-s-2 constraint has jurisdiction only in the MRoot domain, the two MWord-domain prefixes may remain unparsed without violating Parse-s-2. To summarize, then, the morphological structure outlined in (2) correlates with di¤erent prosodic domains. Prefixes in the MWord domain, even multiple prefixes, consistently do not get stressed. The formal interpretation of these facts is that prosodic constraints on footing (Parse-s-2 and Al-L(Ft, MRt)) apply with explicit reference to the MRoot domain. Recall that the initial observation in this section was that no MWord-domain prefix has a lexical full vowel: only epenthetic schwa surfaces in this domain.23 Given the argumentation throughout sections 12.3 and 12.4.1 that Wgt-to-Str is highly ranked (as in (13) and (15)), it is clear that the lexical absence of full vowels in any of the prefixes within the MWord domain is not accidental. Their systematic absence preserves the significant generalization that foot alignment functions optimally to signal the word-internal MRoot edge.
262
Patricia A. Shaw
12.4.2
MStem-Domain Prefixes
There are several di¤erent reduplicative morphemes, distinct in form and function, which may occur as a prefix to a root. These reduplicative prefixes essentially define the morphological content of the MStem domain. The initial argumentation in this section draws on the interaction of stress with examples from each of two principal reduplicative canons on verb roots: an open-syllable pattern, marking Progressive aspect, and a closed-syllable pattern, marking Plural.24 Infixation of an -l- Plural allomorph provides further support for the prosodic relevance of the MStem domain. The first pattern to be considered, building on Shaw et al. 1999, involves a reduplicative CV-prefix that, in conjunction with glottalization of subsequent resonant(s), marks the Progressive aspect of certain classes of verb stems.25 As seen in (31), if the MRoot has a lexically full vowel, the Progressive reduplicative prefix surfaces with a copy of the vowel and concomitant reduction of the MRoot vowel to schwa: Progressive , , pe´-p t’, y , ya´-y k w - t , , q w´ı-q w m , nı´-n w, t xa´-x k w - t e e e e e
Root (-TR) , pe´t’, y ya´k w - t , q w ´ım , nı´w, - t xa´k w - t
e e
e
e
e
a. b. c. d. e.
e
(31)
Gloss ‘sew’ ‘smash it, break it up’ ‘disembark, get out of ’ ‘advise him’ ‘bathe him’
Two properties of these forms need to be accounted for. First, the lexically distinctive vowel quality of the MRoot surfaces only in the reduplicant. Second, the Progressive reduplicative prefix is systematically stressed in such forms, clearly establishing that stress can extend leftward beyond the MRoot-domain edge. In the analysis proposed below, these properties are integrally related. Consider first the fact that in (31) the lexically distinctive quality of the MRoot vowel surfaces only in the reduplicant, not in the base; thus neither Base-Reduplicant (BR) faithfulness nor Input-Output (IO) faithfulness in vowel identity is satisfied. The relevant faithfulness constraint is MaxNuc, which subsumes faithfulness in both moraic and featural content. Assuming the Full Model of reduplication developed in McCarthy and Prince 1995, prefixal vowel quality in Progressive reduplication can be accounted for with a highly ranked Input-Reduplicant (IR) faithfulness constraint, MaxNucIR (32a).26 Setting aside for the moment the question of domain alignment, the tableau in (32b) shows how the previously motivated constraints Wgt-to-Str (13) and Ft-Bin-s (15) interact with MaxNucIR . The other FaithNuc constraints, ranked below Ft-Bin-s, are not shown here. (32) a. MaxNucir A nucleus in the Input has a correspondent in the Reduplicant.
Inside Access
263
vi. vii. + viii.
Ft-Bin-s *
*! *!* *!
*
*!
*
*! *!
e
v.
*!
e e
iv.
MaxNucir
e
iii.
Wgt-to-Str
e
ii.
‘be sewing’ (¼ (31a))
e
i.
MRoot , [p t’ y , [Prog-[pet’ y , , pe-(pe´t’ y ) , , (pe´-pet’ y ) , , (pe`-)(pe´t’ y ) , , p -(pe´t’ y ) , , p -(p ´t’ y ) , , (p ´-p t’ y ) , , (pe´-pt’ y ) , , (pe´-p t’ y ) e
b. MStem , [pe´-
Pairwise comparisons of candidates in this tableau establish that identity of lexically distinctive vowel quality between the Base and the Reduplicant is sacrificed in favor of an optimal Heavy-Light moraic trochee, in conformity with high-ranking prosodic constraints on Wgt-to-Str and Ft-Bin-s. Now let us turn to an account of the stress pattern exhibited by Progressive reduplication. Considered in isolation, the data in (31) might lead one to think that foot alignment extends to the left edge of the MStem domain, thereby including the reduplicative prefixes in its purview. However, two other sets of data show that this is not the case, and that stress aligns with the MRoot edge. The first set of data involves underlyingly vowelless roots (see (22)); in the Progressive stems of such roots, the reduplicative prefix takes the form C -. As expected, where syllabification requires a nuclear head, a default epenthetic schwa appears. Representative data are given in (33). In these data, schwa is not included in underlying representation (see section 12.2.1). e
Reduplicated Progressive stem t -t ´s , , , c -c ´q w , s, -s ´q, k w -k w ´´ , x w -x w ´y ´ -´ ´t’- m t’ -t’ ´p l’ e e e e
Unreduplicated MRoot t ´s , , c ´q w , s, ´q k w ´´ x w ´y ´ ´t’- m t’ ´p l’ e e e
e e
e e e e
a. b. c. d. e. f. g.
MRoot /ts/ ,, /cq w / , /sq , / /k w ´/ /x w y/ /´t’/ /t’pl’/
e e e e e e e e e e e e
(33)
MRoot gloss ‘approach’ ‘get pierced’ ‘get torn, get split’ ‘spill over, tip over’ ‘wake up’ ‘rake herring’ ‘play cards’
264
Patricia A. Shaw
What is of immediate interest here is the locus of stress in the Progressive forms. The data of (31) established that the Progressive reduplicant is, at least potentially, in the stress domain. One might therefore expect a stress pattern like *[t ´-t s] (33a) or *[´ ´-´ t’- m] (33f ), parallel to the stress exhibited by MRoot-initial words beginning with a sequence of schwas—for example, [w ´w s] ‘tree frog’ (6f ) or [ ´n x w y t] ‘stop oneself ’ (9a). The di¤erence between the stress patterns in (31) and (33) shows that stress is not predictable simply on the basis of phonological properties of a word; reference to internal morphological domain structure is also essential. When, as in (33), nuclear weight is not present as a dominant factor, it is apparent that foot alignment to the MRoot is in force. The absence of lexically distinctive vowel quality in (33) means that Wgt-to-Str is not a factor, allowing the lower-ranked Al-L(Ft, MRt) (13b) to emerge. This is illustrated in the following tableaux, where Progressive reduplication of vowelless /CC/ roots in (34a) is compared with Progressive reduplication of /CVC/ roots in (34b). Recall that in the latter case, high-ranking MaxNucIR (32) ensures retention of the underlying vowel quality in the reduplicative prefix: e e
e e e
e e
e
e
[Prog- [t s
‘be approaching’ (MRoot ¼ /ts/) (¼ (33a)) Wgt-to-Str
e
i. [(t ´- [t s)
Al-L(Ft, MRt)
e
e
*
e
e
MRoot , [p t’ y , [Prog- [pet’ y , , i. [(p ´-[p t’ y ) , , ii. [pe-[(p ´t’ y ) , , + iii. [(pe´-[p t’ y )
Ft-Bin-s
*!
+ ii. [t - [(t ´s) b. MStem , [pe´-
e e
MRoot [t ´s
e
(34) a. MStem [t -
, ‘be sewing’ (MRoot ¼ /pet’ y /) (¼ (31a); cf. (32)) Wgt-to-Str
MaxNucir
Al-L(Ft, MRt)
*!
*
e
e
*! *
e
e
e
The incorporation of an MStem syllable into a stress foot in (34b) is the straightforward consequence of the previously well-established ranking of Wgt-to-Str above Al-L(Ft, MRt). The second set of data that confirms the role of the MRoot domain edge even in forms with MStem prefixes involves Plural/Distributive/Iterative (henceforth Pl) re,, , , duplication. Although an analysis of the full range of Pl allomorphy in h nq min m is beyond the scope of this chapter, two of its reduplicative realizations are directly relevant. Their stress behavior supports the morphoprosodic generalizations estabe
e
e
Inside Access
265
lished above. As shown in (35), Pl reduplication prefixes a CVC reduplicant. In stems with an underlying full vowel, high-ranked MaxNucIR preserves the full vowel in the reduplicant and results in stress on the reduplicative prefix, parallel to what was seen with Progressive reduplication in (31): Underlying MRoot , w w a. /k aq / b. /t’ax w / c. /t’ic m/
Reduplicated PL stem , , w´ w w w k aq -k q t’a´x w -t’ x w t’ı´c-t’ c m e
e e e
(35)
MRoot gloss ‘get hit’ ‘go/come down’ ‘swim’
e
Underlyingly vowelless roots exhibit a C C-reduplicative Pl prefix, as seen in (36). Parallel to the Progressive forms in (33), stress falls on the MRoot, rather than on the reduplicative prefix. e
Unreduplicated MRoot , , c ´q w l ´k w p ´n- t , n ´p- x
Reduplicated PL Stem , , , , c q w -c ´q w l k w -l ´k w p n-p ´n- t , , n p-n ´p- x e e e e
e e e e e e
a. b. c. d.
Underlying MRoot ,, /cq w / /lk w / /pn/ , /np/
e e e e e e
(36)
MRoot gloss ‘get pierced’ ‘get broken’ ‘bury -tr’ ‘eat, chew -tr’
The analysis of Progressive stress in (32) carries over directly to these Pl stems: in the absence of an underlying full vowel, Wgt-to-Str is not a factor, and Al-L(Ft, MRt) compels alignment of the foot to the MRoot. Compare the tableau in (37a), in which Wgt-to-Str stresses a full-vowel reduplicant, to that in (37b), in which the impact of the lower-ranked Al-L(Ft, MRt) emerges. (37) a. MStem , [k w a´q w -
MRoot , , [k w q w (MRoot ¼ /k w aq w / ‘get hit’) (¼ (35a)) , Wgt-to-Str Al-L(Ft, MRt) [Pl-[k w aq w , , * + i. [(k w a´q w - [k w q w ) , , ii. [k w aq w - [(k w ´q w ) *! e
e e
b. MStem , , [c q w -
MRoot , , ,, [c ´q w (MRoot ¼ /cq w / ‘get pierced’) (¼ (36a)) ,, Wgt-to-Str Al-L(Ft, MRt) Ft-Bin-s [Pl-[cq w , , , , *! i. [(c ´q w - [c q w ) , ,w , ,w * + ii. [c q - [(c ´q ) e
e e
e
e
e
In all of the data sets in (37), there is a single syllable to the left of the Root. An important question is whether Wgt-to-Str can draw Foot-headedness more than
266
Patricia A. Shaw
one syllable to the left of the Root. Interestingly, various kinds of evidence converge on the conclusion that a foot edge can be misaligned from the MRoot edge by no more than one syllable. One type of evidence leading to this conclusion derives from the interaction of the Plural -l- infix, realized as [- l-] or [-l -] depending on context,27 with CV-diminutive reduplication. Consider the following data: e
(t’ y a´t’ y - (l-a´(t’ y a´-l -)
e
h- (l- e´-
MRoot (t’ y a´m) t’ y m) t’ y m) t’ y m , (he´wt) , h wt) e
e
e
e
e
i. ‘rat’ ii. dimþpl
MStem
e
e
e
e
b.
MWord sss*s-
s-t’ya´m , s-t’ y a´-t’ y m , s-t’ y - l-a´-t’ y m , *st’ y a´-l -t’ y m , he´wt , h- l-e´-h wt
i. ‘bone’ ii. dim iii. dimþpl
e
a.
e e e
e
(38)
As seen in (38aiii,bii), the Plural infix disrupts the contiguity of the reduplicative Diminutive prefix (violating Contiguity). As a consequence, even when both Stem prefixes are present, the full vowel of the Diminutive prefix is not dislocated more than one syllable away from the MRoot edge. The ungrammaticality of the starred form in (38aiii) follows directly from the independently motivated ranking of Wgtto-Str over Al-L(Ft, MRt), as shown in (39).
Wgt-to-Str *!
e
*!
‘little bones’ (¼ (38aiii))
Al-L(Ft, MRt)
Contiguity
*
e
**!
e
e
e
e
e
+ iv.
[s- [dim-l- [t’ y am , [s [t’ y al [(t’ y ´m) , [s [t’ y la [(t’ y ´m) , [s [(t’ y a´l ) [t’ y m , [s [t’ y (la´ [t’ y m)
[st’ y la´t’ y m]
e
iii.
e
ii.
e
i.
MRoot t’ y am/ t’ y m)
e
b.
MStem dim-lt’ y -(la´-
e
(39) a. MWord ss-
*
*
All of the candidates in (39b) satisfy the high-ranked MaxNucIR (not shown) as well as a constraint, not formalized here, aligning the reduplicative Diminutive prefix with the left edge of the MStem. Candidates (39bi,bii) show that, regardless of whether Contiguity is maintained in the reduplicant or not, perfect alignment with the MRoot has the fatal consequence of violating Wgt-to-Str. Candidate (39biii) moves the foot head further left into the MStem domain, thereby maintaining Contiguity and respecting Wgt-to-Str, but at the expense of a double violation of Al-L(Ft, MRt). The candidate in (39biv) misses perfect foot alignment by one syllable, and the plural infix violates Contiguity, but it is optimal compared with the
Inside Access
267
alternatives. Misalignment by only one syllable leaves the bisyllabic foot straddling the MStem and MRoot domains. The only situation that could compel a foot to be realized fully to the left of the MRoot, as in losing candidate (39biii), would be a sequence of MStem prefixes the leftmost of which contained a full vowel. In such a case, Wgt-to-Str would overrule Al-L(Ft, MRt). However, the language conspires to avoid this scenario. The only source for a full vowel in an MStem prefix is reduplication; all realizations of a reduplicative full vowel appear immediately before the MRoot. The internal MRoot edge also plays a crucial role in stress foot alignment in forms where the MStem domain is polysyllabic but contains no full vowel. This situation is exemplified in the following data, where the Plural [- l-] @ [-l -] infix surfaces after the initial Progressive reduplicative prefix of an MRoot lacking a full vowel (Suttles 2004, 170): e
[prog- pl- MRoot [MRoot]] , ‘crack’ [s - l - [(s q) , ‘break’ [t’ - l - [(t’ q)
, [s l s ´q] , [t’ l t’ ´q]
‘several are cracking’ ‘several are breaking’
e ee e ee
e
e e e e
e
MStem , a. /sq/ , b. /t’q/
e
(40)
These forms have primary stress on the final odd-parity schwa-headed syllable, conspicuously inconsistent with the stress pattern previously encountered in trisyllabic words like those in (9), repeated in (41), that also have three schwa-headed syllables in a row: [ ´n x w y t] [´ ´q ll x w ] e e e e e
xw) y t w MRoot [(´ ´q l) l x MRoot [( ´n
e
e
e e e e
e
(41) a. b.
‘stop oneself ’ ‘know’
The stress di¤erences between (40) and (41) are predictable from their respective internal morphological structure, summarized below: .C . MRoot [(C ´C) MRoot [(C ´.C ) .C C
e e
MStem [
e e e
MStem [C
e
(42) a. Examples in (40): b. Examples in (41):
The apparent anomaly in (40) of having a degenerate (C ´C) foot at the right margin of a word thus follows straightforwardly from Al-L(Ft, MRt); what distinguishes (40) from (41) is that the final schwa syllable is initial in the MRoot. A tableau showing the optimal parse of data such as (40), following the established constraint ranking, is given in (43): , (43) s l s ´q ‘several are cracking’ (MRoot ¼ /sqª/) (¼ (40a)) , Wgt-to-Str Al-L(Ft, MRt) MStem [Prog-l- MRoot [sq , a. [(s `l ) [(s ´q) *!* , + b. [s l [(s ´q) e
e ee
e
ee
e
ee
268
Patricia A. Shaw
Finally, bringing together all of the claims of the proposed analysis, consider data where all three morphological domains have overt content. The Progressive reduplicative forms in (44a,b) show that when there is a full vowel in the MStem domain, it will attract stress through the ranking of Wgt-to-Str Al-L(Ft, MRt). The form in (44b)28 a‰rms that stress footing is iterative, with Rightmost assigning relatively greater prominence to a foot at the right edge of the word. In the form in (44c), stress does not fall on the Progressive reduplicant prefix, since it has no full vowel to bring Wgt-to-Str into play. The apparent failure in (44c) to parse the initial sequence of two schwa-headed syllables under Parse-s-2 (15) falls out from the hypothesis that MStem prefixes, like MWord prefixes (see (28)–(30)), are external to the domain of stress footing. Parsing weightless prefix syllables into feet would incur unnecessary violations of Al-L(Ft, MRt).
e
e
e
e e
e
eee
e
e
e
e
e
e e e
e
e
12.5
[t wı´t t] ‘like it’s sleeping’ (WS, 255) , , , , [y pe`p ¨’xe´n m] ‘feeling along with feet’ [w sw wı´st x w ] ‘it was in plain view’ (WS, 542) [st l t tı´w n] ‘a¤ectionate (dim.) term for siblings or living cousin’s children’ (WS, 479) e
d.
e
c.
MRoot [t t) sleep , , , [p ¨’) (¼xe´n- m) feel¼foot-intr [(wı´-st- x w ) reveal-caus-3agr [(tı´w n) sibling e
b.
MStem [(ı´prog, [(pe`prog[w prog[t - l - t Pl- Pl- dim-
e
a.
MWord [t wlike[y along[w -sest-s[snom-
e
(44)
Conclusions
Several types of argumentation have been presented in support of the general claim ,, , , that in h nq min m, word-internal morphological structure plays an independent and critical role in stress assignment. The focus of the present analysis has been the argument that prosodic constraints on the parsing and alignment of stress feet must make explicit reference to the left edge of morphological domains, specifically the MRoot domain, as summarized in (45): e
e
e
(45) [MWord Nonredup prefixes [MStem Redup prefixes [MRoot Lexical root . . . j
Footing domain: Al-L(Ft, MRt) Parse-s-2 The prosodic e¤ects of morphological su‰xes or clitics at the right edge remain unaddressed here, and constitute an important area for further research.29
Inside Access
269
The principal theoretical issue addressed in this chapter is to what degree reference to internal morphological structure such as the MRoot can be constrained. The Generalized Alignment schema by definition (McCarthy and Prince 1993a) allows reference to the domain edge of any PCat or MCat in constraint evaluation, without restriction on access to internally embedded hierarchical domains. In comparison with earlier and much more restrictive theoretical hypotheses like Bracket Erasure and Strict Cyclicity, the relative freedom of access at the Output level to internal constituency permitted by Generalized Alignment might seem unduly powerful. In ,, , , this case, the complex morphoprosodic interactions of h nq min m have provided compelling empirical support for the necessity of recognizing the internal left-edge boundary of the internal major lexical category MRoot. Significantly, Kiparsky’s (2000c) model of the phonology-morphology interface in OT requires that the internal morphological and prosodic constituency of Stem and Word both be accessible to phonology, thus allowing cyclic e¤ects to be captured through IO correspondence. The argumentation for including the MRoot as a fundamental morphological category that may be recognized at the Output level is patently plausible, given that the MRoot constitutes the fundamental lexical unit within any theory of morphology. In fact, if constraints on accessibility are to be constructed with reference to the basic constituent hierarchies of morphology and prosody, it would seem highly implausible for MRoot to have to be excluded. The inclusion of the MRoot in the list of morphological constituents to which phonology may refer also connects to crosslinguistic findings regarding cognitive processing and learnability. It is well known that prefixes frequently do not belong to the principal phonological constitutent of a morphological word (CzaykowskaHiggins 1998, 180; see also Cohn 1989 on Indonesian and Szpyra 1989 on Polish, ,, , , among others). H nq min m prefixes pattern similarly: with the systematic exception ,, of the minimal alignment violations documented in section 12.4.2, prefixes in h nq , , min m are not parsed into stress feet and hence are not part of the Prosodic Word. ,, , , Stress is demarcative in such systems; in h nq min m, a polysynthetic language with complex prefixal concatenation, stress footing functions to identify the major internal morphological category, the MRoot. Any theoretical model must provide clear and constrained mechanisms to capture this kind of ‘‘inside access’’ to morphological constituency. e
e
e
e
-e
e
e
e
e
e
e
e
Notes I am particularly indebted to Darin Howe, Jill Campbell, and Larry Grant for their invaluable contributions to the Musqueam research program. I also thank Luigi Burzio, Laura Downing, Sharon Hargus, Sharon Inkelas, Rene´ Kager, and an anonymous reviewer for helpful discussion, as well as Dale Kinkade, Wayne Suttles, Larry Thompson, and Ewa CzaykowskaHiggins for sharing their thoughts about Salish on many occasions.
270
Patricia A. Shaw
1. For issues related to the endangered status of this and other First Nations languages in British Columbia, see Shaw 2001a and Shaw 2004. ,, , , 2. These data represent the ‘‘downriver’’ dialect of the Hul’q’umin’um’-h nq min mHalkomelem continuum, as traditionally spoken in Musqueam and other villages of the Fraser River delta in British Columbia. This research has been conducted under the auspices of the Musqueam-UBC First Nations Languages program. I am profoundly grateful to the late Adeline Point, as well as to the late Dominic Point, the late Edna Grant, and the late Arnold Guerin Sr. for their insightful teachings. I thank my Musqueam colleagues for their collaboration and deep commitment to language revitalization. Additional data are cited from Suttles 2004 and Guerin n.d. (abbreviated WS and AG respectively). The language content belongs to the Musqueam people and the scholarship belongs to the author; the views presented here are those of the author and not necessarily those of the Musqueam Indian Band. e
e
e
3. Sounds in parentheses are marginal within the native lexicon; [sˇ] occurs only as a rounded backed allophone of /s/. Following NAPA (North American Phonetic Alphabet: see American Anthropologist 36(4), 1934) orthographic conventions, the symbol ‘‘c’’ represents an alveolar a¤ricate [t s ], ‘‘cˇ’’ represents an alveopalatal a¤ricate [tS], and ‘‘sˇ’’ an alveopalatal fricative [S]. Glottalization is represented by an apostrophe directly over the symbol, and labialization by a raised w immediately after. 4. Since Proto-Salish *u > a in the dialect groups to which Musqueam belongs (Boas and Haeberlin 1927, 131–132; Shaw 2001b), the short [u]s that are left in the synchronic lexicon derive primarily from loanwords. 5. Phonemic distinctions in vowel length carry a very low functional load; most, if not all, long vowels are analyzable as derived. 6. Due to space limitations, the present analysis does not extend to right-edge domains. Note that van Eijk (1998, 472) presents a model of morphological constituency for Lillooet that, on the basis of quite independent criteria, is directly comparable in its left-edge demarcations to those posited here. Czaykowska-Higgins (1998, 154, 172) posits a more substantive mismatch in prosodic versus morphological domains in Moses-Columbian Nxaamxcin Salish, particularly within the MStem/PStem domains. However, her conclusion that ‘‘stress is never assigned further left than the root morpheme’’ correlates importantly with the present analysis, requiring for Nxaamxcı´n simply a higher (inviolable) ranking of the same Align-L constraint pro,, , , posed here for h nq min m (see section 12.4). Black 1998 and Bar-el and Watt 2001 explore domains in other Salish languages. e
e
e
7. A characterization of the limited set of contexts where schwa is seemingly lexical is beyond the scope of the present chapter. For discussion of schwa in other Salish languages, see among many others Blake 1992, 1999, 2000; Kinkade 1998; Roberts and Shaw 1994; Urbanczyk 1996; Czaykowska-Higgins and Willett 1997; Rowicka 2000; and references therein. For the diachronic complexities, see especially Thompson 1979. 8. Whereas syllable, nucleus, and mora are treated here as theoretical primatives (Shaw 1992, 1993), the terms onset and coda hold no formal theoretical status, and are simply used mnemonically to refer to nonnuclear segments at the left and right edges of the syllable. 9. Abbreviations: agr ¼ agreement; dim ¼ diminutive; intr ¼ intransitive; ls ¼ lexical su‰x; pl.name ¼ place name; pro ¼ pronoun; prog ¼ progressive; nom ¼ nominalizer; n.tr ¼ ntransitivizer (limited control); obj ¼ object; pl ¼ plural; red ¼ reduplicant; refl ¼ reflexive;
Inside Access
271
s ¼ singular; su ¼ subject; t.tr ¼ t-transitivizer (control); ind ¼ independent; 1s ¼ firstperson singular; 3 ¼ third person; past ¼ past tense; instr ¼ instrumental; 2p ¼ second person plural. 10. Certain higher-ranked constraints will allow a sequence of full vowels to surface—for example, the constraint * ]s against a tautosyllabic schwa-glottal sequence blocks -reduction in words like /s-k w aya/ [sk w a´ya] ‘squirrel’. The basic left-aligned trochaic system prevails; see Shaw et al. 1999. , , 11. Suttles (WS, 140, 177) records this as [xa´-x w k w -]; our consultant gives [xa´-x k w -]. e
e
e
e
12. Contra various claims (e.g., Hayes 1995) that the canonical disyllabic trochee is evenly ,, , , balanced—that is, LL ¼ (m:m)—the h nq min m system actively optimizes uneven trochees (m.Ø). Note, however, that under the hypothesis that schwa is nonmoraic, reduction of a moraic vowel to schwa creates an optimal Strong-Weak rhythmic contour within the foot (cf. ,, , , Kager 1999, 173¤. and earlier work). The present analysis of h nq min m therefore predicts that uneven trochees may in fact be the least marked foot canon in other trochaic systems sensitive to a moraic full vowel–versus–nonmoraic schwa distinction as well. e
e
e
e
e
e
13. As motivated in section 12.4, this holds within the domain initiated by the MRoot. 14. See (26) ¤. for an analysis of systematic exceptions to this generalization. 15. See Shaw et al. 1999 for arguments against a foot-reversal analysis. 16. Unfooted syllables violate Parse-s-to-Foot (not included in the tableaux here), ranked below Foot-Bin-s. 17. There are various formalizations of this generalization, the basic intent being that one syllable may remain unparsed, but a sequence of two cannot. See especially Selkirk 1984 and Elenbaas and Kager 1999, among others. 18. The crucial ranking of Al-L above Ft-Bin-s is established in (34a). 19. See Suttles (2004) for a characterization of aspectual, modal, and derivational properties of these prefixes. Questions of apparent homonymy and cliticization remain important areas for further research. 20. See note 10 for explanation of the final unstressed [e] here. , , 21. The heteromorphemic l- sequence merges to [l] in fast speech: [t l ´n c ]. The posttonic schwa is variably realized or deleted in fast speech. e e ee
22. Although Suttles 2004 and Guerin n.d. both regularly mark primary stress, neither documents secondary stress consistently. The stress patterns in (26) and elsewhere were attested in my consultants’ speech and/or in Guerin’s (n.d.) documentation. 23. See Shaw 2002 for an analysis of word-initial clusters and epenthesis. 24. Although the collocation of more than one reduplicative prefix ,is, highly constrained, , , examples of Diminutive with Progressive—for example, [til m] ‘sing’, [t tit l m] ‘a little one (child) is singing’—are recorded by Suttles (72¤.), as well as Plural plus Diminutive forms, such as (38aiii,bii), from my own database. Such forms obey MaxNucIR in the Progressive and follow the stress generalizations described here. ee
e
e
25. Only Progressives formed by reduplication are considered here. See Hukari 1978 and Suttles 2004 for classification of other Progressives. ,, , , 26. IR faithfulness is not a property of all reduplicative prefixes in h nq min m; (32a) applies to the Progressive as well as to the Plural, in (35). Discussion of the implications of these facts e
e
e
272
Patricia A. Shaw
for recent claims about reduplicative faithfulness by Struike (2000) and Inkelas and Zoll (2000) is unfortunately beyond the scope of the present chapter. 27. See (40). A full analysis of the constraints governing the locus of the Plural infix is beyond the scope of this chapter (see Shaw 2003). The basic generalization is that the Plural infix is aligned as close as possible to the beginning of the MStem domain, or to the beginning of the MRoot if there are no MStem prefixes, subject to higher ranking constraints on foot alignment and reduplicant anchoring. The epenthesis of an adjacent schwa follows from independent constraints on syllabification. 28. The full gloss from Arnold Guerin Sr.’s lexical files is ‘to be feeling along with the feet (in shallow water looking for cockles)’. 29. See Shaw et al. 1999 for preliminary hypotheses regarding prosodic roles of lexical su‰xes.
13
Morphosyntactic Correspondence in Bantu Reduplication
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
13.1
Introduction
Since the pioneering work of Marantz (1982), Kiparsky (1986b), McCarthy and Prince (1986), and others, the primary goal in the study of partial reduplication has been to construct a theory that insightfully captures the full range of considerations that speakers may invoke in determining how a reduplicant will relate to its base. Given that both phonology and morphology are potentially involved, this has meant two things. First, there has been an attempt to characterize the reduplicant in prosodic terms: the shape of a reduplicant is frequently defined by reference to foot, syllable, and/or moraic structure. Second, the literature has shown an increasing awareness of the role of morphological structure in determining the base-reduplicant relationship. Researchers such as Downing (1997a, etc.), Urbanczyk (1996), and McCarthy and Prince (1993b, 1995) have shown that, in addition to prosodic constraints, the realization of a reduplicant may be influenced by purely morphological conditions. In Bantu verb-stem reduplication, for example, simplex stems may reduplicate di¤erently from polymorphemic ones, which may show further di¤erentiations in turn, depending on whether the a‰xes are derivational versus inflectional in nature. Some of these morphological conditions can be quite subtle, and yet, as we will show, provide crucial evidence for our very conception of how and where (partial) reduplication takes place within a grammar. In short, reduplication provides an ideal testing ground for theories of morphology, phonology, and their interface. In this chapter we provide a detailed description of verb-stem reduplication in Ndebele, a Southern Bantu language of the Nguni group, which also includes Zulu, Xhosa, and Swati. We show that the reduplicant in Ndebele is not only conditioned by phonological and morphological factors, as in other Bantu languages, but that these factors are ‘‘abstract’’ in nature: despite surface appearances to the contrary, the reduplicant of an Ndebele verb stem must be analyzed as a verb stem itself (cf. Downing 1997a, etc.). Its surface form is obtained not by surface correspondence to the base output, but rather by direct spell-out of its own (identical) morphosyntactic
274
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
structure, which, in turn, is a direct copy from the base. As schematized in the (typically) prefixal reduplication in (1), (1) Morphosyntactic representation (base)
the input is a left-branching morphosyntactic structure, where the deepest embedded morpheme is the root—for example, a verb root in our study. The surface outputs to the right of the arrow are obtained in the following way. First, we copy this morphosyntatic representation as a reduplicant to the left of the input base. We then spell out both structures independently. What we have identified as the surface base is derived by the normal rules of word formation, constrained by (sometimes conflicting) considerations such as compositionality, morphotactic restrictions, and so on. The surface reduplicant is also subject to these general considerations, as well as those imposed specifically on the reduplicant. The reduplicant may thus be subject to prosodic as well as morphological constraints. In Ndebele, for instance, the reduplicant is limited to two syllables and inflectional su‰xes may not appear within it. Our conception of verb stem reduplication in Ndebele in (1) di¤ers from most other conceptions in two ways. First, we make explicit that reduplication is a morphological process. While no one would contest this conclusion, research on partial reduplication has mostly been conducted by phonologists who emphasize surface base/reduplicant correspondence, and hence view morphology in terms of surface morphs rather than morphosyntactic structure. With Inkelas and Zoll (2005) we do not assume a direct phonological correspondence between the reduplicant and the base. Second, we take the position that partial reduplication is derived from total reduplication. The conception in (1), however, goes beyond phonological ‘‘full-copy’’ theories such as Steriade 1988, in explicitly treating reduplication as the total copy of the abstract morphosyntactic structure of the base. If there are no special phonological or morphological conditions on the reduplicant beyond those characterizing the base, we in fact derive total reduplication on the surface—an apparent compounding of a base with itself. If there are special conditions, we obtain partial reduplication. We thus agree, in part, with Eulenberg (1971, 73), who states that ‘‘cases of so-called partial reduplication are simply phonological reductions, sometimes drastic, from cases of full reduplications.’’ As will be seen below, we propose to revise Eulenberg’s statement to read ‘‘phonological or morphological reductions.’’ In many, if not most cases of partial reduplication, there will be no di¤erence between our morphosyntactic approach versus the ‘‘morph’’ approaches that have
Morphosyntactic Correspondence in Bantu Reduplication
275
characterized the prosodic analysis of reduplication since Marantz 1982 and McCarthy and Prince 1986. This is because in most situations the properties (e.g., linear order) of surface morphs (specifically, a‰xes) generally mirror the underlying morphosyntactic representation. In Bantu, however, there are widespread instances where this is not the case. As we will see in section 13.5, when there is a mismatch between the surface order of su‰xes versus the underlying morphosyntactic representation, it is the latter that determines what can appear in the reduplicant. The chapter is organized as follows. In section 13.2 we provide a basic overview of verb-stem reduplication in Ndebele. In subsequent sections we treat complications arising in the reduplication of stems containing subminimal or ‘‘consonantal’’ verb roots (section 13.3), fusion or ‘‘imbrication’’ of perfective -ile (section 13.4), and the passive su‰x -w- (section 13.5). We then conclude by considering synchronic and diachronic implications of our findings. 13.2
Basic Overview
We begin by considering the basic properties of verb-stem reduplication in Ndebele. In (1) we first consider verb stems that consist of a ‘‘long’’ root (bCVC) and the default inflectional final vowel su‰x -a. The meaning of such reduplications generally is to do the action in little bits, here and there, perhaps not very well. The forms are given in their minimal citation form—that is, minus inflectional prefixes (and with underlying High tone marked on the first vowel of verb roots). With one major exception that we will examine in section 13.3, prefixes are irrelevant to stem reduplication. We will adhere to the notational convention of separating reduplicant and base with the ‘‘þ’’ symbol, reserving ‘‘’’ for internal morpheme breaks:1 (2)
Plain verb stem a. lim-a thum-a (H) b. nambith-a (H) thembuz-a (H/L)
Reduplicated verb stem lim-aþlim-a thum-aþthum-a nambiþnambith-a thembuþthembuz-a
‘cultivate’ ‘send’ ‘taste’ ‘go from wife to wife’
The pattern in (2a), in which the verb root ¼ CVC, shows a total reduplication of the verb stem. In (2b), however, where the verb root > CVC, we see that the reduplicant is limited to two syllables—that is, to a bisyllabic foot. In both sets of examples the reduplicant is identical to the first two syllables of the base verb stem. However, when we turn to cases of productively su‰xed roots in (3), we find that something more subtle is going on. In these examples, where -el- is the applicative su‰x and -is- is the causative su‰x, we see that there are two possible shapes of the reduplicant: it can either be identical to the first two syllables of the base, hence lim-e and lim-i, respectively, or it can end in [a], thus lim-a.
276
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
(3) a. lim-el-a b. lim-is-a
lim-eþlim-el-a lim-aþlim-el-a lim-iþlim-is-a lim-aþlim-is-a
‘cultivate for/at’
(applicative -el-)
‘make cultivate’
(causative -is-)
To account for this variation, consider the traditional view of the internal structure of the Bantu verb stem in (4). (4)
cf. Meeussen 1967
As seen, a verb stem consists of a base and an obligatory final vowel (FV) morpheme, which is [a] in most verb forms. Within what Bantuists refer to as the ‘‘base,’’ a root may be ‘‘extended’’ by derivational su‰xes (or extensions). Among the Ndebele extensions that we consider in this study are applicative -el-, causative -is-, and passive -w-. The second variants in (3a) and (3b) show that the extension vowel may optionally not appear in the reduplicant, in which case the reduplicant is pronounced with a final [a]. As seen in (5a), this second variant ending in [a] is not available if the verb root is polysyllabic: (5) a. nambith-a (H) *namb-aþnambith-a thembuz-a (H/L) *themb-aþthembuz-a b. casuk-a casuþcasuk-a *cas-aþcasuk-a casul-a casuþcasul-a *cas-aþcasul-a
‘taste’ ‘go from wife to wife’ ‘become nauseated’
/cas-uk-a/
‘nauseate’ (transitive)
/cas-ul-a/
A final [a] is also not possible in (5b), where the input verbs carry the nonproductive reversive su‰xes -uk- and -ul-. Finally, to complete this introduction to the basics, note in (6) that the reduplicant cannot include inflectional material occurring in the so-called final vowel slot: (6) a. lim-e
!
b. lim-i
!
c. lim-ile
!
lim-aþlim-e *lim-eþlim-e lim-aþlim-i *lim-iþlim-i lim-aþlim-ile *lim-iþlim-ile
(subjunctive -e) (negative -i) (perfective -ile)2
Morphosyntactic Correspondence in Bantu Reduplication
277
These examples are significant for two reasons. First, they show that inflectional suffixes are outside the scope of reduplication. Second, they show that the source of reduplicant-final [a] is not necessarily the base, as one might have presumed from [a]-final reduplicants in (3). The data in (6) make it clear that there is an independent source for reduplicant-final [a]. In all of the above respects, Ndebele verb-stem reduplication is equivalent to that in Swati, for which Downing (1997a, etc.) proposes the Bantu verb-stem structure in (7). (7) Bantu verb stem according to Downing; cf. Mayers 1987 a. I-stem ¼ Inflectional stem b. D-stem ¼ Derivational stem c. Root ¼ minimal D-stem d. Xþa ¼ ‘‘canonical stem’’ (CS)
As seen, the full verb stem is referred to as an inflected stem (or I-stem), which in turn has two parts: (i) a D-stem, which may be potentially extended by derivational suffixes, and (ii) an inflectional final su‰x. The root is identified as the minimal D-stem. In (7) this structure is exemplified with the root lim- ‘cultivate’, the applicative and causative extensions -el- and -is-, and the inflectional final vowels -e, -i, -ile, and -a. Downing refers to any verb stem that ends in the final vowel -a as a ‘‘canonical stem.’’ To summarize the facts we have seen thus far, and as indicated in the last line of (7), the root -lim- must be copied into the reduplicant, the vowels [e] and [i] of applicative -el- and causative may optionally copy, and the [e] and [i] of the inflectional su‰xes -e, -i, and -ile may not copy. If the structure of the input ‘‘base’’ stem is as in (7), the analytical challenges before us are the following. First, by what means do we ensure that only derivational su‰xes, not inflectional ones, are available to reduplicate?
278
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
Second, by what means do we ensure that the CVC-a reduplication pattern is permitted with CVC-VC-V stems but not with CVCVC-V stems? That is, why are forms such as lim-aþlim-el-a and lim-aþlim-is-a acceptable, while *namb-aþnambith-a and *themb-aþthembuz-a are not? Downing’s (1997a,c) position is that in the case of CVC-a reduplicants, the CVC- string must correspond to an independently existing minimal D-stem, which lim- clearly is. CVCa RED must be a corresponding word (cf. imperative lima!). On this account, the reduplicants *namb-a and *themb-a are disallowed because namb- and themb- do not exist as roots, and therefore the requisite minimal D-stems, *namb-a and *themb-a, do not independently exist. Steriade (1997) slightly amends the story to refer to existing versus nonexisting words, noting that lim-a exists as a corresponding imperative verb, while *namb-a and *themb-a do not. In this chapter we do not speak of a reduplicant as corresponding to a stem. Rather, it is a stem, which is in morphosyntactic featural agreement with the following, ‘‘normal’’ stem that it appears to reduplicate. As seen in (8), (8)
reduplication is stem juxtaposition, where Stem1 , the reduplicant, is subject to a bisyllabic size constraint, and, as indicated by the subscript on its morphosyntactic structure, is in perfect featural agreement with Stem2 , the base. At this point we note that there are two inviolable morphological properties of base verb stems: First, verb stems must contain a verb root. Second, verb stems must be morphologically complex. The first property is self-evident: one cannot have a stem of any sort that does not in turn consist of a root. The second property is what interests us here: Verb stems in Ndebele (and in most Bantu languages) must be bipartite (i.e., su‰xed). That is, there must not be a right alignment of the verb root with the verb stem: *[[verb]root ]stem . While it has been hypothesized that preProto-Bantu may not have required a su‰xal vowel on all verb roots (Gre´gorie 1979), verb stems in Ndebele must end in one of the inflectional endings—the most generally distributed one being -a. In this study we follow the tradition of identifying -e, -i, and -ile as inflectional endings—that is, as Downing’s IFS. We will depart from previous scholarship, however, in treating -a not as an inflectional ending, but rather as the default stem su‰x in Bantu in general:
Morphosyntactic Correspondence in Bantu Reduplication
279
(9)
As indicated, -a may be invoked, typically in final position, to ‘‘fill out’’ any kind of stem: D-stem, I-stem, R-stem. The su‰x -a di¤ers from other verb and noun endings in having no corresponding morphosyntactic feature(s); unlike -e, -i, and -ile, it does not directly realize tense/aspect, mood, or polarity distinctions on verbs. While these morphs spell out feature complexes that include [þsubjunctive], [þnegative], or [þperfective], -a appears only in their absence—that is, as a default. We will henceforth refer to -a as the FV, reserving IFS for the other final su‰xes. With this interpretation of -a established, we can now explain why -a can appear in the reduplicant, while IFSs may not: the latter may occur only in an I-stem, while the reduplicant is a D-stem in Ndebele. Since -a is devoid of inflectional features, it may appear as a default ending on any kind of stem—for example, the reduplicant D-stem, which, by definition, will not carry an IFS. Our proposal in (10) is that the R-stem is a daughter of Downing’s I-stem and sister to her IFS: (10)
Each of the stems under the R-stem in (10) is a D-stem, although the reduplicant is limited to two syllables. We assume, as is the case in D-stems, that -a is always available to reduplicants as well. As seen in (11), this gives rise to the variation seen in reduplication of D-stems that have extension su‰xes. We assume, for the sake of argument, that the FV is present even in (11a), where (due to disyllabic size constraints) it is not parsed: (11) Reduplicant D-stem of base -lim-is-a ‘cause to cultivate for’ a. b.
In (11a), the [i] of -lim-i comes from the causative su‰x -is-, while in (11b) it comes from the FV -a.
280
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
Now note with respect to the articulated structure of the reduplicated verb stem in (10) above that both D-stem2 and the I-stem should have to branch. In case D-stem2 is simplex (e.g., -lim-), it would appear to need the FV -a. However, this morpheme vies for the same ‘‘slot’’ as the IFSs -e, -i, and -ile. Therefore, since -a is featureless, the higher-ranked requirement that the IFS be spelled out takes precedence and, in the subjunctive, we get [[lim-] -e], rather than [[lim-a]]. The fully articulated structure of reduplicated verb stems is seen in (12). (12)
The crucial point in (12) is the disassociation of the FV morpheme -a from the IFS category to which all previous studies have assigned it. As seen, the FV -a is considered to belong to the D-stem, while the inflectional endings -e, -i, and -ile, which mark certain subjunctives, negatives, and perfectives, are a property of the I-stem. Since it is a featurally empty morph, the FV -a is able to occur in D-stemRED without disrupting the required morphosyntactic featural agreement between DstemRED and D-stemBASE. Assuming that the morphological makeup of the reduplicant is as given above, we now turn to its phonological requirements, also indicated in (12). We assume a set of statements roughly like the following, which are phrased for convenience in the terminology of Optimality Theory (Prince and Smolensky 1993). They are part of the cophonology (phonological ‘‘level’’; see, e.g., Inkelas, Orgun, and Zoll 1997) of Stem1 in the reduplication construction. They ensure that the reduplicant is of the proper size and determine which elements of the morphological material under the reduplicative D-stem will actually instantiate the bisyllabic reduplicant. (13) D-stemRED cophonology a. [ss] The reduplicant is bisyllabic (¼ a minimal prosodic word) b. MAX(Root) The reduplicant should parse the root (cf., e.g., McCarthy and Prince 1993; Urbanczyk 1995; Futagi 1997)
Morphosyntactic Correspondence in Bantu Reduplication
281
c. MAX(Ext) The reduplicant should parse extension su‰xes (if any) d. MAX(FV) The reduplicant should be faithful to FV -a (if any) The first constraint requires the reduplicant to assume the exact shape of the minimal prosodic word in Ndebele: one foot consisting of two syllables. This is imperative, and causes the truncation we have observed above. (Later we will see cases where this same constraint causes augmentation.) The second through the fourth constraints are violable, with Max(Root) ranked above the other two. The last two statements in (13) thus come into play when the root does not exhaust the bisyllabic reduplicant (i.e., when the root is smaller than CVCV, e.g., the CVC root lim-). As we have seen, there are two options for realizing the bisyllabic reduplicant in such cases: use material from extension su‰xes, if any (as is done in, e.g., lim-eþlim-ela), or use the FV morph, -a (e.g., lim-aþlim-el-a). The equal viability of these two options can be expressed in OT by freely ranking the corresponding constraints (as proposed by Anttila 1997 and Ito and Mester 1997, among others). The rankings of the constraints in (13) are given in (14). (14) [ss] Max(Root) Max(Ext), Max(FV) The tableaux in (15) and (16) show how these constraints correctly predict the CVCV- and CVCa- reduplicants of extended CVC roots. (Note our assumption that the filler morph [a] is present in the input.3) (15) Derivation of reduplicant in lim-eþlim-el-a lim-el-a
[ss]
Max(Root)
Max(Ext)
Max(FV)
*(l)
*
+ a. lim-e b. lim-a
**!(el)
(16) Derivation of reduplicant in lim-aþlim-el-a lim-el-a a. lim-e + b. lim-a
[ss]
Max(Root)
Max(FV)
Max(Ext)
*!
*(l) **(el)
In neither tableau is there a violation of [ss] or Max(Root), since the reduplicants are both bisyllabic and incorporate the entire verb root lim-. In (15), Max(Ext) is ranked higher than Max(FV). The winning candidate is thus (15a), lim-eþlimela, since the reduplicant in (15b), lim-a, could have parsed the [e] of the stem lim-el-a, but did not. In (16), the constraint Max(FV) is ranked higher than Max(Ext). In
282
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
this case the winning candidate is (16b), lim-aþlim-el-a, since (16a), lim-eþlim-el-a, violates the now higher-ranked constraint Max(FV). We thus see how the free ranking of the two lowest constraints produces the two reduplicant possibilities. Now compare these results with the corresponding tableaux in (17) and (18), where the input nambith-a has a bisyllabic root: (17) Derivation of reduplicant in nambiþnambitha nambith-a
[ss]
+ a. nambi
Max(Root)
Max(Ext)
*(th)
b. namb-a
Max(FV) *
**!(ith)
(18) Derivation of reduplicant in nambiþnambitha nambith-a
[ss]
Max(Root)
+ a. nambi b. namb-a
Max(FV)
Max(Ext)
* *!*(ith)
As seen, *namb-aþnambith-a is ruled out in both (17b) and (18b), because it violates the relatively high ranked Max(Root) constraint—that is, the [i] of the root nambithcould have been parsed but was not. To summarize, we analyze the reduplicant as a D-stem that is constrained in two ways: (i) it is bisyllabic, and (ii) it must match the base D-stem morphosyntactically. Our analysis resembles those of Downing (1997a, etc.) and Steriade (1997) in relating the reduplicant to an existing morphological constituent: like Downing, we take the D-stem to be the relevant level. These studies di¤er from ours, however, in relying on output-output correspondence to relate the reduplicant and base. Thus, in order to get the FV -a in the reduplicant, it is necessary that the preceding -CVC- be an existing verb root in the language. In the following sections we will see that this condition is neither necessary nor su‰cient to predict the full range of facts in Ndebele reduplication. We thus consider, in turn, subminimal (section 13.3), imbricated (section 13.4), and passivized (section 13.5) verb forms. In each case we will observe the overriding importance of reference to roots versus reference to other elements. The unmistakable generalization will be that as long as the root is fully parsed, one has great freedom in how the two syllables of the reduplicant are filled out in Ndebele. 13.3
Subminimal Roots
Up until now we have been considering cases in which the reduplicant must be phonologically truncated in order to achieve the bisyllabic size condition. In this section
Morphosyntactic Correspondence in Bantu Reduplication
283
we consider the reduplication of verb stems that contain phonologically subminimal verb roots such as those in (19a), in which entirely di¤erent behavior is observed. (19) a. ‘‘Consonantal roots’’ -dl- (H) ‘eat’ -m- (H) ‘stand’ b. Monosyllabic stems -dl-a (H) ‘eat’ -m-a (H) ‘stand’
-lw- (H/L) -z- (H/L)
‘fight’ ‘come’
-lw-a (H/L) -z-a (H/L)
‘fight’ ‘come’
As seen in (19b), when these so-called consonantal roots are followed by the FV -a, the resulting stems are monosyllabic. Normally these stems would occur with prefixes. Compare, however, the imperative forms in (20). The singular a‰rmative imperative is the only context in which a stem can occur without prefixes: (20) a. lim-a thum-a (H) b. nambith-a (H) thembuz-a (H/L) c. yi-dl-a (H) yi-m-a (H)
‘cultivate!’ ‘send!’ ‘taste!’ ‘go from wife to wife!’ ‘eat!’ (*dl-a) ‘stand!’ (*m-a)
bamb-a hlamb-a bhavum-a thembis-a (H) yi-lw-a (H/L) yi-z-a (H/L)
‘catch!’ ‘swim!’ ‘growl!’ ‘promise!’ ‘fight!’ (*lw-a) ‘come!’ (*z-a)
The examples in (20a,b) show that the unprefixed verb stem is used in the imperative (singular, a‰rmative), if the verb stem is at least two syllables long. In (24c), however, we see that a monosyllabic stem cannot occur in its bare form in the imperative, but rather acquires an augmentative syllable yi-, known as a ‘‘stabilizer’’ in the Bantu literature, which we analyze as a semantically empty ‘‘filler’’ morph, much like the FV -a. The motivation for augmentation with yi- in (24c) is clear: as in many other Bantu languages (Myers 1987a; Kanerva 1989; Mutaka and Hyman 1990; Downing 1998), Ndebele words are subject to the bisyllabic minimality condition in (21). (21) Minimal prosodic word (o) ¼ [s s]foot Verbs are required to meet the minimal prosodic word condition in (21) and must thus be augmented by the empty morph yi- where subminimality would otherwise result. This situation arises only with monosyllabic stems in the imperative, where there is no prefix to supply the required second syllable. This minimality condition also helps us make sense of the similar size condition on reduplicants: reduplicants must be minimal prosodic words as well. Reduplicated verbs thus display the nested prosodic word structure depicted in (22).
284
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
(22)
That is, the prosodic structure of a reduplicated verb is: [Prefixes [RED]o BASEþIFS]o . 13.3.1
Subminimality and Augmentative -yi
With this established, we now turn to the question of how monosyllabic verb stems are reduplicated. (23) shows that the augmentative [yi] makes its appearance here as well, this time as a su‰x. As shown in square brackets in (23), -yi augments the otherwise subminimal reduplicants of such stems: (23) [dl-a-yi]þdl-a (H) [m-a-yi]þm-a (L) [lw-a-yi]þlw-a (H/L) [z-a-yi]þz-a (H/L)
‘eat!’ ‘stand!’ ‘fight!’ ‘come!’
As seen, -yi appears between the two occurrences of the monosyllabic stem. The (bracketed) reduplicant is subject to the two-syllable requirement, to whose satisfaction the su‰xal -yi is a crucial contributor. The absence of a corresponding -yi in D-stemBASE is no mystery on this account, because D-stemBASE constituents are never subject to the minimality condition that motivates the presence of [yi] in the first place. We know that D-stemRED is required to be exactly bisyllabic. As was seen in (26), the verb as a whole is also required to be minimally bisyllabic, but in the above examples, the whole verb would be bisyllabic anyway even without [yi]. It is thus only the reduplicant that would otherwise be subminimal. The data in (24), which contain the infinitive prefix uku-, confirm that the minimality condition is imposed on the reduplicant and on the whole verb but not on ordinary verb stems. Here, the existence of a prefix blocks [yi] augmentation in the unreduplicated forms, because the verb as a whole is at least bisyllabic. The fact that the verb stem is monosyllabic is of no relevance. When reduplication is present, however, [yi] augmentation does occur. It is required to supplement the otherwise
Morphosyntactic Correspondence in Bantu Reduplication
285
subminimal reduplicant. In (28), constituents subject to prosodic minimality are show in brackets: (24) Verb with unreduplicated stem [uku-dl-a] [uku-m-a] [uku-lw-a] [uku-z-a]
Verb with reduplicated stem [uku-[dl-a-yi]þdl-a] [uku-[m-a-yi]þm-a] [uku-[lw-a-yi]þlw-a] [uku-[z-a-yi]þz-a]
‘to ‘to ‘to ‘to
eat’ stand’ fight’ come’
(*uku-yi-dla) (*uku-yi-ma) (*uku-yi-lwa) (*uku-yi-za)
The greatest relevance of consonantal roots for our analysis comes from their behavior under su‰xation. D-stems consisting of consonantal roots with extension suffixes show a dazzling variety of reduplication possibilities, as illustrated below: (25) a. uku-dl-a (H) b. uku-dl-el-a
c. uku-dl-is-a
‘to eat’ ‘to eat for/at’
‘to feed’
uku-[dl-a-yi]þdl-a uku-[dl-el-a]þdl-el-a uku-[dl-a-yi]þdl-el-a uku-[dl-e-yi]þdl-el-a uku-[dl-is-a]þdl-is-a uku-[dl-a-yi]þdl-is-a uku-[dl-i-yi]þdl-is-a
(cf. [lim-e]þlim-el-a) (cf. [lim-a]þlim-el-a) (cf. [lim-i]þlim-is-a) (cf. [lim-a]þlim-is-a)
Again, reduplicants are bracketed. In (25a), a form with no extension su‰xes, the input to the reduplicant cophonology consists of the consonantal verb root dl-, plus freely available empty morphs -yi and FV -a. As seen, there is only one possible outcome, which is to supplement the consonantal root with both empty morphs, yielding a reduplicant of the shape [dl-a-yi]. In (25b,c), in which the base contains a -VC- extension su‰x, the same reduplication pattern seen in (25a) is still possible: in the reduplicant, -yi supplements the root and final vowel, yielding in both cases [dl-a-yi]. But in addition, two further reduplication possibilities emerge. One possibility is to exhaustively parse the extension suffix, yielding, in (25b), [dl-el-a], and in (25c), [dl-is-a]. The other possibility is to take the vowel from the extension su‰x and use -yi to supply the remaining syllable, thus [dl-e-yi] and [dl-i-yi]. In summary, the use of -yi is on a par with the use of the FV -a and material from extension su‰xes. Any of these options (singly, or in combination) may be used to supplement a consonantal root and bring a reduplicant to the bisyllabic goal: -yi, -a, material from extension su‰xes. Our existing analysis needs to be supplemented with a constraint that makes reference to -yi. Our intuition is that -yi is a morphosyntactically empty morph that can fill an optional position in D-stemRED just like the FV morph -a.4 In parallel with our treatment of -a, for which we proposed the constraint Max[FV] in (13), we invoke the constraint in (26) which mandates the surfacing of the filler su‰x -yi.
286
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
(26) MAX(-yi) Parse the -yi su‰x. Ranked freely with Max(FV) and Max(Ext), this constraint accounts for the new reduplication patterns we have just seen. In the input in (27), both -a and -yi are assumed to be present in input, just to illustrate how the constraints pick and choose among daughters of D-stemRED: (27) Derivation of reduplicants in dlela-dlela, dleyi-dlela, and dlayi-dlela dl-el-a-yi
RED¼ [ss]
a. dl-el-a-yi
Max(Root)
Max(Ext)
Max(FV) Max(-yi)
*!
+ b. dl-el-a
*
+ c. dl-e-yi
*(l)
+ d. dl-a-yi
**(el)
*
The candidate in (27a) loses because it violates the highest-ranked constraint, RED¼[ss]. For each of the candidates in (37b–d), there is some ranking of the bottom three constraints such that that candidate will emerge as the winner: candidate (b) wins when Max(Ext), Max(FV) Max(yi); candidate (c) wins when Max(yi) Max(Ext) Max(FV); candidate (d) wins when Max(FV), Max(yi) Max(Ext). The analysis correctly predicts that -yi will not surface when the root is su‰ciently large. Even if -yi is present in the input, as indicated in (28), Max(Root) forbids it to supplant root material. (28)
RED¼ [ss]
Max (Ext)
Max (FV)
Max (-yi)
+ a. lim-eþlim-el-a
*(l)
*
*
+ b. lim-aþlim-el-a
**(el)
lim-el-a-yi
Max (Root)
c. li-yiþlim-el-a
*!(m)
***(mel)
d. l-a-yiþlim-el-a
*!**(lim)
****(lime)
* *
The candidates augmented with -yi in (28c,d) are ruled out immediately by Max(Root), leaving the freely ranked Max(Ext) and Max(FV) to decide between the two attested candidates, (28a,b). While yielding easily to our analysis, [yi]-augmentation poses problems for the canonical stem and word-based accounts of Downing and Steriade. First the redupli-
Morphosyntactic Correspondence in Bantu Reduplication
287
cant [dla-yi] in (27d) is clearly not an independent word, and is thus problematic for Steriade. [dla-yi] could be considered a canonical stem by Downing, if -yi is ignored. However, in (27c) there is no definition of either the canonical stem or the independent word that the truncating reduplicant [dle-yi] would meet. Even if -yi is stripped o¤, what remains behind, namely [dl-e], is not a possible corresponding stem or word, because it contains some but not all of the applicative su‰x, and no FV.5 It is because our analysis is sensitive to the internal structure of the D-stem, distinguishing roots, extension su‰xes, and filler su‰xes, that we are able to account for reduplication patterns in which the reduplicant contains material other than what is in the ‘‘base,’’ including cases where it does not correspond to a possible or existing stem or word in the language. 13.3.2
Subminimality and the Macrostem
Before leaving the topic of subminimal roots, we turn to two more sets of data that provide further evidence for the view of reduplication developed in this chapter. The first concerns an alternative realization of reduplicated imperatives of subminimal verbs. Consider the forms in (29). (29) Bare stem -dl-a -m-a -lw-a -z-a
Imperative yi-dl-a yi-m-a yi-lw-a yi-z-a
Reduplicated imperative yi-dl-aþyi-dl-a yi-m-aþyi-m-a yi-lw-aþyi-lw-a yi-z-aþyi-z-a
‘eat!’ ‘stand!’ ‘fight!’ ‘come!’
Cf. from (23) dl-a-yiþdl-a m-a-yiþm-a lw-a-yiþlw-a z-a-yiþz-a
This method of reduplicating imperatives di¤ers in two ways from that seen previously. First, augmentative yi- is initial, rather than final, in the reduplicant. Second, it also appears in the base. Similar facts occur in the reduplication of bases with consonantal roots that are su‰xed with the perfective -ile, which occupies the IFS position, outside the domain of reduplication. As (30a) shows, a prefixed, unreduplicated base of this sort—here, m-ile—is not augmentable with [yi]. Yet under reduplication, we find the same two alternatives as in the imperative: either the reduplicant is [yi]-final, and the base is not augmented, as in (30b), or the reduplicant and base are both [yi]-initial (30c). (30) a. (ba-) m-ile (H) b. (ba-) m-a-yiþm-ile c. (ba-) yi-m-aþyi-m-ile
‘(they) stood’
*(ba)-yi-m-ile (single, su‰xed -yi in reduplicant) (double, prefixed yi- in reduplicant and base)
There are two ways of thinking about the pattern in which both base and reduplicant show initial yi- of the type seen in (29) and (30c). One is that the presence of yi- is driven by a prosodic requirement on the base, being passively reflected in the
288
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
reduplicant by some kind of correspondence requirement. The other is that yi- is required by the reduplicant itself, and must then by correspondence appear in the base as well. The latter scenario can be termed morphological ‘‘backcopying,’’ on analogy to the transmission of phonological alternations from reduplicant to base, documented in a small number of reduplication examples by McCarthy and Prince (1995) (but see Kiparsky 1997d and Inkelas and Zoll 2005 for reanalyses). The evidence would appear to favor backcopying in this instance. First, it would be di‰cult to maintain the base-augmentation analysis, given our earlier finding that prosodic minimality is enforced on reduplicants and on whole words but not otherwise on verb stems, as was schematized in (22). Recall that the evidence for this was the behavior of unreduplicated monosyllabic verb stems, as in (24): these are not supplemented by [yi]. If the stem formed a domain for prosodic minimality, an assumption necessary to maintain the reduplication analysis, then the facts in (29) and (30) would be unexplained. The second piece of evidence for backcopying comes from the behavior of bases preceded by a prefixal object marker (OM), such as noun class 10 zi-, the one used in all examples herein. If, and only if, the root of a reduplicated stem is consonantal, then zi- (and other OMs) can be reduplicated, appearing in both reduplicant and base. In (31) we present a case of OM doubling where the consonantal root is not followed by extension su‰xes. The reduplicant is enclosed in brackets: (31) uku-zi-dl-a (H) ‘to eat them’ uku-zi-[dl-a-yi]þdl-a (OM is not reduplicated) uku-[zi-dl-a]þzi-dl-a (OM is reduplicated) These data are quite parallel to the yi-prefixation data seen earlier: exactly when there is room in the reduplicant for the OM, the OM may appear doubled in the base. Its appearance in the reduplicant is being driven by the bisyllabic prosodic size requirement on the reduplicant; the base (here, dl-a) would be well formed without the OM, but the reduplicant would not. Unlike prefixal yi-, OMs can be incorporated into the bisyllabic reduplicant, and ‘‘backcopied,’’ even when the consonantal root is followed by extension su‰xes. As (32) shows, in such cases, the reduplicant can be brought up to two syllables by incorporating either the OM, the extension su‰x, the FV -a (or both), or su‰xal -yi can be used to bring the reduplicant up to two syllables: (32) a. uku-zi-dl-el-a (H)
b. uku-zi-dl-is-a (H)
‘to eat them for/at’
‘to feed them’
uku-zi-[dl-el-a]þdl-el-a uku-zi-[dl-a-yi]þdl-el-a uku-zi-[dl-e-yi]þdl-el-a uku-[zi-dl-a]þzi-dl-el-a uku-[zi-dl-e]þzi-dl-el-a uku-zi-[dl-is-a]þdl-is-a uku-zi-[dl-a-yi]þdl-is-a uku-zi-[dl-i-yi]þdl-is-a
(OM not reduplicated)
(OM is reduplicated) (OM not reduplicated)
Morphosyntactic Correspondence in Bantu Reduplication
c. uku-zi-thum-a (H) ‘to send them’
uku-[zi-dl-a]þzi-dl-is-a uku-[zi-dl-i]þzi-dl-is-a uku-zi-[thum-a]þthum-a *uku-[zi-thu]þzi-thum-a
289
(OM is reduplicated) (OM not reduplicated)
As seen in the unacceptable form *uku-zi-[yi-dl-a]þyi-dl-a (intended: ‘to eat them’), the only option for fleshing out the reduplicant that is not available in these verbs with OMs is prefixal yi-. Prefixal yi- is in complementary distribution with OMs. We thus add the OM, if present in the verb, to the list of options for fleshing out the bisyllabic D-stemRED constituent: (33) Potential daughters of D-stemRED a. Root b. Extension su‰xes c. FV -a d. Su‰xal -yi e. Prefixal yiEntails backcopying; in complementary distribution with OM f. OMs Entails backcopying; in complementary distribution with prefixal yiNote in (32c) that the OM cannot appear in reduplicants whose root is CVC or longer (cf. the correct output, uku-zi-thuma-thuma, where the OM zi- is not included in the bisyllabic reduplicant, thum-a). This follows from our existing analysis. Since reduplicants are required by the high-ranking Max(Root) to exhaust the first CVC portion of the root, in such a case there is no room for any a‰xes in the reduplicant: (34)
zi-thum-a
[ss]
Max(Root)
+ a. zi-thum-aþthum-a b. zi-thuþzi-thum-a
*!(m)
At this point in the analysis, three questions arise: Why are OMs in complementary distribution with prefixal yi-? Why are only OMs and prefixal yi- ‘‘backcopied’’ to the base? What is the morphological analysis of backcopying? Our analysis is the following. First, we propose that OMs and prefixal yi- occupy the same morphological ‘‘slot.’’ They are prefixes that attach to D-stems to create a constituent that Bantuists refer to as the Macrostem (M-stem): (35)
M-stem Prefix
D-stem
where Prefix ¼ OM or yi-
290
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
Prefixal yi- may be semantically empty, but has a morphological e¤ect nonetheless: constituents containing it must be of the level M-stem.6 Given the structure in (35), our next step is to generalize the reduplication construction. In addition to consisting of juxtaposed D-stems, we now allow the R-stem to consist of juxtaposed M-stems. The juxtaposed stems in the reduplication construction must still agree in every feature—including morphological category (Dstem versus M-stem). (36)
R-stem Stem1
Stem2
where Stem1 and Stem2 can be either D-stems or M-stems, M-stems so long as they agree in every morphological feature (including category)
We can now understand why it is the prefixal supplements to the reduplicant that induce backcopying: they induce it by requiring the reduplicant to be an M-stem. Due to the highly agreeing nature of the reduplication construction, if the reduplicant is an M-stem, the base must be as well—and when the base is an M-stem, it of course must have a prefix. Once it is established that both reduplicant and base are M-stems, it is a simple matter of morphosyntactic feature checking to ensure that they each have the same prefix. Thus, OM doubling is not in fact backcopying, but rather simple reduplication. The components involved in the apparent backcopying of an OM are recapitulated as follows. First, the morphology allows a free choice of M-stem or D-stem reduplication constructions. Second, if the M-stem construction is chosen, the obligatory prefix (an OM or yi-) is parsed into the reduplicant. And finally, morphosyntactic feature agreement between reduplicant and base ensures that reduplicant and base end up with the same prefix. The M-stemRED cophonology is the same as the DstemRED cophonology, with one exception: with the M-stemRED cophonology Max(Pfx) and Max(Root) are inviolable. The inviolable nature of both constraints means that if the root is longer than C, the M-stem reduplication construction cannot be used. This is a correct prediction.7 In summary, we have examined further data that confirm our claim that the reduplicant is morphologically complex. We have shown that the truncation-agreement analysis of reduplication can extend nicely to these new data, including to the apparent ‘‘backcopying’’ phenomena that results when the reduplicant is prefixed. By contrast, the canonical stem and corresponding word analyses of Downing and Steriade have little to say about these data. Certainly reduplicants with initial yi(e.g., yi-dl-a) are not ‘‘canonical stems,’’ nor are the truncated reduplicants with initial zi- (e.g., zi-dl-i, zi-dl-a) possible words. We next consider imbricated verb stems, which provide further support for our analysis.
Morphosyntactic Correspondence in Bantu Reduplication
13.4
291
Imbrication
In this section we consider the reduplication of verb stems that have undergone imbrication, a special fusion process triggered by the perfective su‰x -ile. Whereas most CVC roots add -ile in the perfective, as seen to the right in (37a), (37) a. lim-a thum-a (H) b. bonakal-a (H/L) dumal-a (H/L) thath-a (H)
‘cultivate’ ‘send’ ‘appear, be visible’ ‘become depressed’ ‘take’
ba-lim-ile ba-thum-ile ba-bonakel-e ba-dumel-e ba-theth-e
‘they cultivated’ ‘they sent’ ‘they appeared, were visible’ ‘they became depressed’ ‘they took’
others such as those in (37b) instead fuse or ‘‘imbricate’’ the -il- part of the perfective ending -ile. In these examples, the penultimate root vowel [a] is realized as [e], and the stem also ends in the final vowel -e.8 The question is how the ‘‘umlaut’’ in (37b) a¤ects reduplication. As seen in (38a), (38) a. ba-bonakel-e (H/L) b. ba-dumel-e (H/L) c. ba-theth-e (H)
‘they appeared’ ‘they became depressed’ ‘they took’
ba-bonaþbonakel-e ba-duma-dumele ba-dume-dumele ba-thath-aþtheth-e ba-theth-aþtheth-e *ba-theth-eþtheth-e
since the reduplicant is bisyllabic, imbrication has no e¤ect on the reduplicant when the umlauted penultimate vowel does not occur in one of the first two syllables of the stem. In (38b), however, where imbrication a¤ects the second vowel of the root dumal-, converting it to dumel-, we see that the reduplicant can be either duma- or dume-. Similarly, in (38c), where imbrication converts the root thath- to theth-, there are again two possible reduplicants: thatha- and thetha-. The reduplicant thethe-, on the other hand, is not grammatical. The correct observation is that the ‘‘umlaut’’ triggered by the perfective can be reduplicated, while the final -e of the perfective cannot be. Note in the case of ba-dumaþdumel-e in (38b), that the reduplicant duma- is not an existing minimal canonical stem. So it cannot be the case that the sequence -ele of -dumele is simply being truncated. Instead, to account for the perfective, we recognize two ‘‘alloconstructions.’’ The first, schematized in (39), shows the nonimbricating alternate:
292
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
(39)
As shown, the nonimbricating -ile su‰x is unambiguously in the IFS position, hence unavailable for reduplication. Now consider the analysis of perfectives marked by -e with umlaut.9 The question is where the umlauting feature (. . .) should be within the morphosyntactic structure, since it optionally appears in the reduplicants in (38b,c). Our view is that Ndebele speakers themselves are not sure where to assign the umlaut, and hence have the two competing analyses, schematized in (40). (40) a. Imbricating perfective: Umlaut analyzed to be part of IFS
b. Imbricating perfective: Umlaut analyzed to be part of D-stem (e.g., root)
In (40a), the umlaut is assigned to the IFS, identically to -ile in (39), from which it clearly derives historically. In (40b), however, the umlaut is seen as an internal mod-
Morphosyntactic Correspondence in Bantu Reduplication
293
ification on the morpheme to which it attaches—here, the root.10 If assigned the structure in (40a), the umlaut will not be realized in the reduplicant, which bars the realization of inflectional features. If assigned the structure in (40b), the umlaut will be obligatorily realized because of the high-ranking constraint Max(Root). In neither case will the final su‰x -e make it into the reduplicant, since it is unambiguously an IFS. Finally, for the sake of completeness, note in (41) that reduplicated verbs optionally are exempt from imbrication even when their nonreduplicated forms must imbricate: (41) a. ba-bonaþbonakal-ile (H/L) ‘they appeared’ b. ba-dumaþdumal-ile (H/L) c. ba-thath-aþthath-ile (H)
‘they became depressed’ ‘they took’
*ba-bonakal-ile *ba-dumal-ile *ba-thath-ile
(cf. (38a)) (cf. (38b)) (cf. (38c))
We hypothesize that these forms arise in the following way. First, we note that imbrication is not automatic in Ndebele. For example, of the 148 CaCaC- verb roots in the Comparative Bantu On-Line Dictionary (CBOLD) version of Pelling (1971), 39 obligatorily undergo imbrication and 24 do so optionally, while 85 may not imbricate. Roots (and extension morphemes such as the reciprocal -an-) will therefore have to carry lexical marking for imbrication. This lexical marking only optionally percolates up to the R-stem level. When it is present, the appropriate match with the IFS -e will be made. When it is absent, -ile must be used. To summarize, with this last assumption, as well as the two ‘‘alloconstructions’’ in (38), imbrication supports the analysis we have developed thus far. Reduplication is juxtaposition of morphologically identical stems (either M-stems or D-stems), with the first subject to a bisyllabic size condition. 13.5
Palatalization
In the last two sections we have established that a‰xal material in the D-stem may surface in the reduplicant only when the verb root has been exhaustively parsed. We now consider passivized forms, which further support this conclusion and make more precise our proposals about the internal structure of the reduplicant. In (42a) we see that the passive is derived by means of the derivational su‰x -w-, here occurring directly after the root: (42) a. bal-a bik-a (H) phek-a b. boph-a (H) bumb-a (H) thum-a (H)
‘read’ ‘announce’ ‘cook’ ‘tie’ ‘mold’ ‘send’
bal-w-a bik-w-a phek-w-a botsh-w-a bunj-w-a thuny-w-a
‘be ‘be ‘be ‘be ‘be ‘be
read’ announced’ cooked’ tied’ molded’ sent’
294
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
The examples in (42b) show that this -w- palatalizes an immediately preceding labial consonant. However, as the examples in (43) show, palatalization is not restricted to the immediately preceding consonant, but can target any preceding labial, as long as it is not the first consonant in the constituent:11 (43) gombolozel-a vumbulul-a (H) fumbath-a (H)
‘encircle, surround’ ‘uncover, unearth’ ‘clench (hand)’
gonjolozel-w-a vunjulul-w-a funjath-w-a
‘be encircled’ ‘be uncovered’ ‘be clenched’
Like the imbricating alternant of the perfective, we follow Sibanda 2004 in analyzing the passive as having two phonological components. Basing himself on Zoll 1996, Sibanda characterizes these as subsegments, one palatalizing (call it p) and the other labializing (call it o). The palatalizing component attaches, potentially over some distance, to a preceding labial consonant and changes its place of articulation to palatal. The labializing component attaches to the immediately preceding consonant and gives it a labial o¤glide. Now consider the possible reduplicated forms of the verbs in (44). (44) a. bal-a
‘read’
bal-w-a
‘be read’
bik-a (H)
‘announce’
bik-w-a
‘be announced’
phek-a
‘cook’
phek-w-a
‘be cooked’
‘tie’
botsh-w-a
‘be tied’
bumb-a (H) ‘mold’
bunj-w-a
‘be molded’
thum-a (H)
thuny-w-a
‘be sent’
b. boph-a (H)
‘send’
bal-w-aþbal-w-a bal-aþbal-w-a bik-w-aþbik-w-a bik-aþbik-w-a phek-w-aþphek-w-a pek-aþphek-w-a botsh-w-aþbotsh-w-a botsh-aþbotsh-w-a *boph-a-botsh-w-a bunj-w-aþbunj-w-a bunj-aþbunj-w-a *bumb-aþbunj-w-a thuny-w-aþthuny-w-a thuny-aþthuny-w-a *thum-aþthuny-w-a
In each case the -w- may or may not appear in the reduplicant, exactly as we expect of a derivational su‰x. However, any labial palatalization induced by the passive obligatorily surfaces in the reduplicant (cf. the ungrammaticality of *thum-aþthunyw-a, etc.). The same facts are obtained when the final vowel is other than -a—for example, subjunctive -e in (45). (45) a. bal-w-e bik-w-e
‘be read’ (subjunctive) ‘be announced’
bal-w-aþbal-w-e bal-aþbal-w-e bik-w-aþbik-w-e bik-aþbik-w-e
Morphosyntactic Correspondence in Bantu Reduplication
phek-w-e b. botsh-w-e
‘be cooked’
phek-w-aþphek-w-e pek-aþphek-w-e botsh-w-aþbotsh-w-e botsh-aþbotsh-w-e *boph-a-botsh-w-a bunj-w-aþbunj-w-e bunj-aþbunj-w-e *bumb-aþbunj-w-e thuny-w-aþthuny-w-e thuny-aþthuny-w-e *thum-aþthuny-w-e
‘be tied’
bunj-w-e
‘be molded’
thuny-w-e
‘be sent’
295
Treating the passivizing p and o as subsegments under the D-stem parallels our analysis of the umlauting subsegment . . . that results in imbrication, as discussed earlier. Two important di¤erences subdivide this set of subsegments, however. First, while o and . . . are constrained to appear on the immediately preceding segment (the preceding consonant and /a/, respectively), p is permitted to find suitable hosts anywhere within the D-stem. That is, o and . . ., but not p, are subject to morphological adjacency (alignment), whereby the output segment on which it surfaces must correspond to the rightmost C or V of the immediately preceding morpheme in the input. This condition draws support from the data in (46), in which the root is longer than CVCV and must therefore be truncated in the reduplicant. (46) a. gombolozel-a vumbulul-a fumbath-a b. gonjolozel-w-a vunjulul-w-a funjath-w-a
‘encircle, surround’ ‘uncover, unearth’ ‘clench (hand)’ ‘be encircled’ ‘be uncovered’ ‘be clenched’
gomboþgombolozel-a vumbuþvumbulul-a fumbaþfumbath-a gonjoþgonjolozel-w-a vunjuþvunjulul-w-a funjaþfunjath-w-a *funj-w-aþfunjath-w-a
Although these are passive stems, passive o is not able to surface in the reduplicant. This follows straightforwardly from adjacency: the final consonant of the root does not surface in the reduplicant, and therefore, by the terms of adjacency, the passive -w- cannot surface either. As a control, the data in (47) show that labialization that is not from the passive morpheme must be preserved under reduplication: (47) a. yejwayel-a yejwayelek-a yejwayez-a yenway-a
‘get accustomed to’ ‘become customary’ ‘accustom’ ‘scratch an irritation’
yejwaþyejwayel-a yejwaþyejwayelek-a yejwaþyejwayez-a yenwaþenway-a
*yejaþyejwayel-a *yejaþyejwayelek-a *yejaþyejwayez-a *yenaþyenway-a
296
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
b.
yenwab-a @wonwab-a c. qwaqwaz-a tshwatshwaz-a
‘be happy’ ‘be happy’ ‘click (stick on stick)’ ‘make hissing noise’
yenwaþyenwab-a wonwaþwonwab-a qwaqwaþqwaqwaz-a tshwatshwaþ tshwatshwaz-a
*yenaþyenwab-a *wonaþwonwab-a *qwaqaþqwaqwazw-a *tshwatshaþ tshwatshwaz-a
It is not optional to copy only the C portion of a tautomorphemic Cw structure.12 The second dimension of di¤erence among the three subsegments is why imbricating . . . and passivizing o are only optionally parsed in the reduplicant, while passivizing p must surface (assuming, of course, that a suitable host exists in the output). The approach taken to imbrication in section 13.4 readily provides an explanation. Optionality in the case of . . . was said to result from the availability of the two alloconstructions in (40): when . . . is taken to be part of the IFS, it cannot appear in the reduplicant. When it is taken, instead, to belong to the same constituent as the segment on which it is realized (i.e., the root), it must be parsed into the reduplicant. In the passive case, there is only one construction: when a root labial is palatalized, p belongs to the root and hence must appear in the reduplicant because of high-ranked Max(Root). To recapitulate what we have said about the passive thus far, p obligatorily appears in the reduplicant because it is parsed with the root on which its palatalizing e¤ect is realized. On the other hand, o was said to optionally occur in the reduplicants in (44) and (45) because it is a derivational a‰x that can be truncated as in the case of applicative -el- and causative -is-. However, close inspection of more morphologically complex forms shows that additional, and more interesting, conditions hold on the surfacing of the passive o subsegment. Consider in particular the passivized applicative stems in (48). The applicative extension su‰x -el- occurs between the root and the passive su‰x. Note the expected long-distance action of the palatalizing component p of the passive in the nonreduplicated stems in (48b): (48) a. bal-el-a bik-el-a (H) phek-el-a b. boph-el-a (H) bumb-el-a (H) thum-el-a (H)
‘read for/at’ ‘announce for/at’ ‘cook for/at’ ‘tie for/at’ ‘mold for/at’ ‘send for/at’
bal-el-w-a bik-el-w-a phek-el-w-a botsh-el-w-a bunj-el-w-a thuny-el-w-a
‘be read for/at’ ‘be announced for/at’ ‘be cooked for/at’ ‘be tied for/at’ ‘mold for/at’ ‘be sent for/at’
Up to now we have not seen a case in which two semantically contentful su‰xes jockey for a position in the reduplicant. Which will take precedence? Will there be free variation, as we have seen between extension su‰xes and the ‘‘filler’’ morphs [yi] and [a]? The forms in (49) show the reduplicated versions of the passivized applicatives in (48). As seen earlier in the unextended stems, the palatization component p of the passive surfaces in the reduplicant, while the labial component o is optional:
Morphosyntactic Correspondence in Bantu Reduplication
(49) a. bal-el-w-a
‘be read for/at’
bik-el-w-a
‘be announced for/at’
phek-el-w-a
‘be cooked for/at’
b. botsh-el-w-a
‘be tied for/at’
bunj-el-w-a
‘be molded for/at’
thuny-el-w-a
‘be sent for/at’
297
bal-eþbal-el-w-a bal-aþbal-el-w-a bal-w-aþbal-el-w-a bik-eþbik-el-w-a bik-aþbik-el-w-a bik-w-aþbik-el-w-a phek-eþphek-el-w-a phek-aþphek-el-w-a phek-w-aþphek-el-w-a botsh-eþbotsh-el-w-a botsh-aþbotsh-el-w-a botsh-w-aþbotsh-el-w-a bunj-eþbunj-el-w-a bunj-aþbunj-el-w-a bunj-w-aþbunj-el-w-a thuny-eþthuny-el-w-a thuny-aþthuny-el-w-a thuny-w-aþthuny-el-w-a
Note that when the passive o does surface in these reduplicants, it surfaces on a segment to which it is not adjacent in the input, in apparent violation of the Adjacency constraint proposed earlier. That is, it appears that a mapping such as that in (50) is occurring: the passive is surfacing not on the final segment of the immediately preceding morpheme, but on the final segment of the morpheme before that. The applicative su‰x is being skipped over, even though it is closer to the beginning of the stem and (under standard assumptions about multiply extended D-stems) forms a closer constituent with the root than the passive does: (50)
Is Adjacency violable in Ndebele? We suggest that it is not. Rather, we contend that the appearance of ‘‘at-a-distance’’ linking of passive o in (50) is illusory, and that the applicative is not in fact being skipped over in these cases.
298
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
To appreciate this contention, consider the data in (51), which serve as background to a previously undetected morphological condition on the surfacing of passive o in reduplicants. The verbs in both sentences in (51) contain the sequence -el-w-, composed of applicative and passive extension su‰xes: (51) a. abantwana b-a-phek-el-w-a ukudla children they-past-cook-appl-pass food ‘The children were cooked food’ b. ukudla kw-a-phek-el-w-a abantwana food it-past-cook-appl-pass children ‘The food was cooked (for) the children’ Although possessing the same surface morphs in the same linear order, these verb stems realize two di¤erent argument structures. The sentence in (51a) is a passivized applicative, and its verb has the argument structure in (52a). (52) a.
b.
c.
! -el-wThis morphological structure and this order of the su‰xes is what is expected from the scope relations or ‘‘mirror principle’’ (Baker 1985; Alsina 1999; Hyman and Mchombo 1992; Hyman 1994a). The sentence in (51b), on the other hand, is an applicativized passive, and its verb has the argument structure in (52b). In this case, by the mirror principle, the order of the morphs is expected to be passive -w- followed by applicative -el-. However, in Ndebele, as in most Bantu languages, the applicative cannot follow the passive.13 Hence, a number of linguists, particularly of the Tervuren school inspired by Meeussen (1967), have proposed rules that metathesize morphs, as when (52b) is transformed into (52c). It is thus not the surface order of the morphs that interests us, but rather the underlying morphosyntactic structure. Given this, now consider how the verbs in (51) reduplicate: (53) a. abantwana
b. ukudla
b-a-phek-eþphek-el-w-a b-a-phek-aþphek-el-w-a *b-a-phek-w-aþphek-el-w-a kw-a-phek-eþphek-el-w-a kw-a-phek-aþphek-el-w-a kw-a-phek-w-aþphek-el-w-a
ukudla
abantwana
‘The children were cooked food’ ‘Food was cooked for the children’
Morphosyntactic Correspondence in Bantu Reduplication
299
In (53a), the passivized applicative, we see that the reduplicant can be pheke- or pheka-, but not *phekwa. Passive o can not be parsed in the reduplicant. In (53b), by contrast, the reduplicant can be phek-eþ, phek-aþ, or phek-w-aþ. The reduplicant is free to parse o or not, as we are used to seeing. The reason for this surprising discrepancy jumps out once one compares the argument structures of the verb stems in (52): the passive surfaces under reduplication only if it is sister to the root. Thus, phek-w-aþ is well-formed as the reduplicant of -phek-el-w-a only if the subtree rootþpassive occurs in the base to which it corresponds, as in (52b). We conclude that reference to internal morphosyntactic structure is critical to determining the well-formedness of reduplicants. How exactly is this reference to be made? We posit that it follows as a direct consequence from the internal structure of the Ndebele verb stem. Making the standard assumption that D-stems are binary branching, (54) gives the structure of the reduplicant of an extended verb stem of the kind seen above: (54)
Whether it is D-stemREDj or D-stemREDi that is spelled out, only two reduplicants are possible: phek-e and phek-a. The candidate phek-w-a is ruled out because of the adjacency violation that results from skipping over the applicative -el-. On the other hand, in the case of applicativized passives, we would have a structure like the one in (55), in which the lowest D-stem node has the passive su‰x adjacent to the root: (55)
As seen, both phek-w-a and phek-a can be spelled out as the reduplicant at both the D-stemREDj and D-stemREDi levels, while phek-e can result only as the realization of D-stemREDj . Note that phek-w-e is not a possible reduplicant, since the morphotactics prohibit the [e] of applicative -el- from following passive -w- on the surface.
300
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
As seen, in the D-stemRED cophonology in both (54) and (55), copying of labialization is optional, but adjacency is always respected. Thus the passive -w- can, but need not, be parsed at the lowest D-stem RED node, where the segment to which it would be adjacent in output ends the immediately preceding morpheme in the input. Observe that the branching structure we hypothesize for the Bantu verb stem, both ‘‘base’’ and ‘‘reduplicant,’’ is tantamount to a cyclic analysis. Constraints are enforced on every incarnation of a D-stem (or M-stem), regardless of whether it is a daughter or a mother in the verb structure. This e¤ect is predicted by the theory of Sign-Based Morphology (Orgun 1996) and is particularly evident in cases where one valency demands one versus another order of su‰xation. In the example in (56a), for instance, the verb fik-a ‘come’ is intransitive: (56) a. fik-a b. fik-is-a c. fik-is-w-a
‘come’ ‘make come’ ‘be made to come’
In (57b), fik-is-a ‘make come’ is made transitive by the causative su‰x -is-, and, as a result, can be passivized as in (57c). Now note the possible reduplications of (56c) in (57a). (57) a.
fik-iþfik-is-w-a fik-aþfik-is-w-a b. *fik-w-aþfik-is-w-a
As seen, the reduplicant consists of the root fik- plus either the [i] of the causative su‰x -is- or the default FV -a. In (57b) we see that it is not possible to parse the passive -w- in the reduplicant because this would imply that fik- ‘come’ had become passivized in the first cycle—which is, of course, not the case, since the verb is intransitive. This e¤ect is automatically obtained from the approach taken here: rather than predicting the properties of the reduplicant by reference to the surface output of the corresponding base, the full range of facts is accounted for by attributing to the D-stemRED the full morphosyntactic structure of the base. This morphosyntactic structure is in turn spelled out in the normal way, subject to the cophonology of the reduplicant—for example, the two-syllable constraint. Most interesting is the role played by this constraint in producing reduplications such as phek-w-aþphek-el-w-a in (55). While a superficial comparison of base and reduplicant suggests a violation of Adjacency, it is the underlying structure in (55) combined with [s s] that make the reduplicant phek-w-aþ possible. Because of the prosodic constraint, both -wand -el- cannot be parsed. In addition, *phek-w-eþ is not a possible reduplicant because the passive morph -w- is not allowed to precede the applicative morph -el- (or its subpart, as -e would be in this case). The reduplicant phek-w-aþ is possible because it respects the hierarchical morphosyntactic structure in (55), and, crucially,
Morphosyntactic Correspondence in Bantu Reduplication
301
because there is no requirement that the feature applicative be spelled out in the reduplicant. If that option is taken, however, then the only possible grammatical output is phek-eþphek-el-w-a, which more closely resembles the surface order of the -el-w- sequence. Examples of this sort can be produced with respect to other su‰xes and other morphosyntactic situations. Ndebele thus provides strong evidence that the potential spell-outs of a reduplicant are governed by more than superficial resemblances with its base—and, once again, that as concerns the Bantu verb stem, morphology must proceed in a cyclic fashion (Hyman 1994b). 13.6
Conclusion
At the outset of this study, we characterized the study of partial reduplication as having as its goal to construct a theory that insightfully captures the full range of considerations that speakers ‘‘care about’’ in determining how a reduplicant will relate to its base. What has the study of Bantu reduplication contributed toward this goal? Certainly one very important contribution has been made in the work of Downing (1997a, etc.), in recognizing the important role that morphological structure plays in Bantu reduplication, which, therefore, is not solely a prosodic phenomenon. Mutaka and Hyman (1990) showed for Kinande the role that morphological considerations can play, because morpheme integrity can inhibit prosodic templatic e¤ects in certain languages. Downing went further and showed the role that morphological constituent structure plays, noting e¤ects of the derivational stem constituent on Bantu reduplication. In this chapter we have taken this important insight a step further, finding that not only the D-stem as such but also its internal constituents—root, extension su‰xes, and filler su‰xes—play a role in defining the variety of shapes of the Ndebele reduplicant. In making reference to D-stems and their constituents, the ‘‘canonical stem’’ shape to which Downing observed most reduplicants to adhere can actually be derived, rather than stipulated. This is an especially desirable outcome in Ndebele, where not all reduplicants are canonical stems in any case. In fact, the comparative study of verb reduplication in Bantu shows that the phonological and morphological constraints on the reduplicant can be independently controlled to produce most of the possible combinations. In our introduction we cited Eulenberg 1971 as expressing the view that partial reduplication derives from total reduplication—that is, by phonological and morphological paring down of the full base. Diachronic evidence for this position is quite clear as concerns the Bantu verb stem. Numerous Bantu languages require full verb-stem reduplication including all extensions and the IFS—for example, Luganda lim-il-agan-aþlim-il-agan-a ‘cultivate for each other here and there’, and its subjunctive, lim-il-agan-eþlim-il-agan-e.
302
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
It is not di‰cult to explain why this original state of total reduplication should become modified over time: it is hardly necessary to repeat the whole verbal I-stem in order to convey the aspectual idea that the action is done a little here and there. It is therefore possible to economize and limit the reduplicant to a subpart of the base. This can be done either phonologically and/or morphologically, as we have seen. Historical developments can follow one or both of the following scales: (58) a. Phonological scale full > foot > syllable (> mora > gemination) b. Morphological scale full > derivational stem > root The phonological scale consists of a gradual narrowing down of the reduplicant to fit a prosodic template, frequently a bisyllabic foot. The morphological scale consists of first restricting the reduplicant to derivational material, and ultimately to root material. As is known from Kikerewe (Odden 1996; Downing 1997b), there can be ‘‘intermediate’’ stages. In this language, a verb that has two or more extensions can be reduplicated fully, or can copy zero, one or more of these extensions—for instance, lim-il-anþa ‘cultivate for each other’ can reduplicate as lim-il-an-aþlim-il-an-a, lim-ilaþlim-il-an-a, or lim-aþlim-il-an-a. Restricting ourselves to verb-stem reduplication in Bantu, the specific choices speakers make can be listed as parameter settings, as in (59). (59) a. Phonology i. Size constraint: yes/no (1, 2 s’s) ii. Tone transfer: yes/no14 iii. Length transfer: yes/no b. Morphology i. IFS: yes/no ii. OM: yes/no iii. FV ¼ -a: yes/no iv. Max(Root): yes/no Concerning the morphological parameters, we hypothesize that copying the OM is not original. Rather, verb reduplication was originally limited to the I-stem. Reduplication of the OM is thus a subsequent development motivated by two separate phonological conditions. The first, seen in Ndebele, occurs when the verb stem is monosyllabic and the OM is copied in order to fill out the bisyllabic foot condition on the reduplicant.15 This is one of three strategies that have been noted to fill out a [s s] condition on the reduplicant: (i) ‘‘move up’’ to M-stem and copy the OM, if there is one (e.g., Ndebele class 10 zi-); (ii) fill with a ‘‘dummy’’ syllable (e.g., Ndebele yi-); (iii) double reduplicate the monosyllabic stem (e.g., sw-a-sw-aþsw-a ‘grind here and there’ in Kinande) (Mutaka and Hyman 1990).
Morphosyntactic Correspondence in Bantu Reduplication
303
Turning now to the tendency for the reduplicant to end in -a, we note, as does Downing (1997c, etc.), that this is the most frequent verb ending in most Bantu languages. However, rather than seeing this as an indication of a special ‘‘canonical verb stem,’’ we believe it is necessary to have a Bantu-specific constraint FV[-a], whose e¤ects are, in fact, not restricted to verb reduplication—or even verbs—in Bantu.16 There is in fact considerable evidence that -a is the default stem-final vowel, whether its function is inflectional or derivational, and whether applying to nouns or verbs. First, it is true, as Downing points out, that -a is used as an inflectional default FV in many Bantu languages. In Giphende, for example, the recent past tense is marked by the IFS -ı´ if immediately preceded by a CV-, CVC- or CGVC- verb root, as in (60a,b,c) respectively:17 (60) a. tw-a-mb-ı´ tw-a-y-ı´ b. tw-a-som-ı´ tw-a-meng-ı´ c. tw-a-kwec-ı´ tw-a-mwang-ı´ d. tw-a-vumbı´g-a´ tw-a-digı´m-a´ e. tw-a-som-e´l-a´ tw-a-som-e´s-a´ f. tw-a-mb-e´l-a´ tw-a-mb-e´s-a´
‘we ‘we ‘we ‘we ‘we ‘we ‘we ‘we ‘we ‘we ‘we ‘we
put’ went’ loaded (a gun)’ detested’ tied’ scattered’ buried’ were afraid’ loaded for/at’ made load’ put for/at’ made put’
In (60d) we see that when the stem consists of three full syllables, the IFS must be -a´. In these examples the root is morphologically unanalyzable. In (60e), where a -CVCroot is followed by a -VC- extension, the same -a´ is required. Finally, in (60f ), where the first stem syllable consists of a CV- root which has fused with the V of a -VCextension, we also must get the default FV -a´ rather than -ı´. The generalization is, therefore, that -ı´ can occur only within the second syllable of a verb stem whose base is morphologically simplex. Otherwise the recent past tense reverts to default -a´. It is clear from such evidence that the use of the FV -a is logically independent of reduplication. We also suggest that it is independent of the bisyllabic CVC-a shape that Downing has referred to as a the ‘‘minimal canonical stem.’’ To see this, consider the following forms from Lengola (Stappers 1971, 268), where verb-stem reduplication marks the habitual: (61) a. i-kul-a i-£on-a i-tu´m-a i-lı´mb-a
‘acheter’ ‘regarder’ ‘envoyer’ ‘chanter’
i-kul-aþkul-a i-£on-aþ£on-a i-tu´m-aþtum-a i-lı´mb-aþlimb-a
304
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
b. i-kpet-a i-ki-a i-gbok-a c. i-£ı´-a i-Bıˆ-a i-„a´m-a
‘couper’ ‘faire’ ‘trouver’ ‘manger’ ‘parler’ ‘crier’
i-kp-aþkpet-a i-k-aþki-a i-gb-aþgbok-a i-£-aˆþ£i-a i-B-aˆþBi-a i-„-aˆþ„am-a
In (61a) we observe full rootþFV reduplication (preceded by the infinitive prefix i-). However, Stappers also indicates that verbs such as those in (61b,c) form their habitual by means of a [Ca] reduplicant. The most straightforward analysis is that in these forms D-stemRED is limited to one syllable (versus the more usual two) and must end in -a. A similar conclusion is drawn from gerundive reduplication in Lomongo (Hulstaert 1965):
‘lancer’ ‘duper’ ‘eˆtre vivant’ ‘couvrir’ ‘insulter’ ‘piler’
c
-lı´k-le´ng-kng-ku´k-to´l-t ´k-
Gerundive Bombwanja dialect Coquilhatville and Northern dialects n-dı´þlı´k-a n-d-a´þlı´k-a n-de´þle´ng-a n-d-a´þle´ng-a n-kþkng-a n-k-a´þkng-a n-ku´þku´k-a n-k-a´þku´k-a n-to´þto´l-a n-t-a´þto´l-a n-t ´þt ´k-a n-t-a´þt ´k-a c
Verb
c
(62)
c
In this case of verb reduplication, the reduplicant again consists of a single CV syllable. In Bombwanja dialect, the CV corresponds to the initial CV of the base. In Coquilhatville and Northern dialects of Lomongo, however, it has the shape [Ca]. As in Lengola, the fixed vocalism is presumably a reflection of the FV -a, hence the reduplicant can be analyzed as C-aþ. If we were to extend Downing’s analysis of the Kinande ‘‘minimal canonical D-stem’’ CVC-a, we would have to say that in Lengola and Lomongo the minimal canonical D-stem is C-aþ. However, we know of no evidence for this. Rather, in this construction we would simply say that RED ¼ s and the Bantu-specific constraint FV[-a] is ranked high. Finally, note that even in languages where there is a bisyllabic minimum in e¤ect, CVC-a can be a possible reduplicant without there having to be a corresponding CVC- root. The example comes from Kikuyu (Peng 1991; Downing 1999): (63) a. kor-a cin-a b. koor-a Buut-a
‘grow’ ‘burn’ ‘pull out’ ‘depose’
! ! ! !
kor-aþkor-a cin-aþcin-a koor-aþkoor-a Buut-aþBuut-a
Morphosyntactic Correspondence in Bantu Reduplication
! ! ! !
Boc-aþBocor-a h r-aþh rr-a ci'-aþci'erer-a hwr-aþhwrrk-a cc
‘be indented’ ‘be quiet’ ‘encircle’ ‘tilt’
cc
c. Bocor-a h rr-a d. ci'erer-a hwrrk-a
305
cc
When the base is CVC-a or CVVC-a, as in (63a,b), reduplication appears to be total.18 The longer stems in (63c,d) indicate, however, that only the first CV(V)Cmay appear in the reduplicant, and that the FV -a is required. To account for this, Peng refers to the initial CV(V)C- of verbs as a prosodic minimal root (Rmin ) that has the shape [s.C]. In our framework, the FV[-a] simply outranks Max(Root). What these data from Lengola, Lomongo, and Kikuyu clearly show is that each of the properties of reduplicants can be independently manipulated and reranked in terms of the appropriate constraints. Thus, although Max(Root) is very important crosslinguistically, it can be outranked by RED[ss] in Ndebele, by RED[s] in Lengola and Lomongo, or by FV[-a] in Kikuyu. In fact, Max(Root) appears to be even more seriously violated in at least some realizations in Bukusu in (66) below. To appreciate this, however, we need to seek explanations for two crucial—and, as we will suggest—related questions: Why should there be a tendency to realize as much of the root as possible in Bantu verb-stem reduplication? Why is Bantu verb-stem reduplication prefixal? The most immediate answer that linguists will probably advance to the first question is that roots—an open morphological class—are universally more contrastive and salient than grammatical a‰xes. However, a moment’s reflection will reveal that this answer is insu‰cient. We suggest a second contributor to relative high ranking of Max(Root), namely that a Bantu verb-stem reduplication is prefixal. As originally observed by Marantz (1982), prefixal reduplication tends to be left-anchored, while su‰xal reduplication tends to be right-anchored. Schematically, the unmarked mappings (in Marantz’s framework) or Base-Reduplicant correspondences (McCarthy and Prince 1995) are as in (64a), their marked counterparts as in (64b). (64) a. Unmarked Prefixal: Su‰xal: b. Marked Prefixal: Su‰xal:
tuki tuki
! !
tuþtuki tukiþki
tuki tuki
! !
kiþtuki tukiþtu
Niepokuj (1991) speculates that the more widespread unmarked situation in (64a) has a functional explanation: in these outputs the copied sequence in the reduplicant appears adjacent to the same sequence in the base. As a result, the reduplicative process is more transparent—that is, not requiring a long-distance identification as in
306
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
(64b). It is therefore to be expected that there will be greater faithfulness to the root in prefixal reduplication than there will be to su‰xes. To show that this is the correct interpretation, consider what Bantu verb-stem reduplication would look like if it were su‰xal. Let us take the same forms we considered from Ndebele in (3) and assume that the su‰xed reduplicant is subject to Red(ss). Possible outputs might be those indicated in (65). (65) a. lim-el-a lim-is-a b. lim-el-a lim-is-a
lim-el-aþlim-e lim-el-aþlim-a lim-is-aþlim-i lim-is-aþlim-a lim-el-aþm-el-a lim-is-aþm-is-a
‘cultivate for/at’
(applicative -el-)
‘make cultivate’
(causative -is-)
‘cultivate for/at’ ‘make cultivate’
In the mirror-image realizations of Ndebele in (65a), we provide two variants for each reduplication: the first copies the first two syllables of the base, while the second uses the FV -a. Such realizations, which represent the marked base-reduplicant correspondence in (64b), are unattested in Bantu. The unmarked BR correspondence with right-anchoring is given in (65b), where su‰xes are preserved at the expense of (part of ) the root -lim-. The realizations in (65b) violate Max(Root) in a way that (65a) does not, but can we say that they are any less ‘‘natural’’ or expected? While we have not (yet?) found exact outputs as in (65b) in any Bantu language, we have noted some variants in Bukusu that are highly suggestive.19 Thus, consider the reduplications in (66). (66) a. lim-a re´m-a b. lim-il-a re´m-er-a c. kacul-a mulix-a
‘cultivate’ ‘cut’ ‘cultivate for/at’ ‘cut for/at’ ‘chat, talk’ ‘flash’
! ! ! ! ! !
lim-aþlim-a re´m-aþrem-a lim-aþlim-il-a re´m-aþrem-er-a kacul-aþcul-a mulix-aþlix-a
In (66b), where the applicative su‰x has been added, reduplication is prefixal. Consider, however, the forms in (66c), where the stem contains an unanalyzable CVCVC- verb base. While some such verbs reduplicate fully, these reduplicate by means of a truncated ‘‘second part’’—that is, by apparent su‰xation.20 The generalization is that the speaker chooses to isolate the root at the left and the su‰xes at the right of the total form. Why should this be? We suggest that the answer is the same as to our second question: Why should Bantu verb-stem reduplication be prefixal? The Bantu verb stem is su‰xal, as we have seen. Speakers are therefore accustomed to both derivational and inflectional grammatical marking at its right edge. If reduplication were to proceed as in (65a), this general organizing principle would be violated. The resolution is thus as in (66b):
Morphosyntactic Correspondence in Bantu Reduplication
307
the root appears at the left of the structure, where it is expected, and the su‰xes appear at the right, where they too are expected. The resolution in (66c) is thus ingenious in that the second part of the R-stem includes all of the su‰xal material—and only a minimum of root material. The fact that any of the root occurs in the second part at all is presumably due to the otherwise exceptionless property that a‰xes cannot occur without a root. Since this is but one of the reduplicative strategies employed by our consultant, it would be much more interesting if some Bantu language had only this pattern. To summarize this last result, we suggest that Max(Root) is driven by the su‰xal nature of the Bantu verb stem. We therefore make the following general markedness claims concerning the a‰x orientation of the reduplicant crosslinguistically: (i) the reduplicant will tend to be prefixed when the base has a su‰xing structure, and (ii) the reduplicant will tend to be su‰xed when the base has a prefixing structure. We have already seen the evidence for prefixal reduplication in Bantu verb-stem reduplication. On the other hand, Mutaka and Hyman (1990) demonstrate that noun reduplication is a word-level su‰xal process in Kinande: (67) a. ku-gulu mu-go´ngo` ku´-boko b. o.mu´-twe bi-la ka´-tı`
‘leg’ ‘back’ ‘arm’ ‘head’ ‘intestines’ ‘stick’
ku-gulu.gulu mu-go´ngo.go´ngo` ku´-boko´.boko mu´-twe´.mu´-twe bi-la.bi-la ka´-tı´.ka´-tı`
‘a real leg’ (etc.)
The direction of a‰xation is not evident in (67a), where we might think that it is the stem that has been totally reduplicated. However, the nouns in (67b) provide the answer. Here, the noun class prefix is also reduplicated to fill out the bisyllabic prefixal template. Mutaka and Hyman’s analysis is that the prefixþstem is reduplicated as a su‰x, and the noun prefix is copied in (67b) because the stem is monosyllabic. In our view, su‰xal reduplication is possible in noun reduplication because nouns are not characterized by extensive su‰xation. Instead, they are most notably marked by noun class prefixes. As such, they are good candidates for su‰xal reduplication, as per (67b). In this study we have had several goals. A first goal has been to document in some detail the nature of verb-stem reduplication in Ndebele. A second goal has been to contribute to discovering the full class of reduplicative parameters that can be set by means of di¤erent constraint rankings. In so doing, a third goal has thus been to contribute to Optimality Theory in providing yet another example of language-internal variation that can be attributed to free constraint ranking. Finally, we have illustrated the need for enforcement of constraints on subtrees (cyclicity) and have provided a morphological reanalysis of apparent backcopying in reduplication.
308
Larry M. Hyman, Sharon Inkelas, and Galen Sibanda
Notes This is a shortened (and only slightly updated) version of a paper prepared on April 25, 1998, for the Trilateral Phonology Weekend (TREND) meeting in Berkeley. We are grateful for several helpful discussions, especially with Laura Downing. Research on the Ndebele lexicon, supported by the Comparative Bantu On-Line Dictonary (CBOLD) project, was funded in part by National Science Foundation grant SBR96-16330. 1. Verb tone is indicated throughout by (H) following verb stems with High-tone roots and (H/L) following verbs in which the root can take both High and Low tone. Unmarked forms have Low-tone roots. High tone on consonantal roots is realized only in the presence of extension su‰xes. 2. Technically, -ile can be demonstrated to consist of two su‰xes, -il-e, since the passive -wappears between them—for example, si-thu´ny-i-w-e ‘it was sent’ (cf. corresponding active stem: /thum-il-e/). 3. Completeness would require us to consider, in addition, reduplicants lacking the FV -a. This would force us to use the additional notational complication of tableaux des tableaux (Itoˆ, Mester, and Padgett 1995) for dealing with allomorphy. We leave this to the reader. 4. We assume that -yi is disallowed in regular D-stems because it serves no function and thus gratuitously violates *Struc (the general ban on structure of all kinds). Below we relate the presence of prefixal yi- in imperatives such as yi-dla ‘eat!’ to the macrostem level. 5. We are not considering the possibility that the [e] of [dl-e] is interpreted as the subjunctive final vowel -e, since the base in this case is not subjunctive, and since the IFS -e never reduplicates under any condition. 6. In proposing that OMs belong to a Macrostem constituent that excludes the IFS, we depart from the tradition of analyzing them as sisters to the I-stem. In a number of Bantu languages that otherwise restrict the prefix slot to one OM, the first-person singular prefix n- can still cooccur with a second OM. A reasonable conclusion to draw is that it joins the stem, thereby freeing up the one prefix slot for a second OM (Schlindwein 1986). 7. Another option for generating backcopying would be to assume a phonological correspondence between reduplicant and base, following McCarthy and Prince 1993b, 1995, and use an anchoring constraint. This would capture the generalization that the a‰xes that force backcopying are both prefixes: Anchor-L: The initial elements of RED and Base must correspond. 8. The imbricated forms derive historically from *bonakail-e, *dumail-e, and *thaith-e. See Bastin 1983 and Hyman 1995 for further evidence and discussion. 9. It should be pointed out that some verbs do not accept the imbricated form, while still others show variability. In addition, verbs whose last vowel is other than /a/ may also take the ‘‘short’’ -e form in the perfective, but in these cases there is no umlaut, e.g. a-dabul-ile ¼ a-dabul-e ‘he tore’. 10. An alternative would be to analyze the umlaut as an ‘‘extension,’’ where it would be optionally parsed under reduplication, exactly like applicative -el- and causative -is-. While this works for the perfective, it does not generalize to the passive construction analyzed in section 13.5. We therefore prefer to say that speakers are inconsistent in assigning the umlauting feature to either the IFS or to the D-stem, as in (40a,b). The one exception is where the umlaut is assigned to an extension with /a/ (e.g., thum-an-a ‘send each other’ ! ba-thum-en-e ‘they sent
Morphosyntactic Correspondence in Bantu Reduplication
309
each other’). In this case the umlaut is assigned to -an-—that is, under the extension node in (40b). 11. The imperviousness of morpheme-initial labials to palatalization has been addressed by Chen and Malambe (1998), Downing (1998), and Sibanda (1999a); we will pay it no further attention here. 12. The forms in (47) are the only examples we know where tautomorphemic Cw is found in post-stem-initial position. Since the glides of initial [ye] and [wo] in (47a,b) can be treated as epenthetic, the correct generalization may be that tautomorphemic /w/ appears only after the first consonant of a stem. See Sibanda 2004. 13. For a comprehensive study of verb extension combinations in Ndebele, see Sibanda 2004, where it is shown that the ‘‘morphotactically unmarked’’ linear order is -is-el-an-w- (causativeapplicative-reciprocal-passive); also Hyman 2003 for an overview of templatic su‰xing ordering in Bantu. 14. The only Central Bantu language we know with tone transfer in verb-stem reduplication is Chichewa, where each stem is a phonological word (Mtenje 1988; Kanerva 1989; Hyman and Mtenje 1999; Myers and Carleton 1996). 15. A second situation in which the OM sometimes becomes reduplicated is when it fuses with the root—that is, when the OM is either a homorganic nasal, or a CV- prefix followed by Vinitial root, as in Kihehe (Odden and Odden 1985). 16. In Basaa, for instance, nouns are diminutivized by CV prefixal reduplication and a shift to class 19/13 (prefixes hi-/di-). As seen in the following examples provided by Deborah Schmidt, the full form typically ends in -a´: mim ‘body’ ! hi-miþmı´m-a´, £l ‘thigh’ ! hi-£þ£l-a´, l Ð ‘country’ ! hi-l þl ´Ð-a´. c
c
c
17. Study of Giphende was conducted with Mwatha Ngalasso. The final -i of the recent past optionally assimilates to the preceding vowel. Thus, the two forms in (60b) may also be realized tw-a-som-o´ and tw-a-meng-e´. 18. Peng cites forms without tones and without FV. We have added -a to all verb bases to make them more comparable to those seen in Ndebele and elsewhere in this study. 19. These forms were collected by Julia Hill in an undergraduate field methods course at UC Berkeley in fall 1997. The consultant was Wanjala Khisa, who showed a great deal of variation (and with whom Downing 2004, pursued a more detailed study). 20. We use the designations ‘‘first part’’ and ‘‘second part’’ instead of reduplicant and base, since it is not evident which is which in these forms. This problem, anticipated in Niepokuj 1991, is further explored in Inkelas and Zoll 2005.
14
Patterns of Reduplication in Yoruba
Douglas Pulleyblank
The task is thus to harmonize the theory of reduplication with the rest of non-linear morphology as far as possible, in a way which also solves the descriptive problems, while maintaining, and if possible increasing, its desirable restrictiveness. —Kiparsky 1986b, 8 14.1
Introduction
Cases of reduplication can be located on a scale where one logical extreme is to ‘‘compound’’ two completely accurate copies of a base. Moving down the scale, one observes cases involving deviations from such accuracy—cases where only a subset of the base is copied, where the copy is not fully accurate, or where reduplication mixes copied and specified material. The logical extreme as one moves down this scale is an ‘‘ordinary’’ a‰x whose segmental and prosodic makeup is fully independent of the base to which it attaches. An adequate theory of morphology and phonology must be able to account for the entire range of such cases. This chapter is a case study of four types of reduplication in Yoruba, a NigerCongo language of Nigeria, in which reduplication is diverse and productive (see also Abı´o´du´n 1997; Adewole 1997).1 Each of the four reduplication types exhibits ˙ rather di¤erent properties. At one end of the scale, agentive reduplication involves two identical copies of a base, with both segmental and tonal material copied. Partway down the scale, infixing reduplication involves a complete copy of the base supplemented by a segmentally and tonally fixed infix. Further down the scale, distributive reduplication copies part of a base, restricting its copying to segmental material. At the end of the scale, gerundive reduplication copies only part of a base, mixing the copied material with lexically specified material. Much recent work has stipulated that base-reduplicant identity results from special constraints on correspondence relations for reduplication (e.g., McCarthy and Prince 1995). This chapter argues that, at least in Yoruba, such special constraints are unnecessary and redundant. Complete or partial identity between a base and a reduplicant results from either (i) morphological copying or (ii) the satisfaction of
312
Douglas Pulleyblank
prosodic requirements. While special correspondence relations do occur in reduplicative constructions, these relations are shown to be by-products of the phonology seeking to minimize violations of input-output faithfulness, not the result of constraints specifically tailored for reduplication. 14.2
Complete Reduplication: The Agentive
At the ‘‘complete accurate copy’’ end of the reduplicative scale, agentives and instrumentals are productively derived by faithfully reduplicating a verb þ noun combination as in (1) (Folarin 1987; Ola 1995).2 Segmentally, both consonants and vowels ˙ ˙ are copied intact, including vowel height, nasality, and tone. (1) Agentive reduplication a. wole´wole´ ‘sanitary inspector’ b. pejapeja ‘fisherman’ ˙ ˙ c. yoyı´nyoyı´n ‘dentist’ ˙ ˙ d. jedı´jedı´ ‘piles’ ˙ ˙ e. yı´nru`nyı´nru`n ‘meningitis’ ´ ´ ´ ´ f. nawonawo ‘extravagant person’ g. jaye´jaye´ ‘lover of pleasure’
wole´ peja ˙ yoyı´n ˙ jedı´ ˙ yı´nru`n na´wo´ jaye´
‘look at the house’ ‘kill fish’ ‘extract tooth’ ‘eat anus’ ‘twist neck’ ‘spend money’ ‘enjoy life’
In analyzing total reduplication, there are two basic approaches. One possibility is to analyze one half of the reduplicated form as the ‘‘base,’’ the other half as the ‘‘reduplicant,’’ and to require correspondence between the two (McCarthy and Prince 1995). If no constraint interferes with perfect correspondence, then total reduplication results. A second possibility is to attribute reduplication more directly to the morphology. Specifically, reduplication can be analyzed as a form of compounding with the morphological requirement that the two halves of the compound be identical (Yip 1998; Hyman et al., this volume; Inkelas and Zoll 2000). Although both approaches can derive the Yoruba agentives in (1), there are significant di¤erences between them. 14.2.1
Morphological Identity
Under the morphological identity approach, the morphological structure of the reduplicated form is one in which the two halves are identical: (2) Agentive reduplication [ Agentive [VPi ][VPi ]] This approach bears a clear relation to the proposal of Steriade (1988) whereby reduplication results from a complete copy that can be pared down by the application of phonological processes. The crucial phonological constraint under the morphological identity approach is input-output correspondence (McCarthy and Prince 1995):
Patterns of Reduplication in Yoruba
313
(3) Input-output faithfulness a. MAXIO Every segment of the input has a correspondent in the output. b. DEPIO Every segment of the output has a correspondent in the input. Since the postulated morphological structure involves two identical compounded bases, input-output faithfulness will ensure that two phonologically identical strings surface. The base forms in (1) are a simplification of the actual, morphologically complex input forms: wole´ ‘look at the house’ derives from wo ‘look at’ þ ile´ ‘house’, peja ‘kill ˙ fish’ derives from pa ‘kill’ þ eja ‘fish’, and so on. As shown in Pulleyblank and Akin˙ labi 1988, the morphological base for this type of reduplication is a verb phrase. To obtain the correct surface forms in such cases, constraints independent of reduplication are required to ensure vowel deletion, retention of H tones, and so forth (Rosenthall 1997; Pulleyblank in press; Orie and Pulleyblank 2002—see also below). Each half of the reduplicated form will be subject to the same constraints. To see this, consider the form pejapeja ‘fisherman’, derived from the morphological form ˙ ˙ [[paþeja][paþeja]]. A constraint prohibiting vowel hiatus rules out sequences of vow˙ ˙ els (Pulleyblank in press; Orie and Pulleyblank 2002) and a left-edge anchoring constraint favors retention of the second vowel in a VV sequence (Pulleyblank in press).3 (4) a. NOHIATUS It is ill-formed to have two adjacent vowels linked to di¤erent sets of features.4 b. ANCHORL The segment at the left edge of a morpheme in the input has a correspondent at the left edge of the morpheme in the output.5 Together these constraints eliminate candidates (5a–c). (5) Vowel sequences in the input
a. b. c. + d. e. f.
/pa-ejaþpa-eja/ ˙ ˙ [paejaþpaeja] ˙ ˙ [paejaþpeja] ˙ ˙ [pejaþpaeja] ˙ ˙ [pejaþpeja] ˙ ˙ [pahejaþpaheja] ˙ ˙ [pajaþpaja]
NoHiatus
DepIO
MaxIO
AnchorL
*!* *!
*
*!
* ** *!* **
*!*
314
Douglas Pulleyblank
In (5d), the deletion of a vowel to satisfy NoHiatus occurs in both halves of the reduplicated form, while in (5e), a similarly motivated consonant epenthesis also occurs in both halves. Such apparent mimicking of one half by the other follows from the requirement that the morphologically defined input of the reduplicated form contain two identical copies of the same morphosyntactic input, each subject to the same markedness constraints (NoHiatus in this case) and to the same inputoutput faithfulness. Note that this type of example establishes that DepIO is ranked above MaxIO in Yoruba: consonant epenthesis (5e) is a more costly faithfulness violation than vowel deletion (5d). Orthogonal to the issue of reduplication, AnchorL selects the candidate where V2 is retained in a heteromorphemic V1 V2 sequence (candidate (5d) over (5f )).6 14.2.2
Base-Reduplicant Correspondence
Consider now the approach to such reduplication based on base-reduplicant correspondence. Under this analysis, the morphology of the Yoruba agentive would be assumed to have a base to which a reduplicative a‰x is attached.7 (6) Agentive reduplication [RedAgentive [VP]] Identity between the reduplicant and the base would be ensured by a set of basereduplicant faithfulness constraints that parallel the input-output constraints of (3). (7) Base-reduplicant faithfulness a. MAXBR Every segment of the base has a correspondent in the reduplicant. b. DEPBR Every segment of the reduplicant has a correspondent in the base. As with the purely morphological analysis sketched above, NoHiatus, DepIO, and MaxIO would ensure that VV sequences in the base are resolved in favor of vowel deletion. This e¤ect would then be transferred to the reduplicant via the basereduplicant faithfulness conditions. This is illustrated by the same case seen above; the reduplicant is underlined for identification:
Patterns of Reduplication in Yoruba
315
(8) Vowel sequence resolution and base-reduplicant correspondence AnchorL
MaxBR
DepBR
MaxIO
DepIO
a. [paejaþpaeja] ˙ ˙ + b. [pejaþpeja] ˙ ˙ c. [pahejaþpaheja] ˙ ˙ d. [paejaþpeja] ˙ ˙ e. [pahejaþpeja] ˙ ˙ f. [peþpeja] ˙ ˙
NoHiatus
/Red þ pa-e: ja/
*!* * *! *!
*
*
*
*!*
*
*!*
The first three candidates (8a–c) show perfect reduplicative copying: of the three, (8b) has the optimal base since the inevitable violation of faithfulness needed to satisfy NoHiatus (cf. (8a)) is achieved through violation of the more lowly ranked MaxIO rather than the higher ranked DepIO (cf. (8c)). The last three candidates (8d–f ) show imperfect copying (violating either DepBR or MaxBR) in spite of an optimal base; as such all three are nonoptimal. Both the morphological identity approach and the base-reduplicant correspondence approach thus successfully derive simple patterns of total reduplication such as the Yoruba agentive. 14.2.3
Comparison: Conceptual Considerations
Before considering empirical reasons for dispreferring the base-reduplicant faithfulness approach, note a curious conceptual redundancy inherent in that theory. As expressed by McCarthy and Prince (1995, 249), ‘‘Constraints of faithfulness demand that the output be as close as possible to the input, along all the dimensions upon which structures may vary.’’ They go on to argue that ‘‘input-output faithfulness and base-reduplicant identity . . . are controlled by exactly the same set of formal considerations, played out over di¤erent pairs of compared structures’’ (p. 250; italics in original). The leading idea of this approach is thus to account for patterns of reduplication through recourse to an independently motivated constraint structure. If faithfulness constraints such as Max, Dep, Contiguity, Linearity, Uniformity, Integrity (McCarthy and Prince 1995) are required to govern input-output relations, then the same constraints should be applicable to the faithfulness relations holding in reduplication. Phrased di¤erently, the goal is to account for reduplication without postulating special reduplication-specific processes or constraints.
316
Douglas Pulleyblank
One could imagine, therefore, that a grammar in which the constraints governing reduplication would literally be the same as those governing nonreduplicative inputoutput configurations would be the best possible realization of this goal. In the proposal that McCarthy and Prince (1995) present, however, this is not the case. Rather, a wholesale reduplication of faithfulness constraints is proposed, where there is a set of faithfulness constraints to govern input-output relations {MaxIO, DepIO, ContiguityIO, . . .}, a complete copy of this set to govern base-reduplicant relations {MaxBR, DepBR, ContiguityBR, . . .}, and a second copy of the set to govern input-reduplicant relations {MaxIR, DepIR, ContiguityIR, . . .}. The basereduplicant set is referred to as the ‘‘basic model,’’ while the inclusion of the inputreduplicant set gives the ‘‘full model.’’ The irony is that while attempting to account for reduplication via independently motivated constraint types, the basic model actually ends up with twice as many faithfulness constraints as are motivated independently of reduplication, and the full model ends up with three times as many constraints. While the correspondence relations postulated fall out from the free assignment of correspondence between any specified phonological elements, the actual number of correspondence constraints needed to deal with the postulated relations triples—with fully two thirds of the full set of correspondence constraints referring specifically to reduplication. Moreover, these reduplication-specific constraints depend on the establishment of an abstract morphological category ‘‘Reduplicant’’ (¼ ‘‘Red’’) to which the reduplication-specific faithfulness constraints refer. As will be seen shortly, this multiplicity of constraints is unnecessary. It is important to distinguish between the structural phonological relations posited by the theory and the constraint set that is proposed to govern those relations. There is no question that two corresponding representations can di¤er in terms of the particular elements contained, the order of such elements, and so on. These relations can then be referenced by constraints such as Max, Dep, Linearity, and so forth. Given multiple representations, these relations can be examined between some input and some corresponding output, between some output and some other corresponding output, between some a‰x and some corresponding base, and so on. At issue for reduplication is whether the substantive set of faithfulness constraints that govern such relations should include constraints specifically referring to reduplicative structures. Allowing construction-specific constraints referring to reduplication not only results in the proliferation of faithfulness constraints just noted, but also raises the possibility of other types of construction-specific constraints. I argue below that it is therefore conceptually desirable to eliminate the class of reduplication-specific faithfulness constraints. For the agentive in Yoruba, this means adopting the morphological identity analysis sketched in section 14.2.2. I therefore also show below how such an approach to reduplication can be applied to cases of partial reduplication.
Patterns of Reduplication in Yoruba
14.3
317
Reduplication with Morphologically Fixed Segmentism: kı´ Infixation
A fundamental property of the proposed analysis is that a single constraint ranking is posited for the full range of reduplicative constructions. While particular reduplicative patterns exhibit di¤erent surface properties (copying of tone, hiatus resolution, and so on), it will be shown that these di¤erences do not require the postulation of di¤erent constraint rankings for di¤erent reduplicative types. In this section, a reduplicative pattern is considered that involves a concurrent ‘‘infix.’’ Exactly the same constraints seen in the previous section account for these ‘‘infixing’’ patterns. The observed patterns di¤er somewhat, however, due to phonological properties of the input forms. 14.3.1
Data
Awoyale (1974) describes a class of reduplicative cases as involving word linkers between two identical noun components. The most widely discussed case involves the morpheme /kı´/. The derived noun that results has the meaning ‘any X’ or pejoratively ‘any old/bad X’. In addition to Awoyale’s treatment, the pattern has been discussed in various places including Ogunbowale 1970, Awobuluyi 1985, Bamgbose ˙ ˙ 1986, and Ola 1995. The following data are taken from Ola 1995. ˙ ˙ (9) /kı´/ infixation Noun Gloss /kı´/ form Gloss a. fı`la` ‘cap’ fı`la`kı´fı`la` ‘any type of cap’ pa´ta´ko´ ‘wood’ pa´ta´ko´kı´pa´ta´ko´ ‘any type of wood’ ja`gu`da ‘thief ’ ja`gu`dakı´ja`gu`da ‘any thief ’ ja`n`du`ku´ ‘dubious person’ ja`n`du`ku´kı´ja`n`du`ku´ ‘any dubious person’ so`we´dowo´ ‘check’ so`we´dowo´kı´so`we´dowo´ ‘any check’ ˙ ˙ ˙ b. omo ‘child’ omoko´mo ‘any child’ ˙ ˙ ˙ ˙ ˙ ˙ ere´ ‘play’ ere´ke´re´ ‘any play’ olorı` ‘queen’ olorı`ko´lorı` ‘any queen’ ake´koˇ ‘student’ ake´koˇka´ke´koˇ ‘any student’ ˙ ˙ ˙ ˙ ˙ ˙ ´ ` ´ ´ko ` ´ ´la ´ ` ´ alakowe ‘educated person’ ala weka kowe ‘any educated person’ ˙ ˙ ˙ It is clear that the presence of kı´ is due to the morphology, not to the phonology. Phonologically, there is no reason a reduplicant cannot occur immediately adjacent to a base, as seen with the agentive forms of section 14.2. If the morphology were to create a sequence such as /pa´ta´ko´þpa´ta´ko´/, there would be no reason to expect it to undergo any phonological modification (cf. ko´le´ko´le´ ‘thief ’). Morphologically, kı´ (plus reduplication) indicates the meaning of ‘any’ or ‘whichever’. Other a‰xes occur in this type of construction with di¤erent meanings. Ogunbowale (1970) lists a vari˙ ety of examples including the following:
318
Douglas Pulleyblank
(10) Various reduplication-triggering morphemes Noun Gloss Reduplicated form a. de´ omo ‘child’ omodo´mo ˙ ˙ ˙ ˙ ˙ ˙ owo´ ‘hand’ owo´do´wo´ ˙ ˙ ˙ ˙ ˙ ˙ b. ko` e`gbe´ ‘side’ e`gbe´ke`gbe´ ˙ ˙ ˙ ˙ ˙ ˙ oju´ ‘face’ oju´koju´ c. nı´ eran ‘animal’ eranle´ran ˙ ˙ ˙ omo ‘child’ omolo´mo ˙ ˙ ˙ ˙ ˙ ˙ d. sı´ ile´ ‘house’ ile´sı´le´ abu´le´ ‘village’ abu´le´sa´bu´le´
Gloss ‘from child to child’ ‘from hand to hand’ ‘side by side’ ‘face to face’ ‘another person’s animal’ ‘another man’s child’ ‘house to house’ ‘from village to village’
Each of these a‰xes can be related semantically, at least diachronically, to a full verb. Ogunbowale (1970) notes the relation of the de´ forms to de´ ‘arrive/reach/extend ˙ to’, the ko` forms to ko` ‘meet’, the nı´ forms to nı´ ‘have/possess’, and the sı´ forms to sı´ ‘to’ (see also Awoyale 1974). Each morpheme therefore has a particular meaning that it contributes to the reduplicative construction. There is an important phonological point to note about these forms, namely that they involve exact copying of the base in conjunction with a lexically idiosyncratic linker. There is no necessary or systematically recurring property in the phonological shape of these a‰xes. Neither their segmental makeup nor their tonal makeup is predictable: segmentally, coronals and noncoronals are observed, as are voiceless and voiced segments, stops and fricatives, oral and nasal segments; tonally, both H and L tones are attested. This variety in specified phonological properties is consistent with the analysis of these segments as morphologically determined, not as representing some sort of emergent phonological property—a point that will be important for cases of reduplication discussed below. 14.3.2
Analysis
The proposed analysis of kı´-reduplication is analogous to that already presented for agentive reduplication where the morphology specifies a compounding structure requiring that both conjuncts be identical. The important di¤erence is that kı´reduplication also stipulates a morphological linker: (11) kı´-reduplication, ‘any old X’ [[Ni ] þ kı´ þ [Ni ]] In its essence, this analysis is one where there is simultaneous morphological specification of reduplication and of a specified a‰x—an overwriting case of the ‘‘table-schmable’’ type (McCarthy and Prince 1995; Alderete et al. 1999, etc.). The di¤erence is that the entire base is copied in addition to the prespecified material. That is, the prefixal kı´ does not get substituted for base material to produce a form like *ja`gu`kı´ja`gu`da or *kı´gu`daja`gu`da. This follows directly from the proposed mor-
Patterns of Reduplication in Yoruba
319
phological structure and an absence of any prosodic shape constraints governing kı´reduplication.8 Since there is no pressure to have the output form conform to a particular prosodic shape, the constraints combine to retain the morphologically specified kı´, except where it is a¤ected by hiatus avoidance, implemented according to the ranking already established in (5). Cases involving a C-initial noun are phonologically transparent: an underlying structure [[ ja`gu`da] þ kı´ þ [ ja`gu`da]] results in the surface form [ ja`gu`dakı´ja`gu`da]. Of more phonological interest are those cases where the base is vowel-initial (9b). In such cases, retention of the full prefix in conjunction with maximal base copying would result in forms like *omokı´omo ‘any child’, *omode´omo ‘from child to child’, ˙ ˙ ˙ ˙ ˙ ˙ ˙ ˙ and so forth instead of the forms in (9) and (10). Outputs with VV sequences are illformed because they violate the prohibition on hiatus (12b); an optimal form like omoko´mo ‘any child’ exhibits vowel deletion as in (12a), where by deleting the first ˙ ˙ ˙ ˙ vowel in the sequence a violation of AnchorL (12c) is avoided. (12) Vowel deletion in reduplication /omo-kı´-omo/ ˙ ˙ ˙ ˙ + a. omoko´mo ˙ ˙ ˙ ˙ b. omokı´omo ˙ ˙ ˙ ˙ c. omokı´mo ˙ ˙ ˙
NoHiatus
DepIO
MaxIO
AnchorL
* *! *
*!
One issue needs to be addressed with regard to the tonal properties of kı´reduplication. As seen in (12), the H tone of the reduplicative prefix is retained, even though it surfaces on the first base vowel rather than the prefix itself. This point is of some importance in the general context of fixed segmentism, particularly relevant for reduplicative patterns such as the gerundive discussed below. The issue is whether the H tone is a lexically specified tonal property or a redundant unmarked tonal category. Independent evidence squarely points to an analysis of the H tone as a lexically specified property (Akinlabi 1984; Pulleyblank 1986, 2004). Evidence from a variety of phenomena—including tonal stability, tones under epenthesis, OCP e¤ects— argues for an analysis where the retention of H and L tones dominates the retention of an M tone, and where an M is the preferred default. For kı´-reduplication, this means that the fixed H must be lexically specified, a pattern that we will see below is paralleled tonally in gerundive reduplication. 14.3.3
Alternations between [i] and [u]
One final point is relevant for this particular type of reduplication. There is a class of cases involving base-initial i@u alternations. These cases illustrate an important
320
Douglas Pulleyblank
diagnostic for distinguishing between reduplication induced by morphological copying and reduplication induced by the satisfaction of prosody. As seen in the examples of (13), the initial [i] of the unreduplicated base sometimes appears as [u] after the kı´ prefix (Bamgbose 1986). Some reduplicated forms are typically produced with the [u] ˙ (13a), while others exhibit variants with [i] and with [u] (13b).9 (13) [i]@[u] alternations Noun Gloss a. `ıse ‘doing’ (ı`-se) ˙ ˙ `ıje ‘eating’ (ı`-je) ˙ ˙ `ıta` ‘selling’ (ı`-ta`) b. `ılo` ‘use’ (ı`-lo`)
/kı´/ form Gloss `ıseku´se ‘bad doing’ ˙ ˙ `ıjeku´je ‘bad eating’ ˙ ˙ `ıta`kuta` ‘bad selling’ `ılo`kulo` ‘bad use’ `ılo`kilo` ‘any use, bad use’ `ırı`n `ırı`nkurı`n ‘walking’ (ı`-rı`n) ‘bad walking’ `ırı`nkirı`n ‘any walking, bad walking’ `ıfe´ `ıfe´ku´feˇ ‘loving’ (ı`-fe´) ‘bad loving’ ˙ ˙ ˙ ˙ `ıfe´kı´feˇ ‘any loving, bad loving’ ˙ ˙ Bamgbose (1986) observes that a pejorative reading is particularly observed with the ˙ forms exhibiting [u]. In addition, he observes that the appearance of [u] in these reduplicative forms is dependent on the form being derived by the prefixation of the abstract nominal prefix `ı-. In underived nouns with an initial [i], only a reduplicative form with [i] is possible (14a); he notes a single exception to this generalization, namely the form for ‘‘anytime’’ in (14b). (14) Underived forms with initial [i] Noun Gloss /kı´/ form a. isu ‘yam’ isukı´su ˙ ˙ ˙ ile´ ‘house’ ile´kı´le´ `ılu´ `ılu´kı´luˇ ‘town’ `ıgba`kugba` b. `ıgba` ‘time’ `ıgba`kigba`
Gloss ‘any yam, bad yam’ ‘any house, bad house’ ‘any town, bad town’ ‘anytime’ ‘anytime’
Two types of analysis have been proposed for these data. One possibility is to attribute the change of [i] to [u] to a process of coalescence (Awobuluyi 1985) or consonant-induced assimilation (Pulleyblank 1988; Ola 1995). This type of solution ˙ is essentially phonological in nature. A purely phonological solution raises a problem, however, in that the alternation in question takes place robustly only in a single environment, an environment that is defined morphologically, not phonologically. Specifically, as already seen, the only robust environment for the i@u alternation is when the second noun is a deverbal nominative, formed by prefixation of `ı-. If
Patterns of Reduplication in Yoruba
321
framed in phonological terms, there is no explanation for why [i] does not change to [u] in the large number of i-initial nouns that do not involve the abstract nominal prefix as in (14a). An alternative is Bamgbose’s (1986) proposal that the u-initial forms involve the ˙ retention of an u`- prefix found in various Southeastern dialects of Yoruba. He notes forms such as the following in the Ijebu dialect: (15) Ijebu dialect: u-initial Noun Gloss u`-lo` ‘using’ u`-se ‘doing’ ˙ u`-je ‘eating’ ˙ u`-ta` ‘selling’
forms permitted /kı´/ form Gloss u`lo`ku´lo` ‘bad use’ u`seku´se ‘bad doing’ ˙ ˙ u`jeku´je ‘bad eating’ ˙ ˙ u`ta`ku´ta` ‘bad selling’
According to this analysis, Standard Yoruba forms like `ılo`kulo` ‘bad use’ derive from /ı`lo`þkı´þu`lo`/. That is, the presence of [u] in the reduplicated forms is attributed to the preservation of the u-prefix on the second noun of the reduplicant, as observed in dialects like Ijebu. A variant of this proposal is to analyze the Standard Yoruba forms as derived directly from an input analogous to the dialectal form: /u`lo`þkı´þu`lo`/. Given the independent prohibition on u-initial forms in Standard Yoruba discussed by Bamgbose, such inputs would be minimally changed into the ˙ actually attested output `ılo`kulo`. Such an approach is not adopted here since it too encounters serious di‰culties. Like the assimilation and coalescence approaches, this analysis fails to account for the limited distribution of the i@u alternation. An initial [u] is common in Southeastern dialects, by no means restricted to nouns involving the abstract deverbal prefix. If u-initial nouns are carried over from such dialects, then why do more cases of the ‘‘u’’ type not occur? Why are forms such as those in (16a) systematically bad if the attested [i] is replaced by [u]? In addition, Olanike Ola Orie (personal communi˙ ˙ ˙ cation) has noted that Standard Yoruba forms of the `ılo`kulo` type may never have polysyllabic bases. Polysyllabic forms are only possible with kı´, not ku´: (16) Polysyllabic Noun a. `ımo`ra`n ˙ b. `ıgba`gbo´ ˙ c. `ıle´rı´
bases: Only Gloss ‘advice’ ‘belief ’ ‘promise’
kı´ possible /kı´/ form `ımo`ra`nkı´mo`ra`n ˙ ˙ `ıgba`gbo´kı´gba`gbo´ ˙ ˙ `ıle´rı´kı´leˇrı´
*/ku´/ form *ı`mo`ra`nku´mo`ra`n ˙ ˙ *ı`gba`gbo´ku´gba`gbo´ ˙ ˙ *ı`le´rı´ku´leˇrı´
Gloss ‘any advice’ ‘any belief ’ ‘any promise’
The polysyllabic restriction is not found in the Southeastern dialects that constitute the putative source for the ku´ forms. An additional problem with the appeal to u-initial nouns is that such forms do not appear to be possible with any other type of reduplication. Recall from (10) that kı´
322
Douglas Pulleyblank
is only one of the possible word linkers in the NþN reduplication type. With other formatives, both nouns must be i-initial; the second noun cannot be u-initial. This is illustrated with forms involving sı´ (cf. 10d). (17) Impossibility of u-forms with sı´ Noun a. igbo´ `ılu´ ipo` `ıjo ˙ b. `ıpa`de´ `ıgba`gbo´ ˙ `ıgbı`mo` ˙ `ıjo`ye`
Gloss ‘forest’ ‘town’ ‘position’ ‘church’ ‘meeting’ (cf. pa`de´ ‘meet’) ‘belief ’ (cf. gba`gbo´ ‘believe’) ˙ ‘committee’ (cf. gbı`mo` ‘confer’) ˙ ‘chief ’ (cf. joye` ‘appoint to o‰cial position’)
/sı´/ form: ‘from X to X’ igbo´sı´gbo´ `ılu´sı´luˇ ipo`sı´po` `ıjosı´jo ˙ ˙ `ıpa`de´sı´pa`de´ `ıgba`gbo´sı´gba`gbo´ ˙ ˙ `ıgba`gbo´sı´gba`gbo´ ˙ ˙ `ıjo`ye`sı´jo`ye`
*/su´/ form *igbo´su´gbo´ *ı`lu´su´luˇ *ipo`su´po` *ı`josu´jo ˙ ˙ *ı`pa`de´su´pa`de´ *ı`gba`gbo´su´gba`gbo´ ˙ ˙ *ı`gba`gbo´su´gba`gbo´ ˙ ˙ *ı`jo`ye`su´jo`ye`
Whether the base of reduplication is an underived noun (17a) or a noun derived by a‰xation of the abstract nominal prefix `ı- (17b), reduplication with sı´ never involves the replacement of [i] by [u]. I propose therefore that the special u-forms observed in kı´-reduplication are a property of the linking prefix itself. That is, the ‘any (old) X’ morpheme has two allomorphs: kı´- and ku´-. The kı´- allomorph has both a neutral ‘any X’ reading and a pejorative ‘any old X’ reading, while the allomorph ku´- is restricted to the pejorative reading. In addition, the kı´- allomorph subcategorizes for the category ‘‘noun’’ while the ku´- allomorph has two more specific restrictions: (i) the noun that ku´- attaches to must be exactly one minimal word, and (ii) ku´- subcategorizes for the abstract nominal prefix ´ı-. The first restriction rules out reduplicative forms with [u] that are polysyllabic (16). The second restriction rules out reduplicative forms with [u] that involve underived nouns (14a) and would also rule out ku´-forms involving consonant-initial bases. The impossibility of u-forms with other linking morphemes such as sı´ is guaranteed by the absence of allomorphy for such morphemes. This analysis interacts with hiatus avoidance. Awobuluyi (1967, 1978) and Pulleyblank (1988, in press) show that the vowels of Yoruba behave asymmetrically with respect to processes of assimilation, deletion, epenthesis, and so on. In the type of contact situation illustrated in (19), nonhigh vowels and back vowels win out over the high front vowel [i] even at the expense of anchoring violations. Following Pulleyblank in press, this asymmetry is attributed to the encoding of sonoritybased distinctions directly into the set of faithfulness constraints. I assume that along the high-nonhigh dimension, nonhigh vowels are more sonorous than high vowels; along the front-back dimension, back vowels are more sonorous than front vowels:10
Patterns of Reduplication in Yoruba
323
(18) Sonority through faithfulness a. Constraint set i. MAXNONHIBK Every NonHi Back element of the input has a correspondent in the output ii. MAXNONHIFR Every NonHi Front element of the input has a correspondent in the output. iii. MAXHIBK Every Hi Bk element of the input has a correspondent in the output. iv. MAXHIFR Every Hi Fr element of the input has a correspondent in the output. b. Harmonic ranking MaxNonHiBk, MaxNonHiFr MaxHiBk MaxHiFr The e¤ect of these constraints is illustrated in (19), setting aside tones that are the focus of the next section. (19) Vowel deletion in reduplicated forms with ku´/ı`lo`-ku´-ı`lo`/ a. `ılo`ku´`ılo`
No Hiatus
DepIO
MaxIO HiBk
HiFr
AnchorL
*
*
*!*
**
**
**
*!
b. `ılo`kilo`
*!
+ c. `ılo`kulo` d. lo`kulo` e. ku´lo`kulo`
NonHi
*!*
The optimal candidate must respect NoHiatus. Deletion of the initial high front vowel of `ılo` (19c) is preferable to deletion of the high back vowel of ku´- (19b) because of the articulated MaxIO constraints. As candidates (19d,e) show, the ku´-forms also have interesting implications for the treatment of reduplication itself. In (19), the surface forms are straightforwardly derived under the assumption that reduplication is the result of morphologically imposed identity. The alternative, base-reduplicant correspondence, is less straightforward. Base-reduplicant correspondence per se is inadequate to account for forms like `ılo`kulo`. If ku´ were analyzed as part of the base then it should be copied: this would incorrectly derive a form like *ku´lo`kulo` (19e). If ku´ were not analyzed as part of the base then it should not be copied: this would incorrectly derive a form
324
Douglas Pulleyblank
like *lo`kulo` (19d). Since part of the input is copied but does not survive on the surface, it would be crucial to invoke input-reduplicant correspondence in addition to the more general base-reduplicant correspondence. Hence this pattern of reduplication would require the ‘‘full model’’ of reduplication, with its proliferation of reduplicant-specific constraints. 14.4
Prosodically Limited Reduplication of Segments Alone: The Distributive
In the two types of reduplication discussed so far, reduplication is total. The next two types to be discussed are partial reduplication, and raise the interesting question of whether reduplication of such a partial nature is consistent with the requirements of morphological identity. Hyman et al. (this volume) and Inkelas and Zoll (2000) argue that it is; it will be argued here, however, that the cases of partial reduplication in Yoruba do not involve morphological identity. But it will also be argued that base-reduplicant and input-reduplicant correspondence is unnecessary. I begin by outlining the facts of distributive reduplication, and then develop the theoretical approach as the analysis of this case is presented. The proposal builds on Kiparsky’s (1986b) proposal that total reduplication corresponds to compounding (the reduplicative types seen in sections 14.2 and 14.3), while partial reduplication corresponds to a‰xation (the reduplicative types about to be examined).11 14.4.1
Distributive Reduplication
Distributive reduplication is a foot-based pattern, involving the copy of an initial VCV sequence (Folarin 1987; Pulleyblank 1988; Ola 1995; among others).12 ˙ ˙ (20) Distributive reduplication a. eweewe´ ‘every leaf ’ ewe´ ‘leaf ’ odoodu´n ‘every year’ odu´n ‘year’ ˙ ˙˙ ˙ osoosu` ‘every month’ osu` ‘month’ ˙ ˙ ˙ ojooju´mo´ ‘every day’ oju´mo´ ‘day’ ˙ ˙ `ıla` b. `ılı``ıla` ‘every line’ ‘line’ o`ro`o`ru ‘every midnight’ o`ru ‘midnight’ `ırı``ıro`le´ `ıro`le´ ‘every evening’ ‘evening’ ˙ ˙ ˙ ˙ o`wu´ro` ‘morning’ o`wo`o`wu´ro` ‘every morning’ ˙ ˙ The distributive pattern exhibits interesting properties. First, it rigidly involves the copy of a single foot.13 If the base constitutes a single binary foot, then the entire foot is copied—for example, eweewe´ ‘every leaf ’. In contrast, if the base is longer than a binary foot, then only the initial VCV sequence is copied, as in ojooju´mo´ ˙ ‘every day’. Second, the distributive exhibits tonal neutralization. If the first mora of the base is M, then both vowels of the distributive are M; if the first mora of the base
Patterns of Reduplication in Yoruba
325
is L, then both vowels of the distributive are L.14 These two core properties of the distributive will be considered in turn, showing also that the vowel assimilation these forms exhibit follows from the analysis proposed. 14.4.2
Integrity Violations
The distributive manifests itself as a single foot, the Yoruba minimal word. Rather than have this constitute a formal condition on some category ‘‘reduplicant,’’ I suggest here that this characterization is by itself su‰cient to derive reduplication without any formal constraints actually referring to reduplication—a return to the type of prosodic characterization of reduplicants found in Marantz 1982, McCarthy and Prince 1986, and so on. Specifically, I propose that the phonological representation of the distributive morpheme is underlyingly /F/ (‘foot’) and that /F/ must be featurally specified on the surface in order to satisfy a family of constraints requiring that segments be featurally specified (Itoˆ and Mester 1993; Padgett 1995; Pulleyblank 1997; and so forth). (21) HAVESPEC A prosodic category must dominate featural specifications. Itoˆ and Mester (1993) explore the possibility that this type of constraint is part of the family of properheadedness constraints. Just as a foot must be headed by a syllable and a syllable headed by a mora, so must a mora be headed by a root node, a root node headed by a place specification, and so on. The precise formulation of this constraint type is not crucial here; what matters is only that they supply pressure to assign features to prosodic categories such as foot and syllable. To see how the analysis works, consider the derivation of a form like ojooju´mo´ ˙ from the input /F-oju´mo´/. HaveSpec rules out the candidate output in which the ˙ foot surfaces without any phonological content at all (22a). MaxIO bans deletion of the a‰xal foot (22b). A third possibility would be to insert default phonological content, a violation of DepIO (22c): (22) Basic options for realization /F-oju´mo´/ ˙ a. F [ ]-oju´mo´ ˙ b. oju´mo´ ˙ c. F [ihi]-oju´mo´ ˙
HaveSpec
DepIO
MaxIO
* * ***
Clearly, none of the outcomes entertained in (22) is reduplicative. A third class of options exists, however. Consider candidates that satisfy the requirement of being a
326
Douglas Pulleyblank
foot, but where each output segment corresponds directly (23a) or indirectly (23b) to some input segment: (23) a.
b.
Such candidates would satisfy the lexical specification that the distributive is a foot, and would do so without incurring any violations of DepIO. DepIO is satisfied in (23a) because every segment in the output directly corresponds to a segment in the input. Struijke (1998) refers to this type of faithfulness as ‘‘broad’’ input-output faithfulness.15 In (23b), DepIO is indirectly satisfied: the reduplicated form corresponds to part of the output base form, which in turn corresponds to part of the input. If the correspondence relations of (23b) are interpreted transitively, then all surface segments correspond to some input segment, hence DepIO is satisfied (see Spaelti 1997). E¤ectively, therefore, there is no need for a special set of constraints to ensure correspondence between the ‘‘Base’’ and the ‘‘Reduplicant’’—such correspondence is simply one way of satisfying the prosodic condition on the distributive without incurring violations of DepIO. An output form like (23a) involves correspondence between the input oju´mo´ and ˙ the stem output oju´mo´, as well as between the input oju´mo´ and the output redupli˙ ˙ cant oju. While this representation is essentially the same as the one that would result from invoking input-output and input-reduplicant faithfulness (McCarthy and Prince 1995, 1999), the nature of the form is actually quite di¤erent. To achieve the result in (23a), input-reduplicant faithfulness would have to (i) posit a category of morpheme ‘‘reduplicant’’ (Red) that is subject to a special class of faithfulness constraints, and (ii) posit a special class of faithfulness constraints relating inputs to the ‘‘reduplicant’’ category so established. The input-reduplicant correspondence relation is directly mediated by the special constraints. In contrast, the analysis proposed here accomplishes a similar relation indirectly. The correspondence relation between the input oju´mo´ and the prefix oju exists not to satisfy some special class of input-reduplicant ˙ faithfulness, but to satisfy the requirement that the prefix have featural content while avoiding violations of the input-output faithfulness requirement Dep. What in general then would prevent such candidates from occurring, unmotivated by any ‘‘reduplicative’’ morpheme? Why do we not observe unprovoked duplication of base material willy-nilly? An answer to this question noted by Spaelti (1997) is to
Patterns of Reduplication in Yoruba
327
be found in the constraint Integrity, proposed and defined by McCarthy and Prince (1995):16 (24) INTEGRITY No element of the input has multiple correspondents in the output. Representations such as (23) violate Integrity since the three input segments /oju/ each have two surface correspondents. The resulting Integrity violations will be tolerated only if forced by some higher constraint—the standard OT scenario. In the case of the distributive morpheme, Integrity violations are forced by the need to satisfy the constraint requiring that the lexically specified foot of the distributive dominate featural content. Consider three possible candidates for the distributive form ojooju´mo´ (base oju´mo´ ), ˙ ˙ first under the assumption of indirect correspondence (comparable to (23b)): (25) Basic options for realization (continued) /F-oju´mo´/ ˙ a.
HaveSpec
DepIO
FootBin
*!**
+ b.
c.
Integ
***
*!
*
328
Douglas Pulleyblank
/F-oju´mo´/ ˙
HaveSpec
DepIO
d.
FootBin
Integ
*!
*****
In (25a), the first three segments of the distributive are accidentally homophonous with segments of the base. Accidental homophony, however, does not constitute formal correspondence, hence each such segment constitutes an unforced violation of DepIO (fatal violations given the ranking). A candidate such as *ihioju´mo´ would be ˙ similarly excluded by DepIO. Comparing candidates with partial copying (25b,c) versus complete copying (25d), we see that copying will take place to the point where prosodic constraints are satisfied—in this case, the requirement that a foot be binary (FootBin—see Ola 1995; Orie and Pulleyblank 2002). ˙ The tableau in (25) indicates only the ‘‘reduplicative’’ core. All four candidates shown violate NoHiatus. These vowel contact violations can be avoided by epenthesizing a consonant (26a), by copying an additional segment (26b), or by copying less and postulating a multiply linked segment (26c). Note that a ‘‘V-deletion’’ option (*F [oj]-oju´mo´)/*F [ojo]-ju´mo´) would not be viable since it would either violate left˙ ˙ edge anchoring or prosodic conditions on feet. (26) Resolving the hiatus problem /F-oju´mo´/ ˙ a. F [oju]h-oju´mo´ ˙ b. F [oju]m-oju´mo´ ˙ + c. F [ojo]-oju´mo´ ˙
NoHiatus
DepIO *!
MaxIO
AnchorL
Integ *** ***!* **
MaxIO plays no role in distinguishing (26b) and (26c) since all input segments have output correspondents in both. The basic conclusion is that constraints with specific reference to a category type ‘‘reduplicant’’ are redundant. Given prosodically specified morphemes with incomplete specifications (Marantz 1982), appropriately ranked constraints on inputoutput correspondence will evaluate reduplicative forms as optimal. Under this view,
Patterns of Reduplication in Yoruba
329
reduplication is an accidental by-product of underspecified templates with independently motivated constraints. A formal issue arises under this view as to whether reduplication should be treated as involving direct (23a) or indirect (23b) correspondence to the input form. In (25), indirect correspondence was assumed, but the alternative of direct correspondence was not considered. Abstracting away from all aspects of (26c) other than the two segments that are in correspondence, consider how the constraint set would evaluate the two forms in (23). (27) Direct versus indirect correspondence in reduplication /F-oju´mo´/ ˙
DepIO
MaxIO
AnchorL
Integ
a.
**
b.
**
As can be seen, the constraint set does not distinguish between the two patterns of correspondence.17 This changes, however, if Linearity is brought into play (McCarthy and Prince 1995). (28) LINEARITY The linear order of elements in S1 is identical to the linear order of their corresponding elements in S2 . Because of the transitive correspondence relation between the elements in correspondence in (29b), Linearity is fully respected. In contrast, the direct correspondence relation observed in (29a) brings about Linearity violations. For example, in the input oju´mo´, o precedes j but in the output, a direct correspondent of the input o ˙ follows j (ojuoju´mo´). Such contradictory precedence relations are an indication that ˙ Linearity has been violated.18
330
Douglas Pulleyblank
(29) The relevance of linearity /F-oju´mo´/ ˙
DepIO
MaxIO
AnchorL
Integ
Lin
a.
**
*!
+ b.
**
Note that the ranking of Linearity with respect to the other constraints is not crucial here because the two candidates di¤er only in their satisfaction/violation of this one constraint. While the correspondence relations of (29a) are comparable to those of inputreduplicant correspondence, and those of (29b) are comparable to those of basereduplicant correspondence, it is important to keep in mind that the status of such relations is fundamentally di¤erent in this approach to that proposed in McCarthy and Prince 1995. Unlike the McCarthy and Prince proposal, which posits faithfulness constraints that explicitly govern these relations, in the approach taken here the correspondence relations are indirectly governed through reference to nonreduplicative constraints such as Integrity and Linearity. To summarize, the proposal made here is that phonological representations do not need to postulate an ad hoc category ‘‘Reduplicant’’ in order to derive reduplicated forms. There is no need for special reduplication-specific classes of correspondence constraints of the Base-Reduplicant and Input-Reduplicant types as proposed in McCarthy and Prince 1995. Instead, forms with correspondence relations between one morpheme and another are possible candidates produced by Gen, and they will be analyzed as optimal in situations where a morpheme is prosodically characterized, incompletely specified, and where Dep and Integrity are appropriately ranked. This proposed elimination of the class of reduplication-specific faithfulness constraints requires, however, consideration of several further issues, to which I now turn: (i) why prosody-induced reduplication cannot be reduced to reduplication by morphological identity, (ii) why a return to a prosodic characterization of reduplication does not induce base erosion, and (iii) how one accounts for e¤ects of the emergence of the unmarked.19
Patterns of Reduplication in Yoruba
14.4.3
331
Reduplication Not Morphologically Induced
Vowel contact arising from distributive reduplication is unlike the general case where two vowels abut, and also unlike cases of abutting vowels in kı´-reduplication. As shown in this section, the observed behavior is, however, precisely the expected behavior where reduplication results from the satisfaction of prosodic requirements, not from morphological identity. The crucial point for hiatus resolution is that the reduplicant is defined by a particular prosodic shape, as argued for the distributive, and not by morphological identity in conjunction with a prosodic delimiter.20 Consider distributive reduplication cases where the initial vowel of the base is [i] (reproduced from (20)). (30) Distributive reduplication with i-initial bases `ılı``ıla` `ıla` ‘every line’ ‘line’ `ırı``ıro`le´ `ıro`le´ ‘every evening’ ‘evening’ ˙ ˙ ˙ ˙ If reduplication were morphologically induced, we would have inputs such as /ı`la`-ı`la`/. Given the phonological grammar of Yoruba developed in sections 14.2 and 14.3,21 this would lead to the incorrect prediction that the surface form should be *ı`la`a`la`. The tableau in (31) illustrates this and should be compared with the tableau in (19); here and elsewhere, a bomb (M) in a tableau designates a candidate that is evaluated as optimal by the analysis but is in fact ungrammatical. (31) Retention of base vowel in distributive reduplication: Morphological identity /ı`la`-ı`la`/ a. `ıla``ıla`
NoHiatus
DepIO
NonHi
HiFr
AnchorL
*
*
*!
M b. `ıla`a`la` c. `ılı``ıla`
MaxIO HiBk
*!
For additional discussion of hiatus resolution in sequences of underlyingly specified vowels, see Pulleyblank 1988, in press; Rosenthall 1997; Orie and Pulleyblank 2002. If reduplication is the result of prosodic specification of a foot, then the correct result is predicted:
332
Douglas Pulleyblank
(32) Retention of base vowel in distributive reduplication: Prosodically specified a‰x /F-ı`la`/ a. `ıla`ı`la`
NoHiatus
DepIO
NonHi
MaxIO HiBk
HiFr
AnchorL
*!
b. `ıla`a`la`
*!
+ c. `ılı``ıla` Because the distributive has no segmental specifications in the input, MaxIO is irrelevant for the prefixal foot;22 because each of the segments of the input `ıla` has one or more surface correspondents in the three output candidates of (32), MaxIO is fully satisfied by all three candidates. As a result, MaxIO is irrelevant and left-edge anchoring of the morpheme `ıla` ensures selection of `ılı``ıla` as optimal rather than *ı`la`a`la`. Morphologically induced reduplication thus must be distinguished from reduplication that is prosodically induced. While this is a return in spirit to the proposal of Marantz 1982, it is a return mediated by the theory of correspondence proposed by McCarthy and Prince 1995, yet not requiring the reduplication-specific constraints proposed there. 14.4.4
Why the Base Does Not Erode
It is argued in Spaelti 1997 and McCarthy and Prince 1999 that it is undesirable to characterize reduplicative morphemes prosodically due to the potential for base erosion—a problem referred to by McCarthy and Prince as the ‘‘Kager-Hamilton problem.’’ If constraints of base-reduplicant correspondence are posited, then it should be possible for these to be in a language in which MaxBR outranks MaxIO. Given such a ranking, if reduplicants are prosodically characterized, then it should be possible for satisfaction of a prosodic Red constraint (e.g., Red ¼ Foot) and MaxBR to force erosion of the base. For example, if this ranking were to apply to an input like RED-oju´mo´, then the optimal output (ignoring constraints such as ˙ NoHiatus) would be ojuoju´ (the reduplicant is underlined): (33) Base-reduplicant correspondence resulting in base erosion /Red-oju´mo´/ ˙ a. [ojumo-oju´mo´] ˙ ˙ b. [oju-oju´mo´] ˙ + c. [oju-oju´]
MaxBR
Red ¼ Foot
MaxIO
*! *!* **
Patterns of Reduplication in Yoruba
333
Spaelti (1997) and McCarthy and Prince (1999) assume that base erosion is nonexistent and seek to eliminate it as a possibility by prohibiting constraints like Red¼ Foot from referring to prosodic constituents. ‘‘Generalized template’’ theory (see McCarthy and Prince 1999 and references therein) is one approach to this type of restriction. Two points are important in addressing this issue. First, it must be established whether there are valid cases of base erosion. If, for example, Hausa is correctly considered to exhibit base erosion (Downing 2000), then there is a need for an analysis of base erosion. Second, whatever the answer to the question of whether base erosion exists, it would appear that such erosion should not be the result of phonological constraints. That is, if base erosion exists, then it must be of a type that is morphologically rare. The proposal elaborated here is that base erosion should be accounted for by the sorts of morphological mechanisms responsible for subtractive morphology. Just as subtractive morphology in general is rare, so then should base erosion be rare. The crucial aspect of the current phonological proposal is that base erosion cannot result from the properties of the purely phonological constraints. By eliminating from the theory all faithfulness constraints referring explicitly to the category of ‘‘reduplicant’’—that is, by eliminating all constraints of the BR and IR classes— base erosion is completely excluded as a phonologically induced phenomenon. Consider candidates such as oju-oju´mo´ and oju-oju´ from the perspective of the ˙ analysis presented here, abstracting away from issues such as hiatus avoidance: (34) Absence of erosion in a theory without reduplicant-specific constraints /F-oju´mo´/ ˙ a. [ojumo-oju´mo´] ˙ ˙ + b. [oju-oju´mo´] ˙ c. [oju-oju´]
DepIO
FootBin
MaxIO
*!
Integ ***** ***
*!*
***
MaxIO is violated by deletion of base segments but is una¤ected by the number of corresponding segments found in the reduplicant (each of which violates Integrity). The relative ranking of MaxIO and Integrity is not crucial since a candidate with an eroded base (34c) incurs as many violations of Integrity as a form that is faithful to the base (34b). As a simple consequence of not having a category ‘‘reduplicant,’’ base-reduplicant correspondence cannot be ranked above Input-Output correspondence since there is no base-reduplicant correspondence at all. Phonologically induced erosion is eliminated.
334
14.4.5
Douglas Pulleyblank
Tonal Properties: Emergence of the Unmarked
The final point to be considered with regard to distributive reduplication is the construction’s tonal properties. As pointed out with reference to (20), the tone of the distributive reduplicant is completely predictable: both moras of the reduplicant are M if the first mora of the base is M (20a) and L if the first mora of the base is L (20b); recall from section 14.3.4 that H tones may not occur on a word-initial vowel, hence will not occur on a relevant reduplicative base. In earlier accounts of the distributive within a theory of tonal underspecification (Akinlabi 1984; Pulleyblank 1986), this pattern has been accounted for by (i) not copying tone during the reduplicative process, (ii) postulating a rule of leftward L tone spreading, and (iii) assigning midtones by default. The crucial aspect of these analyses is the rule of leftward spreading. Whether tone is copied or not during reduplication would determine whether the rule must be feature-changing or featurefilling, but would not a¤ect the necessity of the rule itself. Similarly, if M is a default value then the rule could spread L tone only, whereas if M was a specified value then the rule could be formulated so as to spread all tones. While the exact treatment of tone spreading may not be crucial, let us assume for concreteness that the grammar of Yoruba has a constraint-based equivalent of tone spreading as in (35). (35) ALIGNLEFT[TONE] AlignLeft(Tone; Word) The e¤ects of this constraint will be superseded in nonreduplicative forms by the exigencies of regular input-output faithfulness. Following work such as Lombardi 1995, 1998, Pulleyblank 1996, and Myers 1997, I assume here feature-based faithfulness and reference in faithfulness constraints to feature paths (Archangeli and Pulleyblank 1994). With regard to tone, therefore, the following constraints are relevant for an evaluation of faithfulness: (36) a. MAXTONE Every tone of the input has a correspondent in the output. b. DEPTONE Every tone of the output has a correspondent in the input. c. MAXPATHTONE Every path between a tone and an anchor in the input has a correspondent in the output. d. DEPPATHTONE Every path between a tone and an anchor in the output has a correspondent in the input. MaxTone will prevent the elimination of tones and MaxPathTone will prevent the redistribution of tones. DepTone will prevent the insertion of tones and DepPath-
Patterns of Reduplication in Yoruba
335
Tone will prevent the spreading of tones. If the tonal faithfulness constraints outrank AlignLeft[Tone], then there will be no leftward spreading. This can be seen by consideration of a word like `ıro`le´ ‘evening’ (from (20)). ˙ ˙ (37) Tonal faithfulness irole ˙ ˙ l h
MaxTone
DepTone
MaxPath Tone
+ a. irole ˙ ˙
AlignL Tone
DepPath Tone
**(H)
l h b. irole ˙ ˙
*!
*!
*
l *!
c. irole ˙ ˙
*(H)
*
l h d. irole ˙ ˙
*!
*!*
**
h Note in particular that the attested surface form `ıro`le´ violates AlignLeft twice. ˙ ˙ Since this form is nevertheless optimal, this means candidates respecting or minimally violating AlignLeft must be ruled out by more highly ranked constraints. Specifically, candidates (37b,d) must be ruled out by the higher ranking of at least one of MaxTone, MaxPathTone, and DepPathTone, while candidate (37c) must be ruled out by the higher ranking of either MaxPathTone or DepPathTone. The actual assumption made in (37) is that MaxTone and MaxPathTone outrank AlignLeft and that AlignLeft in turn outranks DepPathTone. While the low ranking of DepPathTone would have no e¤ect on simple forms, it would have a significant e¤ect on a reduplicated form. Consider the following tableau illustrating the reduplicated form `ırı``ıro`le´ ‘every evening’. (Consistent with the results of (37), only ˙ ˙ candidates that respect root tone assignments are illustrated.)
336
Douglas Pulleyblank
(38) Tonal alignment in reduplication Fþirole ˙ ˙ l h
MaxTone
DepTone
MaxPath Tone
AlignL Tone
DepPath Tone
*!*(L) ****(H)
a. iriirole ˙ ˙ l l h + b. iriirole ˙ ˙
****(H)
**
**(L) ****(H)
**
l h *!
c. iriirole ˙ ˙ m l h
The first candidate (38a) contains a tone in correspondence with the root; the second candidate (38b) spreads the tone of the root; the third candidate (38c) assigns a default tone to the reduplicant. Note that the associations of the initial L in (38a) do not constitute DepPathTone violations because both the L and the associations have correspondents in the input. Although the choice of candidate (38b) over (38a) does not result in a surface tonal di¤erence, the comparable choice in a form like ojooju´mo´ ‘every day’ is crucial: ˙ (39) Tonal alignment in reduplication (continued . . .) Fþojumo ˙ m h
MaxTone
DepTone
MaxPath Tone
AlignL Tone
DepPath Tone
*!(H) **(M) ***(H)
a. ojo ojumo ˙ m hm h + b. ojo ojumo ˙
***(H)
**
*(H)
***
m h c. ojo ojumo ˙ m
h
*!(M)
Patterns of Reduplication in Yoruba
337
By spreading the initial M of oju´mo´ (39b), AlignLeft is better satisfied than if tones ˙ are in correspondence (39a). Candidate (39c) shows that in a derived form just as in an underived form, spreading is constrained by faithfulness so as not to a¤ect root specifications. The failure to copy tone has been derived by ranking a constraint on tonal distribution (AlignLeft[Tone]) above a relevant faithfulness constraint. As such, this case exhibits a pattern of the ‘‘emergence of the unmarked’’ (McCarthy and Prince 1994, etc.). Note that many languages systematically require tonal specifications to be leftaligned—the left-edge asymmetry is built into early autosegmental treatments of tone such as Leben 1973 and Goldsmith 1976. While not generally true of Yoruba due to the high ranking of particular tonal faithfulness conditions, we see left-edge preference emerging under reduplication.23 The logic of the proposed analysis of tone in the distributive is that a set of faithfulness constraints (Faithi ) prevents some (markedness) constraint (C) from having any e¤ect in basic forms. (40) Faithi C With regard to some more lowly ranked faithfulness constraint set (Faithj ), however, C is dominant. (41) C Faithj Putting the two subrankings together, we get the following: (42) Faithi C Faithj This schema is precisely the ranking that derives emergence of the unmarked: the marked case is tolerated generally, but the unmarked case ‘‘emerges’’ in contexts governed only by Faithj . In the case of the Yoruba distributive, Faithi is the set {MaxTone, DepTone, MaxPathTone}, C is {AlignL[Tone]}, and Faithj is {DepPathTone}. The general point is that emergence of the unmarked e¤ects in reduplication does not depend on the postulation of a special set of base-reduplicant correspondence constraints. Even in a theory that does not include such construction-specific constraints, unmarked e¤ects may emerge from the way di¤erent input-output faithfulness constraints are ranked with respect to particular markedness constraints. Note also that whether tone is copied in a particular type of reduplication or not is straightforwardly derived from the constraint grammar. Contra Walsh-Dickey (1992, 1993), to distinguish between cases where tone is or is not copied, there is no need to assign tone to di¤erent structural tiers. Moreover, it is not the case that all reduplicative processes will behave the same way in the same language. We have seen that agentive reduplication and kı´-reduplication copy tone, for example, while distributive
338
Douglas Pulleyblank
reduplication does not. Such di¤erences in whether tone is copied or not are consistent with the constraint-based approach taken here but inconsistent with a theory where tone must be represented on di¤erent tiers to account for its variable behavior in reduplication. 14.5
Gerundive Reduplication
The last case of reduplication to be considered in this chapter is the gerundive. In addition to issues of descriptive interest that arise in connection with this pattern of reduplication, two issues of theoretical significance are addressed, namely fixed default segmentism and underapplication. These issues are of particular interest because the Yoruba gerundive has been taken as a paradigm case of base-reduplicant correspondence in Alderete et al. 1999. Gerundive reduplication is illustrated in (43):24 (43) Gerundive reduplication Verb Verbal noun/gerund a. rı´ [rı´] rı´rı´ [rı´rı´] ´ ´ ´ ´ b. gbe [gbe] gbıgbe [gbı´gbe´] c. je [J] jı´je [Jı´J] ˙ ˙ d. da´ra [da´ra] dı´da´ra [dı´da´ra] fı´fo´ [fı´f ´] e. fo´ [f ´] ˙ ˙ f. to´bi [to´bi] tı´to´bi [tı´to´bi] g. bu` [bu`] bı´bu` [bı´bu`] h. sı´n [sı˜´ ] sı´sı´n [sı´sı´˜ ] i. gba´n [gba˜´] gbı´gba´n [gbı´gba˜´] j. du`n [du`˜ ] dı´du`n [dı´du˜`]
Gloss ‘see’ ‘take’ ‘eat’ ‘be good’ ‘shatter’ ‘be big’ ‘dip out’ ‘sneeze’ ‘scoop out’ ‘be sweet’
c
c
This reduplicative pattern has been used to illustrate the analysis of reduplication with fixed segmentism in various prominent treatments of reduplication within prosodic models—for example, Marantz 1982, Kenstowicz 1994, and Alderete et al. 1999. One of the reasons for choosing the Yoruba pattern to illustrate fixed segmentism is its apparent regularity and simplicity: the gerundive forms are derived by a‰xing a [Cı´-] prefix, where the ‘‘C’’ is a copy of the first consonant of the verbal base. The fixed segmentism is the H-toned high front oral vowel [ı´]. Deeper investigation into the prosodic, tonal, and nasal properties of the gerundive prefix, however, reveals a much more complex situation, and a useful test of the approach to reduplication advocated in this chapter. This section is structured as follows. First, the prosodic properties of gerundive reduplication are examined. Second, the marked tonal features are incorporated into the analysis. Third, the vocalic properties, though unmarked, are argued to require
Patterns of Reduplication in Yoruba
339
prespecification. Finally, general properties of nasality in Yoruba are laid out, followed by a treatment of nasality in reduplication. 14.5.1
Prosodic Aspects of Gerundive Reduplication
Prosodically, the gerundive consists of a single syllable. There are three types of explanations for this shape: (i) it could be specified as part of the lexical entry of the gerundive that it be one syllable, (ii) there could be morphological reasons why the gerundive a‰x might be one syllable, and (iii) there could be phonological reasons why the morpheme might be one syllable. The possibility that the gerundive is a syllable due to the lexical specification ‘‘s’’ (Ola 1995) might appear to be the most straightforward approach. In fact, however, ˙ it seems redundant. In order for the template ‘‘s’’ to result in a CV surface form, it would be necessary for some constraint to force the presence of an onset consonant. The constraint ‘‘Onset,’’ requiring that all syllables have an onset, would be a plausible source of pressure. As argued by Orie and Pulleyblank (2002), however, general considerations of hiatus resolution in Yoruba argue that Onset is ranked below both Max and Dep—too low to force the presence of an onset in a morpheme specified only by a prosodic template. It might nevertheless be imagined that Onset could force the presence of an onset in the gerundive if we were to assume that it outranks Integrity. (44) Prosodic shape of the gerundive /s-to´bi/
HaveSpec
DepIO
a. to´bi b. s-to´bi c.
MaxIO
Onset
*! *!
*
´ ]bito´bi s [to
***!*
+ d. s [to´]to´bi e.
´ bi s [hı´]to
f.
´ ]to´bi s [o
Integ
** *!* *!
*
The optimal result would be a candidate filling the gerundive syllable with features in correspondence to the base (44c,d); since the specification of such features involves Integrity violations, however, there would be a preference for including the minimal number of specifications necessary to fill the lexically assigned syllable—hence (44d) would be preferred to (44c). (Note that the actual output, tı´to´bi, has fixed segmentism; this is addressed below.)
340
Douglas Pulleyblank
Attributing the impossibility of *o´to´bi to Onset, however, produces an immediate problem. If Onset is ranked above Integrity, then vowel-initial forms should be ruled out in general—a prediction roundly falsified in Yoruba (consider (9), (13), (20), and so on). To prevent such overproduction of reduplicative onsets, it is crucial that Integrity outrank Onset, and therefore that some other constraint rule out a form like *o´to´bi in (44f ).25 As will be seen below, the identification of the relevant constraint renders the lexical postulation of a syllable for the gerundive redundant. Turning to other possible explanations for the gerundive shape, one might imagine, following generalized template theory (Urbanczyk 1996; McCarthy and Prince 1999, etc.), that it falls out of general considerations of a‰x shape. Attempting to derive the CV form of the gerundive by stipulating that ‘‘Gerundive ¼ A‰x’’ turns out to be somewhat problematic, however. First, note that one of the strongest motivations behind generalized templates (namely how to prevent base erosion) has been removed by the absence of base-reduplicant correspondence constraints. As discussed in section 14.4.3, the problem of base erosion does not arise if there is no basereduplicant faithfulness. Without base erosion as a problem, a central reason for generalized templates disappears. In any event, a more direct problem arises when we actually consider the prosodic properties of Yoruba a‰xes. Consider the list in (45) (Bamgbose 1967; Rowlands ˙ 1969):26 (45) Yoruba prefixes a`nominalizer a`lo ‘going’ < lo ‘go’ ˙ ˙ aagentive akorin ‘singer’ < korin ‘sing’ ˙ ˙ `ıabstract nominal `ıda´wo´ ‘subscription’ < da´wo´ ‘subscribe’ a``ınegative a``ıda´ra ‘not being good’ < da´ra ‘be good’ a`tisise´ ‘to work’ a`ti- infinitival ‘work’ < sise´ ˙ ˙˙ ˙ ˙˙ onı´le´ ‘landlord’ < ile´ ‘house’ onı´- owner of X o`/o`- agentive o`sise´ ‘worker’ < sise´ ‘work’ ˙ ˙ ˙˙ ˙ ˙˙ olu`- agentive olu`ko´ ‘teacher’ < ko´ ‘teach’ ˙ ˙ On the basis of this list, the primary generalizations governing Yoruba a‰xes are as follows: (i) all a‰xes in nonreduplicative forms are vowel-initial, and (ii) the first (and sometimes only) mora of an a‰x is nonhigh (M or L). Phrased negatively, a‰xes are never consonant-initial and the initial mora of an a‰x is never H.27 These requirements mirror general conditions on the optimal shape of a noun, conditions that define that shape as [VCV . . . ] (Ola 1995). But if being vowel-initial and bearing ˙ a nonhigh tone are the optimal properties for an a‰x, then why does the gerundive reduplicant do exactly the opposite? Why does the gerundive begin with a consonant and bear an H tone? Moreover, there would be only weak language-internal grounds
Patterns of Reduplication in Yoruba
341
for analyzing a canonical a‰x in Yoruba as monomoraic when roughly 50 percent of the attested a‰xes are bimoraic. The basic point is that the prosodic properties of a‰xes in Yoruba, however ensured by constraints, cannot be responsible for deriving the gerundive shapes in (43) since the gerundive shape is not canonical. The gerundive shape could only arise in spite of general constraints on the prosodic realization of a‰xes. Hence generalized template theory must look to constraints other than those defining a‰x shape to account for the prosodic properties of the gerundive. We are therefore left with the possibility that the syllable shape of the gerundive is due to phonological properties interacting with its lexical specification. The crucial observation, made by Ola (1995), is that a word-initial bare V cannot carry an H ˙ tone. She suggests that this constraint accounts for the initial C of gerundive reduplication. By assuming (i) that the gerundive morpheme underlyingly contains an H tone, (ii) that the constraint demanding faithfulness to such an H (MaxH) is undominated, and (iii) that the constraint prohibiting an H-toned bare initial V is also un´ ’’), it follows that some modification dominated (a constraint that will be labeled ‘‘*[V of an H tone input is required. In exploring the implications of a lexical H tone specification, I assume an input for the gerundive that consists simply of an H tone mora; assuming the presence of a syllable as well would play no crucial role though it would be fully compatible with the surface pattern observed. (46) Underlying H tone initial vowel /mH-to´bi/
MaxH
a. [ı´-to´bi] b. [i-to´bi]
´ *[V *
*
Ola proposes that the necessary modification involves reduplication. By adding a ˙ reduplicative consonant to the H-toned prefix, both violations of (46) disappear: (47) Inclusion of reduplicative consonant /mH-to´bi/
MaxH
´ *[V
a. [tı´-to´bi] Ola suggests that this analysis fills out an apparent gap in the paradigm of tonal ˙ possibilities on a‰xes. Data such as in (45) show that of the three lexical tones of Yoruba, H, M, L, only two, M and L, appear on prefixes. Given the combinatorial possibilities that such tones would allow, we would expect H-toned prefixes to be possible—the gerundive reduplicative prefix constitutes exactly that case. Since an
342
Douglas Pulleyblank
H tone is marked (section 14.3.4), its presence in the gerundive must be lexically specified—an ‘‘overwriting string’’ in the terminology of Alderete et al. 1999. To summarize, it has been shown that neither the lexical specification of ‘‘s’’ nor the general properties of being an a‰x are su‰cient to account for the prosodic properties of the gerundive prefix. To derive the observed CV shape of the gerundive, it is proposed that the reduplicant bears an H tone underlyingly, a feature whose presence on a vowel forces the appearance of an onset due to the particular con´ . Of direct importance in this context is the conclusion reached in section straint *[V 14.3.2 concerning the markedness relation between H, M, and L tones. If, as concluded there, the H tone is the most marked tone (at least for Yoruba), then its presence in the gerundive reduplicant can only be due to the lexical presence of the H: it cannot be a default or unmarked realization if H is the most marked tone, and there is no general rule of H-insertion that could be considered to apply to the reduplicative prefix of the gerundive. The lexical representation of the gerundive must therefore include an H tone mora (mH). 14.5.2
The Segmental Nature of the Gerundive
As noted above, the gerundive exhibits the fixed vowel quality [i], a quality that Alderete et al. (1999) attribute to emergent unmarkedness in the reduplicant. Their proposal for the gerundive reflects the general schema of ‘‘emergence of the unmarked’’ e¤ects (McCarthy and Prince 1994, etc.), where a markedness constraint is ranked between input-output faithfulness and base-reduplicant faithfulness: FaithIO Markedness FaithBR. Specifically for the Yoruba case, constraints favoring [i] over other vowels are ranked below MaxIO but above MaxBR. Because they are ranked below MaxIO, they have no e¤ect on simple inputs; because they are ranked above MaxBR, they cause the vowel of the gerundive to be [i]. Just as they can within a theory assuming base-reduplicant faithfulness, patterns of ‘‘emergence of the unmarked’’ can be derived within the approach taken here. Nevertheless, when the large set of reduplicative patterns in Yoruba is analyzed as a whole, it is implausible that the [i] of the gerundive is emergent. To understand these points, consider the typology generated by DepIO above Integrity: in a grammar that ranks Integrity above DepIO, epenthesis is the optimal way to resolve the problem of an incompletely specified morpheme; in a grammar that ranks DepIO above Integrity, reduplication is the optimal strategy. Since Yoruba adopts the reduplicative strategy, it is necessary to rank DepIO above Integrity. Consider an example of the distributive, ojooju´mo´, along with an example of the gerundive, ˙ tı´to´bi.
Patterns of Reduplication in Yoruba
343
(48) Divergent reduplicative patterns /F-oju´mo´/ ˙ + a. ojo-oju´mo´ ˙ b. ijo-oju´mo´ ˙ c. iho-oju´mo´ ˙ /mH-to´bi/
DepIO
Integ **
*!
*
*!*
M d. to´-to´bi
**
e. tı´-to´bi
*!
f. hı´-to´bi
*!*
*
The problem with such pairs of forms is that the ranking of DepIO above Integrity ensures reduplication, and also results in the reduplication of both consonants and vowels. This is correct for the distributive, but incorrect for the gerundive. To resolve this problem, two approaches are possible. The first is to lexically specify the gerundive for the vowel [i], rather than just for a mora (Marantz 1982). This is the solution adopted here: (49) Divergent reduplicative patterns /ı´-to´bi/ a. to´-to´bi + b. tı´-to´bi c. hı´-to´bi
MaxIO
DepIO
*!
Integ **
*
*
*!*
This prespecification approach correctly derives the gerundive without jeopardizing the analysis already given of the distributive. Alderete et al. (1999, 358) mention this possibility but note that ‘‘a‰xation would fail to account for the correlations with independently motivated defaults.’’ Correlations, of course, may be accidental. Interestingly, Akinlabi (2000) argues for independent reasons that the [i] of the gerundive must be underlying, not default. He notes that assimilatory processes treat the gerundive vowel di¤erently than cases that are unambiguous instantiations of default [i]. Abı´o´du´n (1997) also suggests that the [i] of the gerundive is underlyingly specified. ˙ He observes that a pattern of ideophonic reduplication systematically specifies the vowel [a], noting therefore that the quality of the vowel in a context of fixed segmentism is not predictable.
344
Douglas Pulleyblank
(50) Fixed [a] in ideophonic reduplication a. fakafı`kı` ‘up and down movement’ b. wa`ra`we´re´ ‘very quickly’ c. ya`da`yo`do` ‘in a joyful mood’ d. gagagu´gu´ ‘plenty of load’ In spite of these arguments in favor of prespecification, it is worth considering an alternative whereby faithfulness is divided into faithfulness to vowel features and faithfulness to consonant features (Alderete et al. 1999). This alternative could derive the gerundive by ranking Integrity below input-output faithfulness to consonants (deriving the reduplication of consonants) and above input-output faithfulness to vowels (deriving default vowel ‘‘epenthesis’’). The split-faithfulness approach fails, however, when attempting to derive both the gerundive and the distributive since to succeed on the gerundive (copying consonants but not vowels) means to fail on the distributive (which copies both consonants and vowels). In spite of its overall failure, it is worth considering the split-faithfulness approach briefly since it illustrates one way of achieving emergence of the unmarked e¤ects even without base-reduplicant faithfulness constraints. Consider the tableau in (51): (51) Emergence of the unmarked /mH-to´bi/
DepConsonant InputOutput
a. to´-to´bi
Integ
DepVowel InputOutput
**!
+ b. tı´-to´bi
*
c. ho´-to´bi
*!
d. hı´-to´bi
*!
*
* *
The optimal candidate would reduplicate a consonant to avoid a violation of highranking input-output constraints on faithfulness to consonants, but would insert an epenthetic vowel to prevent gratuitous violations of Integrity. Note, however, that (51b) actually stands for a class of candidates where the vowel of the gerundive consists of [i] (tı´-to´bi), [e] (te´-to´bi), [a] (ta´-to´bi), and so on. However they are formulated, constraints on vowel markedness would intervene to select the candidate with the least marked vowel (tı´-to´bi) since all such constraints tie on all other relevant constraints. The nonviability of the emergence of the unmarked analysis of the gerundive lies not in some inability to produce such e¤ects, but in independent properties of reduplication in general in Yoruba and of the gerundive in particular.28 In conclusion, I analyze the fixed [i] of the gerundive as involving the prespecification of an a‰xal, H tone high front vowel, returning us to the traditional view that
Patterns of Reduplication in Yoruba
345
the gerundive involves a tonally and vocalically specified vowel along with a reduplicated consonant (Marantz 1982). This analysis is inconsistent with Alderete et al.’s (1999) specific claim that the Yoruba gerundive involves fixed default segmentism, although it is consistent with their overall proposal that fixed segmentism may result either from the emergence of the unmarked or from a‰xation. Note also that this analysis is incompatible with views of underspecification that would force the lexical absence of unmarked or default values (e.g., Kiparsky 1982d; Pulleyblank 1986), while it is consistent with a proposal such as combinatorial specification (Archangeli and Pulleyblank 1994). 14.5.3
Underapplication
The gerundive exhibits interesting and complex patterns with regard to nasalization. First, the vowel of the gerundive is oral whether the base vowel is oral or nasal. Second, the reduplicated consonant of the gerundive is oral if the base consonant is oral, nasal if the base consonant is a nasal stop, and oral if the base consonant is a nasalized continuant. The first two cases constitute patterns of accurate matching between the reduplicant and the base, while the third type is a nasal-oral mismatch. Third, the gerundive exhibits a pattern of underapplication with respect to a process of liquid nasalization: a rule that nasalizes a lateral generally before high front vowels is blocked from applying before the high front vowel of the gerundive (Clements and Sonaiya 1990). In this section, background regarding the analysis of nasality in ˙ Yoruba is presented, followed by a discussion of the reduplicative data. Background on Nasality Standard Yoruba exhibits seven oral vowels: {i, e, , a, , o, u}. Only three, the high and low vowels, exhibit contrastive nasality: {ı˜, a˜, u˜}.29 Both oral and nasal vowels are common, and minimal pairs are frequent.
14.5.3.1
c
(52) Contrastive nasality a. dı´n [dı˜´ ] ‘fry’ b. ku´n [ku˜´] ‘fill’ c. san [sa˜] ‘pay’
(cf. dı´ ‘close up holes (as of a net)’) (cf. ku´ ‘die’)
The prohibition on nasalized midvowels is widely attested in Kwa (Hyman 1972) and will be analyzed here through a markedness scale (Prince and Smolensky 1993) defining the relation of nasality to vowels of particular heights.30 In addition, we will see that it is important to further divide the constraint on high vowels, incorporating a front/back distinction. (53) Prohibition on nasalized vowels *Mid/Nas *Lo/Nas *HiBk/Nas *HiFr/Nas Midvowels Low vowels High back vowels High front vowels cannot be nasal. cannot be nasal. cannot be nasal. cannot be nasal.
346
Douglas Pulleyblank
To derive the contrastive vowels of Yoruba, MaxNas must be ranked below *Mid/Nas and above the three other constraints:31 (54) Nasal faithfulness and ranking a. MAXNAS Every nasal specification in the input has a correspondent in the output. b. *Mid/Nas MaxNas *Lo/Nas *HiBk/Nas *HiFr/Nas Nasality is realized on all sonorants (vowels and consonants) in any syllable in which it is found. The data in (55a) show nasality a¤ecting a sonorant onset. (55b) shows both that obstruents are una¤ected by a following nasal vowel (already seen in (52)) and that a syllable preceding a nasal vowel is una¤ected. (55c–e) continue to make the point that nasal harmony is syllable-bounded; even in a string of sonorants, nasality does not a¤ect preceding or following syllables. (55) Syllable-bounded nasal harmony a. rı`n [r˜`˜ı ] ‘to walk’ ‘to dispense’ yu´n [y˜u˜´] ´ ´ b. orogun [orogu˜] *[or˜o˜gu˜] ‘cowife’ a`gu`ta`n [a`gu`ta`˜ ] *[agu˜ta˜] ‘sheep’ c. `ıbı´nu´ [ı`bı´nu´˜ ] *[ibı˜nu˜] ‘anger’ a`wo`ra´n [a`wo`r˜a˜´] *[aw ‘picture’ ˜ o˜r˜a˜] d. o`yı`nbo´ [o`y˜˜`ı bo´] *[oy˜ı˜bo˜] ‘European’ ´ ra´ntı´ [ra˜tı´] *[ra˜tı˜] ‘remember’ e. `ıda´nwo` [ı`da´˜ wo`] *[ida˜w ‘examination’ ˜ o˜] o`mo`we´ [ `m ˜` we´] *[ `m ˜` w ‘educated person’ ˜ e˜] ˙ ˙ This pattern of nasal harmony can be captured by the high ranking of constraints requiring the alignment of nasality with syllable edges: c c
c c
(56) Nasal harmony a. ALIGNRIGHT[NASAL] Align[Nasal, Right; s, Right] b. ALIGNLEFT[NASAL] Align[Nasal, Left; s, Left] These constraints interact crucially with MaxNasal. The ideal syllable in Yoruba must contain segments that agree in nasality, both segments oral or both nasal. Where an onset consonant is not nasalizable (e.g., ku˜´ ‘fill’ (53b)), nasality is retained on the vowel at the expense of an AlignLeft[Nasal] violation. That is, MaxNasal must outrank AlignLeft[Nasal]. In contrast, syllables of the type *ne, *no, and so forth are unattested: when the vowel is not nasalizable, the onset consonant cannot itself be nasal. This gap is accounted for by requiring that AlignRight[Nasal] outrank MaxNasal. Hence an input like /ne/ would be optimally realized as [le]
Patterns of Reduplication in Yoruba
347
(violating MaxNasal) rather than [ne] (violating AlignRight[Nasal]). Hence the following ranking is established: (57) Retention/loss of nasality in harmonically imperfect syllables AlignRight[Nasal] MaxNas AlignLeft[Nasal] Lateral Nasalization Given this preliminary background, we can now consider patterns directly involving lateral nasalization. As discussed in Clements and Sonaiya 1990 and elsewhere, the full range of vowels is attested with an obstruent ˙ onset (59a), the full range with the liquid /r/ (59b), but a restricted range with the liquid /l/ (59c). Specifically, there is a gap with /l/: the sequence li is not found (except in restricted contexts to be discussed below).
14.5.3.2
(58) Lateral nasalization a. kı´ ke´ ke´ ka˙` ko ko´˙ ku´ kı´n kan ku´n
‘greet’ ‘shout’ ‘pet’ ‘read’ ‘write’ ‘take’ ‘die’ [kı˜´ ] ‘scrub’ [ka˜] ‘be sour’ [ku˜´] ‘be full’
b. rı´ ‘see’ re´ ‘pare o¤ ’ re ‘soak’ ra˙` ‘buy’ ro` ‘be soft’ ro`˙ ‘think’ ru´ ‘sprout’ rı`n [r˜`˜ı ] ‘walk’ ra`n [r˜a˜`] ‘help’ run [r˜u˜] ‘perish’
c. *li le´ le la˙´ lo lo`˙ lu` lin [nı˜ ] la´n [na˜] lu`n [nu˜]
GAP ‘drive away’ ‘be lazy’ ‘lick’ ‘go’ ‘use’ ‘hit’ ‘worry someone’ ‘spend’ ‘become lost’
Adapting the analysis proposed by Clements and Sonaiya, this gap is accounted ˙ for by constraints whose combined e¤ect is to cause a lateral to be nasal when the syllable containing it can be harmonically nasal. That is, a lateral will nasalize when the following vowel can also nasalize. Crucial to the account is the phonologization of an observation made in Piggott (1992, 48): ‘‘A spontaneously voiced segment contains a nasal phase, if it is also characterized by complete oral occlusion.’’ Since a sonorant requires an open vocal tract, it follows that if there is a closure in the oral cavity ([continuant]) then the nasal cavity must be open.32 (59) NASALPHASE If [þsonorant, continuant] then [þnasal]. The relative ranking of NasalPhase and DepNasal (DepNasal: Every nasal specification in the output has a correspondent in the input) will determine whether a language tolerates oral sonorant stops. Unconstrained by anything else, where NasalPhase outranks DepNasal, [l] is impossible; where DepNasal outranks NasalPhase, [l] is tolerated. For Yoruba, the basic ranking is one where NasalPhase outranks DepNasal, but there are two additional conditions. First, the syllable containing the derived nasal must be nasalized throughout. That is, NasalPhase has an e¤ect only when it is also possible to satisfy AlignRight[Nasal]. Second, the
348
Douglas Pulleyblank
impact of NasalPhase is restricted by the markedness constraints on nasalized vowels, as follows. (60) *MidNas *LoNas *HiBkNas NasalPhase *HiFrNas Since all but *HiFrNas are ranked above NasalPhase, the introduction of nasality will only be possible when a lateral is followed by a high front vowel. For an input like /li/, the optimal output nasalizes the consonant and extends the nasality throughout the syllable. (61) Appearance of nasality where not in input /li/
AlignR [Nas]
Max Nas
Nasal Phase
a. [li]
*!
b. [lı˜ ]
*!
c. [ni]
*!
Dep Nas
*HiFr Nas
*
*
*
+ d. [nı˜ ]
*
*
AlignR[Nasal] is violated if the syllable contains a nasal-oral sequence (61c) and NasalPhase is violated if the syllable contains a lateral (61a,b). The only relevant markedness constraint on nasal vowels is *HiFr/Nas, which is ranked too low to have any e¤ect on the outcome. Note that NasalPhase will have no e¤ect on a syllable containing a consonant other than /l/ since only /l/ is both a sonorant and a stop. For example, NasalPhase does not exert pressure on syllables containing /k/ (58a) or /r/ to nasalize (58b). When /l/ precedes any other vowel (high back, mid, or low), the appropriate markedness constraint will rule out nasalization. To exemplify this, consider a case involving a midvowel. As discussed above, mid–nasalized vowels are impossible, prohibited by ranking *Mid/Nas above MaxNas (see (54)). (62) Maintenance of orality due to a midvowel /le/
*Mid Nas
AlignR [Nas]
+ a. [le] b. [le˜]
Nasal Phase
Dep Nas
* *!
c. [ne] d. [ne˜]
Max Nas
* *!
*!
* * *
Patterns of Reduplication in Yoruba
349
Nasalizing the midvowel is ungrammatical (62b,d); nasalizing the sonorant consonant without harmony is ungrammatical (62c). The optimal output is therefore fully oral: [le] (62a). Nasalized high back vowels and nasalized low vowels are possible in Yoruba, but they cannot be derived by the pressure of NasalPhase. MaxNas exerts no pressure to retain nasality when there is no lexically specified nasal value, and the constraints against nasality on a vowel (*HiBk/Nas & *Lo/Nas) prevent the introduction of nasality since they outrank NasalPhase. Note that [nu˜] and [na˜] are both fully grammatical surface forms, but only for inputs where nasality is specified—that is, /lu˜/ and /la˜/. 14.5.3.3 Orality of the Reduplicant Vowel As expected given the proposed analysis of the gerundive, the prefix vowel is oral, even if the base vowel is nasal.
(63) Orality of reduplicant a. sı´n [sı˜´ ] sı´sı´n b. gba´n [gba˜´] gbı´gba´n c. du`n [du`˜ ] dı´du`n
[sı´sı´˜ ] [gbı´gba˜´] [dı´du`˜ ]
‘sneeze’ ‘scoop out’ ‘be sweet’
This follows from the analysis developed here, in which the vowel of the gerundive prefix is underlying, rather than copied from the base. The optimal candidate simply maintains the gerundive vowel. As we will see in the next section, however, the situation with the consonant of the gerundive is more variable. Variable Nasality of the Reduplicant Consonant A stem-initial oral consonant reduplicates faithfully, but the situation is di¤erent for stem-initial nasals. Nasal stops reduplicate as nasal (64a), but, as noted in Clements and Sonaiya 1990, nasal ˙ continuants reduplicate as oral (64b).33
14.5.3.4
(64) Variable nasality of the consonant of the reduplicant a. na` [na˜`] nı´na` [nı˜´na˜`] ‘flog’ mı´mo` [mı˜´m ˜` ] ‘know, understand’ mo` [m ˜` ] ˙ ˙ ´ ´ b. wı´n [w wı´wı´n [wı´w ‘borrow’ ˜ ˜ı ] ˜ ˜ı ] yı´ya´n [yı´y˜ ´˜ ] ‘yawn’ ya´n [y˜ ´˜ ] ra´ntı´ [r˜a´˜ tı´] rı´ra´ntı´ [rı´r˜a˜´tı´] ‘remember’ c
c
c
c
Although both nasal stops and nasalized continuants are derived in Yoruba, the two types of segments behave in di¤erent ways. Clements and Sonaiya approach ˙ this problem by assigning the rule that derives [n, m] to a lexical stratum, while the rule that derives nasalized continuants is assigned to the postlexical stratum. We may treat the di¤erence between the two types of nasal consonants as emergence of the unmarked—assuming, quite plausibly, that nasal stops are less marked
350
Douglas Pulleyblank
than nasalized sonorant continuants. Within a theory postulating reduplicationspecific faithfulness constraints, the pattern can be derived by ranking a constraint prohibiting nasalized continuants below input-output faithfulness and above basereduplicant faithfulness. In the approach taken here, without reduplication-specific faithfulness, the same e¤ect is a by-product of the fact that NasalPhase constitutes pressure for sonorant stops to be nasal while it does not constitute pressure for a similar nasalization of continuants: (65) Nonretention of nasality on sonorant continuant consonant
*!
+ c. [yı´-y˜a˜´]
* *
d. [yı˜´-y˜a˜´]
f. [y˜˜´ı-y˜a´˜ ]
*[HiFr Nas
*
b. [yı´-ya´˜ ]
e. [y˜´ı-y˜a´˜ ]
Integ
*!
DepIO Seg, Nas
a. [ı´-ya´˜ ]
Nasal Phase
AlignL[Nas]
Max Nas
AlignR[Nas]
´ *[ V /ı´-ya´˜ /
*! *!
*
*
* *
*!
Candidates involving tautosyllabic nasal-oral or oral-nasal sequences are excluded by AlignR[Nasal] (65e) or AlignL[Nasal] (65b,d). The ultimate factor therefore is whether some constraint forces a violation of the constraint prohibiting nasalized high front vowels. With a sonorant continuant, neither NasalPhase nor any other constraint applies such pressure. The result is consequently that the *HiFr/Nas constraint rules out a candidate with the marked property of nasality in the reduplicant (65f ). The less marked candidate with an oral reduplicant wins (65c).34 In marked contrast with such cases, forms with a sonorant stop induce nasality in the reduplicant. Consider the case of nı´na` ‘flog’, where the relevant assessment will be identical whether the input verb is considered to contain /n/ (na˜`) or /l/ (la`˜ ). (For concreteness, the input with a lateral is assumed.)
Patterns of Reduplication in Yoruba
351
(66) Retention of nasal consonant
*
*
*
c. [lı´-la`]
*! *!
*
+ e. [nı˜´-na`˜ ] f. [lı˜´-na`˜ ] g. [lı´-na`˜ ]
*[HiFr Nas
*!
*!
b. [lı´-la˜`]
d. [nı´-na`˜ ]
Integ
Nasal Phase
a. [ı´-na˜`]
DepIO Seg, Nas
AlignL[Nas]
Max Nas
AlignR[Nas]
´ *[ V
*!
/ı´-la`˜ /
*!
*
*
*!
**
*
*!
*
Retention of nasality on the low vowel is ensured by the high ranking of MaxNas (66c); AlignL[Nasal] and NasalPhase ensure that the nasality so retained occurs on both the vowel and the consonant of the base (66b). That is, the base must be [na`˜ ], as it would be in isolation. This leaves four prefix possibilities: nasal-oral, nasal-nasal, oral-nasal, and oral-oral—the last four candidates of (66). Of the four, *[nı´-na`˜ ] (66d) is ruled out by the high-ranking constraint, AlignR[Nasal], which prohibits an oral-nasal sequence within a syllable. Both [lı´˜-na`˜ ] (66f ) and [lı´-na˜`] (66g) are nonoptimal because they violate NasalPhase; [lı˜´-na˜`] also violates AlignL[Nasal]. The optimal candidate is therefore the fully nasal nı˜´na˜` (66e). Underapplication of Lateral Nasalization Of particular interest are the gerundives of lateral-initial bases where the vowel is oral but not [i].
14.5.3.5
(67) Underapplication a. le´ lı´le´ b. le lı´le ˙ ˙ c. la´ lı´la´ d. lo lı´lo ˙ ˙ e. lo` lı´lo` f. lu` lı´lu`
of lateral nasalization ‘drive away’ ‘be lazy’ ‘lick’ ‘go’ ‘use’ ‘hit’
Since the base contains a vowel resistant to nasality, we correctly expect that such forms should be immune to NasalPhase. Perhaps surprisingly, however, we observe
352
Douglas Pulleyblank
that the gerundive forms are similarly immune, in spite of the fact that the reduplicant contains a sequence of a lateral followed by a high front vowel. Lateral assimilation underapplies in such cases. Within an approach positing reduplication-specific constraints, such forms could be derived by requiring that base-reduplicant correspondence outrank NasalPhase: MaxBR NasalPhase. However, BR correspondence is not necessary to the analysis of such cases. Consider the analysis of such forms according to the proposals outlined here. A tableau for lı´lo` is used for illustration. (68) Underapplication of lateral nasalization
*!
*
+ b. [lı´-lo`] c. [lı´-no`]
* *!
d. [lı´-no˜`]
*!
e. [nı˜´-lo`] f. [nı˜´-no`] g. [nı˜´-no`˜ ]
Integ
a. [ı´-lo`]
DepIO Seg, Nas
Nasal Phase
AlignL[Nas]
Max Nas
*Mid Nas
AlignR[Nas]
´ *[ V /ı´-lo`/
*! *!
*
*
*
*
*
*
*
*
*!
*
*
*
*
*
*MidNas rules out forms involving a nasalized midvowel (68g), (68d). As always, a base is ruled out if it contains a nasal-oral sequence, eliminated by AlignR[Nasal] (68c,f ). This leaves two plausible candidates: [lı´-lo`] (68b) and *[nı˜´-lo`] (68e). Crucially, NasalPhase is violated in both and therefore serves no selective function. With the e¤ect of NasalPhase neutralized, the optimal form becomes the one best respecting faithfulness: lı´lo`, the form that maintains the orality of the input. One might ask, however, whether the optimal form lı´lo` should be assessed two violations of NasalPhase. The assumption made here is that feature specifications in correspondence with each other are assessed as a unit for the determination of constraint violations. For a correspondence set to satisfy a constraint, all members of the set must satisfy the constraint. The consonants in correspondence in [ni ´˜ı-ni o`] satisfy NasalPhase because both instances of ni satisfy NasalPhase; the correspondence
Patterns of Reduplication in Yoruba
353
pairs in candidates like [li ´ı-li o`], [li´ı-ni o`], and [ni ´˜ı-li o`] all violate NasalPhase because it is not the case that the two segments in correspondence both satisfy NasalPhase. In conclusion, the gerundive illustrates an important general point about the proposed treatment of reduplication. The analysis of nasality in reduplication follows straightforwardly from the treatment of nasality motivated generally. That is, the complexities of nasality under reduplication fall out from general properties of the grammar. In particular, the underapplication of nasalization in the gerundive provide no compelling evidence for reduplication-specific correspondence constraints. Input-output correspondence constraints are adequate for the evaluation of both nonreduplicative and reduplicative forms. 14.6
Conclusion
This chapter has examined a range of reduplication processes in Yoruba. As noted in the introduction, the four cases considered constitute only a subset of the large class of reduplicative patterns observed in Yoruba. Nevertheless, they serve to exemplify a range of behavior from total copy to fairly minimal copying in conjunction with lexical prespecification. Two points of particular importance emerge from the analysis of these forms. At a theoretical level, it has been demonstrated that complex reduplicative patterns taken to require ‘‘base-reduplicant’’ and ‘‘input-reduplicant’’ correspondence can be accounted for without the postulation of constraints that formally refer to the abstract category ‘‘reduplicant.’’ Returning to the sort of prosodic characterization of reduplicants proposed in Marantz 1982, it has been shown that partial reduplication can be accounted for without reduplication-specific constraints, but also without the negative theoretical consequences of base-reduplicant approaches to templatic reduplication. In particular, the problem of ‘‘base erosion’’ does not arise when there are no constraints forcing faithfulness of base and reduplicant. The proposed theory postulates the possibility of two fundamentally di¤erent types of reduplication, echoing an earlier observation by Kiparsky (1986b). ‘‘Compounding’’ reduplication involves the imposition of morphological identity (Hyman et al., this volume; Inkelas and Zoll 2000), while ‘‘a‰xing’’ reduplication involves minimal violations of Integrity. At an analytic level, the importance of considering reduplication within the broad context of a language’s phonology has been argued to be crucial for an adequate analysis. The Yoruba gerundive, for example, can be analyzed in the abstract as involving either a prespecified /i/ or as involving the emergence of an unmarked default [i]. Consideration of general properties of the phonology of Yoruba in conjunction with consideration of additional reduplicative types demonstrates the superiority of the former hypothesis over the latter.
354
Douglas Pulleyblank
Notes Thanks to Oladiipo Ajiboye, Akinbiyi Akinlabi, Rose-Marie De´chaine, Laura Downing, ˙ Masaru Kiyota, Uyi Stewart, and Suzanne Urbanczyk for discussion of variEun-Sook Kim, ous aspects of this chapter, as well as to an anonymous reviewer. Special thanks to Sharon Inkelas and Kristin Hanson for their extensive comments, both substantive and editorial, on an earlier version of the chapter. 1. Yoruba has numerous additional reduplicative constructions not discussed here—for example, adverbial reduplication (da´rada´ra ‘very well’, dı´e`dı´e` ‘slightly’, etc.); ‘‘all X’’ reduplication ˙ (me´je`e`jı` ‘both’, me´te`e`ta ‘all three’, etc.); ‘‘groups of˙ so-and-so many’’ reduplication (me´jı`me´jı` ˙ ˙ ˙ ‘two by two’, me´tame´ta ‘three by three’, etc.); inherent reduplication (du´du´ ‘black’, gbogbo ˙ ˙ form of banking’, agogo ‘bell’, etc.); ideophonic reduplication (mo`no`‘every’, e`su´su´ ‘traditional ˙` ‘of ˙ mo`no` ‘of repeated flashing’, ro`go`do`ro`go`do` ‘of several items big and round’, fe`re`ge`de`fe`re`ge`de ˙ ˙ ˙ ˙ ˙ ˙ ˙ ˙ ˙ ˙ several things being wide’, ke´e´ke`e`ke´e´ ‘in small bits’, ko´gbe´ko`gbe`ko´gbe´ ‘of clumsy, slow move˙ ˙ ˙ 1958; ˙ ˙ Bamgbos ˙ ment’, la´n´ko´nla`n`ko`nla´n´ko´n ‘of limping’, etc.). See Abraham e 1967; Ogunbo˙ ˙ ˙ ˙ ˙ wale 1970; Awoyale 1974; Awobuluyi 1978; Akinlabi 1984; and so on. 2. Transcriptions are in standard Yoruba orthography throughout, except where specially indicated. In Yoruba orthography e ¼ [], o ¼ [ ], Vn ¼ nasalized vowel, s ¼ [S], p ¼ [kp], ˙ ¼ H, ¼ L, or unmarked for tone˙ ¼ M, a˙ tone-marked nasal ¼ syllabic nasal. Certain entirely predictable tonal processes are not indicated or addressed. For example, a lexical H tone systematically appears as rising after an L tone, and an L tone appears as falling after an H tone. c
3. See Rosenthall 1997 and Orie and Pulleyblank 2002 for further discussion of vowel deletion, a detailed account of which is orthogonal to the concerns of this chapter. 4. There are foot-based restrictions on NoHiatus that are peripheral to the concerns of this chapter. See Pulleyblank in press for discussion. 5. Throughout this chapter, I assume unless otherwise indicated that when two segments are in correspondence they are identical. Patterns of coalescence and so on play no role in the cases considered. The assumed identity could be the result of undominated Ident or by featuresensitive Max and Dep (McCarthy and Prince 1995). 6. For at least some dialects/idiolects, there is evidence for a shape delimiter on agentive reduplication. Contrast the data presented in this section—all of which involve a disyllabic base— with longer examples (Pulleyblank and Akinlabi 1988): so`dodoso`dodo ‘truthful person’ ˙ evil’), ˙ ya´nila´soya´nila´so ‘cloth (< so`dodo ‘be truthful’), se`ba`je´se`ba`je´ ‘evil-doer’ (< se`ba`je´ ‘do ˙ ˙˙ ˙cloth’), fe´nilo´mo˙fe´nilo˙´ mo ‘someone who ˙ ˙ takes ˙ ˙people’s lender’ (< ya´nila´so ‘lend ˙ someone ˙ ˙ ˙ ˙ ˙ ˙ ˙ ˙ daughters and marries them’ (< fe´nilo´mo ‘marry someone’s child’). For some speakers, such ˙ reduplicated forms are acceptable;˙ for˙ others, they are not. Speakers who do not accept such long reduplicated forms derive the relevant agentives by simple prefixation: aso`dodo, ase`ba`je´, ˙ ˙ re˙ aya´nila´so, afe´nilo´mo. For such speakers, agentive reduplication exhibits a phonological ˙ ˙ ˙ ˙ ˙ quirement that each half of the reduplicated form must constitute a minimal word—that is, a disyllabic foot (Ola 1995). To block the reduplication of forms that are more than two syllables long, rather ˙than truncate them, the minimal word requirement for agentive reduplication and input-output faithfulness would have to outrank the requirement of morphological parsing (Prince and Smolensky 1993).
Patterns of Reduplication in Yoruba
355
7. For concreteness, I assume that Red is a prefix, although a su‰xal analysis would also be possible. Possible motivation for a prefixal analysis would be that Yoruba morphology is uniformly prefixing; see below. 8. While there appears to be variability as to whether a minimal word requirement is observed as a contingent condition for agentive reduplication, there seems to be no such variability with kı´-reduplication. 9. Principal tonal alternations are discussed in the next section. One alternation not discussed is the realization of an input HLL sequence as ML in forms like `ılo`kulo` ‘bad use’ (<ı`lo`þkı´þı`lo`). See Akinlabi 1984 for discussion. 10. Note that there are no asymmetries between mid and low vowels in Yoruba, and there are no asymmetries between front and back vowels within the set of nonhigh vowels. The former e¤ect is achieved by not invoking any constraint on the feature low in Yoruba; the latter is achieved by collapsing MaxNonHiBk and MaxNonHiFr (labeled MaxNonHi). Note that an analysis in terms of HNuc constraints (Prince and Smolensky 1993) would be adequate for kı´-reduplication, as in (19), but inadequate for distributive reduplication as in (31); see Pulleyblank in press. 11. These two reduplicative types potentially correspond to two morphosyntactic types, as pointed out to me by Rose-Marie De´chaine (personal communication). In the syntactic literature (Chomsky 1995, etc.), some instances of ‘‘doubled’’ constituents involve the coindexation of two independently inserted forms (perhaps analogous to morphological identity?) while others have been argued to derive from movement (perhaps analogous to phonological correspondence?). I leave this possibility for future investigation. 12. There do not appear to be distributive forms derived from consonant-initial bases. 13. See Ola 1995 for evidence that prosodic constraints on feet hold of the reduplicative ˙ constituent. 14. Vowel-initial forms of the type that can undergo distributive reduplication cannot have an initial H tone. 15. The proposal made here is rather di¤erent from Struijke’s proposal since Struijke maintains base-reduplicant correspondence in addition to broad input-output faithfulness. 16. See also Downing 2000 and Rose 2000 for the use of Integrity in reduplication. 17. It is possible that a distinction between the two representations would be made by AnchorL, which requires that ‘‘the segment at the left edge of a morpheme in the input has a correspondent at the left edge of the morpheme in the output.’’ Assuming that the morphology identifies the initial string oju as the ‘‘distributive’’ and the following oju´mo´ as the root for ˙ ‘day’, then the initial o of the input morpheme in (27b) is straightforwardly identified as having a correspondent at the left edge of oju´mo´ in the output. The same identification is perhaps less ˙ direct correspondents. Nevertheless, if AnchorL is clear in (27a) because the initial o has two interpreted as existential quantification over the output (Ex[(Input A x) & (x ¼ leftmost in morpheme) ! by[(Output A y) & ( y ¼ leftmost in morpheme)]]), then AnchorL is satisfied in both and consequently does not distinguish between the two possibilities. 18. Archangeli and Pulleyblank (1994) argue that all phonological representations must respect linear precedence relations—that such relations are inviolable because they constitute absolute well-formedness conditions. In the cases under consideration involving correspondence
356
Douglas Pulleyblank
relations, it should be noted that each individual representation by itself is well formed, hence respects the condition proposed by Archangeli and Pulleyblank. It is only in the correspondence of one well-formed representation with another that discrepancies between precedence relations are observed (and tolerated, depending on the grammar). 19. It is beyond the scope of this chapter to provide a general discussion of reduplicative patterns involving overapplication and underapplication of phonological rules (though see section 14.5.3.5). Such cases can be accounted for by assuming a general constraint (PathIntegrity) requiring identity between correspondent paths. If path integrity is ranked below both a relevant markedness constraint (C) and a relevant faithfulness constraint (F), then regular application results: C F PI. If the markedness constraint and path integrity outrank faithfulness, overapplication results: C, PI F. Underapplication is similar to overapplication, but Integrity (I) with respect to the potentially overapplying feature prevents its copying, and high ranking path integrity prevents regular application: I, PI C F. 20. I argue here that the two mechanisms must be kept distinct and that both are required. This does not eliminate the possibility of invoking both simultaneously; see Hyman et al. (this volume) and Inkelas and Zoll 2000. 21. In vowel contact contexts, there is deletion of the less sonorous vowel in some cases (gbe´ ina´ @ gbe´na´ ‘lift the lamp’) and assimilation of the less sonorous vowel in others (eru` ˙ igi @ eru` ugi ‘bundle of wood’). For an analysis of the prosodic distinction between the two types ˙of cases, see Orie and Pulleyblank 2002. 22. The e¤ect here is comparable to Struijke’s (1998) notion of broad faithfulness. 23. Note that if this analysis as emergence of the unmarked is correct, then it provides evidence for viewing left-edge preferences as the result of a constraint such as alignment (Akinlabi 1996; Pulleyblank 1996; and so on), not as the result of a ‘‘positional faithfulness’’ constraint (e.g., Beckman 1995). Since the initial vowel in the reduplicant is not underlyingly specified for tone, there is no tone to be ‘‘positionally faithful’’ to. Pressure to spread the base tone would be induced by alignment, but not by positional faithfulness. 24. Verbs are normally consonant-initial in Yoruba (see Ola 1995). For gerundives of the mar˙ 25. ginal vowel-initial class resulting from borrowings, see note 25. There are certain morphologically defined contexts where word-initial consonant epenthesis is observed (Ola 1995). One such context involves gerundive reduplication and potentially ˙ of backcopying. Ola (2000, personal communication) describes a class of constitutes a case ` ba´a`mu` vowel-initial verbs borrowed from ˙English including o´gı`lı` ‘be ugly’, e´n´fı` ‘envy’, e`m ˙ ‘embalm’. Such cases form gerunds by reduplication in conjunction with˙ h-epenthesis:˙ hı´ho´gı`lı`, ˙ It ` ba´a`mu`. As can be seen, an [h] occurs both in the reduplicant and in the base. hı´he´n´fı`, hı´he`m ˙ ˙ ´ , with subsequent backcopyis possible that an [h] is motivated in the reduplicant to satisfy *[V ing into the base. Ola (1995) notes, however, that the isolated verb forms may optionally surface with an initial˙ [h]. It is therefore quite plausible that there is no backcopying, rather that the consonant-initial form constitutes the base for reduplication. If the base is consonantinitial, then the reduplicated forms require no special treatment whatsoever. To prevent the possibility of forming a gerundive from the marked class of vowel-initial forms, parsing the ´ , and DepIO gerundive could be analyzed as having the contingent property that Onset, *[V must be satisfied. 26. Not included in this list are clitics (the only ‘‘su‰xal’’ forms in the language; see section 14.3.4), infixes (section 14.3), and reduplicative prefixes (section 14.4).
Patterns of Reduplication in Yoruba
357
27. The ‘‘progressive’’ a‰x is the sole exception to this pattern. It is a syllabic nasal on an H ´ þba´), n´ta` [n´ta`] ‘selling’ (
30. For additional discussion of the necessity for distinguishing between three vowel heights with respect to nasality in Yoruba, see Ajı´bo´ye` 2002. 31. Nothing in the analysis presented here depends crucially on analyzing the feature of nasality as monovalent Nasal or as binary [Gnasal]. 32. As noted above, nothing hinges on whether these features are unary or binary. NasalPhase could be restated equivalently as ‘‘If [Sonorant, Stop] then [Nasal].’’ 33. Ola´diı´po` Ajı´bo´ye` (personal communication) notes that for the Mo`ba` dialect, the initial ˙ sonorants of˙ verbs like ya´n [y˜a˜´] ‘yawn’, yu´n [y˜u˜´] ‘scratch/go’, and hu´n [h˜˙u˜´] ‘give’ are nasalized: yı´nya´n [y˜˜´ıy˜a˜´] ‘yawning’, yı´nyu´n [y˜˜´ıy˜u˜´] ‘scratching/going’, and hı´nhu´n [h˜˜´ı h˜u˜´] ‘giving’. This may be due to more general transfer of nasality as seen in a form with an obstruent onset: dı´n [dı´˜ ] ‘fry’ versus dı´ndı´n [dı´˜dı˜´ ] ‘frying’. A detailed description of nasalization in the Mo`ba` dialect, along with its analysis and implications, is not undertaken here. See Ajı´bo´ye` 2002. ˙ 34. For the relevant candidates in (65), I assume that there is a single nasal specification in the prefix, multiply linked in a case like (65f ). As such, candidates like (65c) and (65f ) are comparable with regard to Integrity.
15
Multiple Tenses in the Malayalam Verb
Tara Mohanan and K. P. Mohanan
15.1
Introduction
Malayalam verb morphology exhibits a somewhat intriguing phenomenon: a single verb form in a clause often contains multiple tense markings. Closely tied to this are two other phenomena. One is that of morphological restrictions on concatenation that result in a mismatch between purely morphological features, on the one hand, and syntactic and semantic features, on the other. The other is that of ‘‘periphrastic paradigms’’ yielding what look like morphological idioms in which grammatical markers in a sequence jointly express a single paradigmatic choice or meaning that is not transparently inferable from the meanings of the individual markers. As far as we know, this configuration of phenomena has not been subjected to theoretical analysis in the literature. The example in (1) illustrates the phenomena while o¤ering a glimpse into the rich complexity of Malayalam verb morphology and its potential interest for linguistic theory.1 (1) kuTTi ooTikkoNTirikkaNamaayirunnu. child should have continued to run ‘The child should have kept running.’ The morphological breakdown of the verb in (1) can be given as (2). (2) ooTikkoNTirikkaNamaayirunnu2 oot-i-kon-t-irikk-anam-aa-i-irun-nu3 run-pa-KOL-pa-IRIKK-mod-AA-pa-IRIKK-pa run - in the process of - should have been Despite its formidable complexity, evidence from various phonological sources points to the assumption that the string in (2) is indeed a single word. First, such verb complexes in Malayalam form single phonological words with a single stress and word melody, thereby behaving like subcompounds (Mohanan 1986). Second,
360
Tara Mohanan and K. P. Mohanan
word-internal phonological processes such as the gemination of initial stops apply internal to these forms. Third, they do not permit internal pauses in normal speech. The verb in (2) contains four occurrences of past tense marking (-i/-tu) in addition to a modal.4 Examples like (2) involving multiple tense marking are part of a robust paradigm, are entirely productive, and form an integral part of the language.5 The tense system of Malayalam has a three-way contrast of past, present, and future. However, in verb forms with multiple tense markings, the full range of choices in the paradigm is available only to the tense marker at the right edge. Thus, the nonfinal past tense a‰xes in (2) cannot be replaced by present or future tense. The modal can be replaced by the future tense marking but not past or present. The occurrence of nonfinal tense a‰xes is the result of purely morphological requirements. As we will see, nonfinal tense markers do not necessarily contribute toward tense meanings; they are not associated with finiteness or with the meanings carried by finite tense. Hence, the morphological requirements that demand the presence (or absence) of particular tense a‰xes result in a mismatch between morphological features, on the one hand, and syntacticosemantic features, on the other. As for the third phenomenon, what we call morphological idioms are fixed sequences of morphological parts. These parts cannot be changed, omitted, or reordered. They jointly carry a single nontransparent meaning that cannot be inferred from the meanings of the individual morphological parts. Despite semantic opacity, however, the sequences cannot be treated as morphologically atomic: the emphatic marker ee, for instance, can intervene between the parts in the sequence. This emphatic marker cannot, however, occur within a stem.6 In this chapter, we try to provide an account of the phenomena mentioned above by describing the rudiments of a grammar that can adequately account for verb forms like (2) and looking at what kinds of assumptions in a theory of morphology are best suited to yield the description. Given that a single paradigmatic choice is expressed by a sequence of morphological pieces in Malayalam verbs, we may regard them as instances of periphrastic paradigms in the sense of Ackerman and Webelhuth (1998), Sadler and Spencer (2001), Ackerman and Stump (2004), and Spencer (2003), illustrating ‘‘morphological periphrasis’’ as opposed to ‘‘syntactic periphrasis’’ (Spencer 2002). As Arono¤ (1994, 8) points out, morpheme-based theories of morphology raise the default expectation of ‘‘a one-to-one match between a stretch of sound and a meaning.’’ Where languages fail to fulfill this expectation, theories postulate zero morphs (Lieber 1980, 1992; Kiparsky 1982d, 1986b; Marantz 1982), rules that change structure (Anderson 1992), and the like. In contrast, lexeme-based theories (Matthews 1972; Arono¤ 1994; Beard 1995; among others) recognize the separation of word-
Multiple Tenses in Malayalam Verb
361
internal representations motivated by morphophonological considerations from those motivated by morphosyntactic and morphosemantic considerations. The need to factor apart two such levels of representation internal to the structure of words as well as above the word level is familiar in morphology and syntax from parallel correspondence theories, including Lexical-Functional Grammar (LFG) (Bresnan 1982, 2001; Dalrymple et al. 1995; Bo¨rjars, Vincent, and Chapman 1997; Nordlinger 1998; among others). These two levels in LFG are c(onstituent)-structure and f(unctional)-structure. Word internal c-structure is a relatively concrete morphologically motivated level that encodes overt order, multiple tense marking, and complex markers. The terminal elements of this level are drawn from a language-specific list of morphemes specified for their grammatical properties. Even though languages vary in c-structure representations, LFG assumes that their units of representation are drawn from a universal inventory. F-structure is a relatively abstract syntacticosemantically motivated level that expresses language-independent grammatical information. C-structure and f-structure are linked to each other through general principles of correspondence, allowing for mismatches between them. In the language of the Principles and Parameters model, ranging from GovernmentBinding (GB) to Minimalism, the substance of c-structure and f-structure is translated as phonetic form (PF) and logical form (LF) representations, respectively (the latter subsuming syntactic information), and the correspondences between them as movement. The principles of correspondence then translate as constraints on movement/traces. Thus, the interface between PF and LF in Minimalism parallels the interface between c-structure and f-structure in LFG. Just as above the word level the separation of c-structure and f-structure eliminates the need for empty categories and movement rules, word internally it eliminates the need for zero morphs and rules of structure change such as insertion and deletion. Given such a dual structure model, an adequate morphological theory must address the issue of the invariance, variability, and limits of variation in the crosslinguistic correspondences between the levels by postulating a set of substantive universal correspondence principles. Recent work in LFG (Choi 1996; Sells 1997, 1998, 2001; Bresnan 2000a, 2000b), which incorporates Optimality Theory (Prince and Smolensky 1993), assumes that the constraints on f-structure and on pairings between fstructure and c-structure are universal and that language-specific variability results from di¤erences in the ranking of these universal constraints. The word-internal c-structure that emerges as part of our account of Malayalam verb morphology suggests that even though the correspondence constraints that hold on c-structure-fstructure pairs may be statable in Universal Grammar, the well-formedness of word-internal c-structure itself may not be deducible from the combination of universal constraints and language-particular ranking but must allow for language-specific local constraints.
362
Tara Mohanan and K. P. Mohanan
The chapter takes the following path. Section 15.2 lays out the facts of tense, modals, and negation in Malayalam and shows how they exhibit the phenomenon of mismatch. In section 15.3, after a quick look at some relevant aspects of the auxiliary system including multiple tense marking, we go on to the facts of complex verb sequences that we claim are morphological idioms. Section 15.4 spells out the conclusions and the theoretical implications of the analysis. 15.2
The Verbal Stem and Mismatches
The negative su‰x -illa and the relativizing su‰x -a in Malayalam can be attached to the past and present tense forms of verbs but not to the future tense form. The future negative is expressed by the form containing just the bare verb and -illa, while the future relative clause is expressed by the present tense verb form and -a. A di¤erent type of morphological restriction is imposed by modals on their hosts. Thus, the modal -aam ‘may/promise to’ requires that its host be tenseless, while the modal -eene ‘would have’ requires it to be in the past tense form, even though it does not carry the past tense meaning. In this section, we show that, given such arbitrary morphological requirements and gaps, the correspondences between c-structure and f-structure follow from the general principles of paradigm completion. 15.2.1
The Simple Verb Stem
The simplest finite verb form in Malayalam that can act as an independent word consists of a single stem morpheme and one of the three tense a‰xes (past, present, future). (3) a. ooTi run-pa ‘ran’
b. ooTunnu run-pr ‘runs’
c. ooTum run-fu ‘will run’
Instead of a tense a‰x, the verb may take a nonfinite a‰x. (4) a. ooTaan run-nf1 ‘to run’
b. ooTuka run-nf2 ‘running’
c. ooTaaRrun-nf3 ‘about to run’
The su‰xes illustrated in (3) and (4) stand in paradigmatic relation to one another, forming a natural distributional class; we will refer to them as t-a‰xes, where t is a function whose value can be specified as [þtense] or [tense]. The su‰xes in (3) are finite t-a‰xes (t, [þtense]); those in (4) are nonfinite (t, [tense]). We can characterize the structure of the verb in (3) and (4) with the rule in (5).7 (5) v ! stem (t-affix)
Multiple Tenses in Malayalam Verb
15.2.2
363
Tense and Negation
Malayalam has a verb illa, the negative counterpart of the copula uNT ‘have/be’. The form illa also functions as a negative verbal su‰x as illustrated in (6). e
(6) a. ooTiyilla run-pa-neg ‘didn’t run’
b. ooTunnilla run-pr-neg ‘isn’t running’
c. *ooTumilla run-fu-neg Intended: ‘won’t run’
d. ooTilla run-neg ‘won’t run’
While illa can attach to a verb with the past and present tense su‰xes, as in (6a and 6b), it cannot attach to a verb with the future tense su‰x -um, as shown by (6c): negative future is expressed by concatenating illa with a bare verb, as in (6d). The generalization illustrated by (6) is not idiosyncratic to the verb ooT or even a subset of verbs; it holds for all verbs in Malayalam.8 The unacceptability of (6c) is a morphological gap in the paradigm.9 The fact that the bare form in (6d) is interpreted as expressing the future tense is an instance of mismatch between morphology and syntax.10 Under normal circumstances, futurity is marked by -um, as in (3c); in the special circumstance of negation, it is marked by the una‰xed form, despite there being a future tense su‰x in the language. To account for the morphological facts in (3), (4), and (6), we begin with the following specifications for the tense a‰xes (t-af). (7) Inherent specification a. i: t-af [þtense b. tu: t-af [þtense c. unnu: t-af [þtense d. um: t-af [þtense
[þpast, fut]] [þpast, fut]] [past, fut]] [past, þfut]]
We also revise (5) as (8), adding a statement to incorporate the negation su‰x into the c-structure of the verb. In order to distinguish the di¤erent hierarchical levels for attaching morphological units, we follow Selkirk (1982) in extending X-bar notation to word-internal structure. (8) a. v 0 ! v1 (neg) b. v1 ! v2 (t-af) x 0 is a lexical category that functions as a terminal element in a phrase structure representation. The levels below it, namely, x1 , x2 , and so on, appear in morphological constructions but do not have the status of words. The lexical specifications in (7), together with the statements in (8), yield the cstructure representations illustrated in (9) for the forms in (3) and (6). For convenience of representation, we label the feature complexes in (7a and 7b), (7c), and (7d) as {past}, {pres}, and {fut}, respectively.
364
Tara Mohanan and K. P. Mohanan
(9) a.
b.
c.
d.
e.
There are at least three possible ways of thinking about why (9d) is ill formed and what sanctions the interpretation of (9e) as future. In a derivational model of morphology (Anderson 1992, for instance, or Halle and Marantz 1993) on par with a derivational model of phonology in which phonetic representations are derived from phonological representations by rules that change structure, we might think of ooTilla as being the ‘‘surface’’ (concrete) form derived from an ‘‘underlying’’ (abstract) form ootþumþilla through the deletion of the future su‰x: ootþumþilla ! ootþqþilla. In a nonderivational model of the kind assumed in Lieber (1980), Kiparsky (1982d, 1993a), Marantz (1982), and others, we would assume that the future morpheme has a zero allomorph that the negative su‰x is constrained to select. Both these accounts assume three morphemes in ooTilla. In contrast to both, the ac-
Multiple Tenses in Malayalam Verb
365
count assumed in (9) postulates two morphemes corresponding to the two overt morphs. (10)
A Feature Morpheme Morph
oot oot oot # oot/
B
C
fu {fut} um
neg {neg} illa
oot oot
fu {fut}
neg {neg}
oot oot
q
/illa/
/oot/
q
/illa/
/oot/
fu
neg {neg}
/illa/
Model C, in contrast to A and B, allows an abstract feature that has no corresponding concrete morphemic representation, analogous to a grammatical function without a corresponding phrase structure node. The motivation for C will become apparent at a later point. Within model C, the unacceptability of (9d) can be made to follow from a morphological restriction on sister selection. Take, for instance, the su‰x -ity in English: that it attaches to an adjective and yields a noun can be stated as the requirement that its sister must be an adj and its mother an n. This strategy can be adopted to rule out (9d). (11) -illa {neg}:
a. mother ¼ [þtense] b. sister ¼ [fut]
The specification in (11b) yields the idiosyncratic morphological gap of (9d). Given (11a and 11b), the c-structure representations in (9c) and (9e) can be expanded as (12a) and (12b), respectively, incorporating the features specified in (11). (12) also provides the corresponding features in f-structure. As mentioned in the introduction, despite significant empirical di¤erences, the levels of c-structure and f-structure are conceptually the approximate equivalents of the modules of PF and syntax, respectively, in Principles and Parameters frameworks. (12)
c-structure a.
f-structure 2 3 pred: ‘run’ ! 4 tense: past 5 neg
366
Tara Mohanan and K. P. Mohanan
b.
2
3 pred: ‘run’ ! 4 tense: fut 5 neg
There is an unmarked feature-to-feature correspondence between f-structure and c-structure in (12a) but not in (12b). The unmarked pairings (which would hold on (9a and 9b) as well) can be viewed as being governed by the principle in (13a), and the marked pairing in (12b) by that in (13b).11 (13) a. Every c-structure feature complex must be associated with a corresponding f-structure feature complex, and vice versa. b. An unassociated f-structure carrying a paradigmatic contrast is paired with an available well-formed unassociated c-structure. We assume that a paradigm involves a pairing of a set of f-structure contrasts paired with the corresponding c-structure realizations. The unmarked state of a paradigm is one in which each of the f-structure options is paired with one and only one c-structure realization. Gaps in the paradigm occur when the f-structure options outnumber available c-structure realizations. (13a) captures the biuniqueness of unmarked c-structure-f-structure correspondences. (13b) expresses as an unmarked principle the tendency of grammars to avoid paradigmatic ambiguity. The morphological restriction on illa in (11b) permits three well-formed cstructures, namely, v[þtense [þpast, fut], neg]; v[þtense [past, fut], neg]; and v[þtense [fut], neg]. The unmarked pairing with the f-structures v-pa-neg and v-pr-neg makes the first two constructions unavailable for pairing with v-fu-neg. By (13b), then, this f-structure is paired with the remaining c-structure, v[þtense [fut], neg]. The explanation given above for the facts in (3) and (6) raises a further question. Given the morphological restriction in (11b), there are two other logically possible scenarios. One is that the future negative is simply inexpressible in the language. The other is that the verb with the future tense su‰x, but without the negative su‰x, expresses negative future, that is, given the co-occurrence restriction *{fut}{neg}, {fut} rather than {neg} is absent in the verb form. Why does the morphological restriction not lead to either of these scenarios? This question can be answered by postulating the following principles.
Multiple Tenses in Malayalam Verb
367
(14) a. All elements of a paradigm must be expressed. b. Paradigmatic contrasts must be preserved. (14a) is a prohibition against defective paradigms, that is, those in which one or more elements are inexpressible. (14b) avoids ambiguity resulting from the neutralization of paradigmatic contrasts. Thus, (14a) requires that the future negative of a verb be expressible, and (14b) prevents the neutralization of the future negative with the future a‰rmative.12 The principle in (14b) is an instance of the more general principle of maximizing and preserving the contrasts in a system, applicable not only to morphology but also to phonology and to other areas of language structure (Zipf 1949; Martinet 1952; Mohanan 1993; Flemming 1995; among others). In fact, it may turn out to be a general principle of cognitive systems (Ramachandran and Hirstein 1999). Note that while the prohibition against the co-occurrence of {fut} and {neg} in cstructure in (11b) is an idiosyncratic language-specific condition, the correspondence principles in (13) and (14) are universal. The principles in (13) and (14), which express the unmarked association between pairs of representations, have a long history. Chomsky (1964) identified linearity, biuniqueness, invariance, and local determinacy as the conditions that hold in the pairing of phonetic and phonemic representations in classical phonemics and argued against classical phonemic representations by demonstrating that these conditions are untenable. Postal’s (1968, 56) naturalness condition, which holds that underlying and phonetic representations should be as close as possible, expresses the essence of the classical phonemic conditions as a markedness condition. The same idea underlies faithfulness constraints in Optimality Theory (OT) (McCarthy and Prince 1995). For instance, linearity-io requires precedence relations in the input and the output to be idential (i.e., no metathesis), while max-io and dep-io require input segments to have corresponding output segments and vice versa (i.e., no deletion, no epenthesis). The association conventions in Autosegmental Phonology (Leben 1973; Goldsmith 1976) also belong to the general class of conditions that express the unmarked relation between paired representations, as do the principle of function-argument biuniqueness in LFG (Bresnan 1980), the projection principle and bijection principle in GB (Chomsky 1981), Kiparsky’s (1968a) prohibition against absolute neutralization, and so on. It would be desirable to formulate a set of abstract conditions that apply to any two pairs of representations rather than to stipulate unrelated sets of conditions on each pair separately. Such an attempt is made, for instance, in the generalized correspondence constraints in OT, thus marking the beginning of a possible theory of unmarked correspondences. The generalized version of Max relates any two strings, S1 and S2 , and requires every element of S1 to have a correspondent in S2 ; the same holds for Dep and Ident (Kager 1999, 205).
368
Tara Mohanan and K. P. Mohanan
In addition to conditions on the pairing of representations, we need another broad class, namely, conditions that apply to the inventories of paired levels of representation (e.g., Kiparsky’s 1968a prohibition against absolute neutralization). We also need to distinguish between levels of representation that carry unmarked correspondences between units in pairs of inventories (e.g., phonetic and phonological levels, fstructure and c-structure, or s-structure and pf) and those that do not (e.g., tone and tone-bearing units, or lf and pf).13 For lack of space, however, we reserve further discussion of the issue to another occasion. 15.2.3
Tense and the Modification Suffix
Another example of morphological gaps and mismatches is found in the relative clause construction, which is marked by the modification su‰x (ms) -a.14 (15) a. weegam ooTiya kuTTi fast run-pa-ms child ‘the child who ran fast’ b. weegam ooTunna kuTTi fast run-pr-ms child ‘the child who runs/will run fast’ The modification su‰x -a, like the negative su‰x -illa, cannot co-occur with the future tense su‰x -um, as shown by (16a). Note, however, that the two constructions di¤er in how this gap is filled. Unlike the future negative, the future relative clause has no distinct form: the present tense verb form expresses both present and future (as in (15b)). (16) a. *ooTuma run-fu-ms
b. *ooTa run-ms
The generalization illustrated in (15) and (16) is representative of all verbs in Malayalam. The unacceptability of (16a) illustrates the phenomenon of morphological gaps, and the interpretation of the present tense su‰x as expressing future in (15b) illustrates the phenomenon of mismatch. To disallow (16a) and allow the ambiguity in (15b) without allowing it in (3b) or (6b), we stipulate a restriction on the su‰x -a as in (17). (17) -a ‘COMP’:
a. mother ¼ [þtense] b. sister ¼ [þtense, fut]
As with negation, the specification on the sister in (17) involves an idiosyncratic gap: it disallows verbal hosts with future marking, thus disallowing (16a). However, unlike the negative, given that the sister of the modification su‰x must be [þtense], the infinitival form is unavailable. This yields the unacceptability of (16b).
Multiple Tenses in Malayalam Verb
369
The c-structure and f-structure features of the relevant forms in (15) are given in (18). c-structure a.
b.
f-structure pred: ‘run’ ! tense: past
!
!
(18)
pred: ‘run’ tense: pres pred: ‘run’ tense: fut
Note that (18b) involves a new type of mismatch. In the case of negation, as discussed above, the future in f-structure has no corresponding c-structure tense feature. In (18b), on the other hand, the future in f-structure does have a corresponding cstructure tense feature, but it is [past, fut], rather than the expected [past, þfut]. As the sister of the modification su‰x must be tensed, there is no unassociated cstructure that v-fu-ms can be paired with. Hence (13b) is inapplicable. By (14a) the f-structure v-fu-ms must necessarily be expressed; it is therefore paired with an associated c-structure, namely, v-{pres}-{-a}. This pairing results in a neutralization of tense in the pairing of c-structure and f-structure (18b), violating (14b). This result would follow if we assumed that (14a) is stronger (or ranked higher) than (14b). Given the restriction in (17b), there are two morphological forms available to express a three-way paradigmatic contrast, namely, the past and the present. fu is associated with {pres} rather than {past} because, given the feature complexes of the three tenses (shown in (7)), the closest to {fut} is {pres}: they di¤er in only one feature value. The relevant principle for this choice can be stated as (19). (19) A nondefault pairing of representations must be as close to the default pairing as possible (measuring closeness in terms of the identity of feature values). The result of combining relative clauses and negation exhibits a further morphological gap. The morph that negates a relative clause is -aat, as illustrated in (20).15
370
Tara Mohanan and K. P. Mohanan
(20) weegam ooTaatta kuTTi (oot-aat-a) fast run-neg-ms child ‘the child who didn’t run/doesn’t run/won’t run/isn’t running fast’ While illa permits the host to contain past or present tense (as in (6a and 6b)), the a‰x -aat- cannot co-occur with any of the tense a‰xes, as illustrated in (21). (21) a. *ooTiyaatta oot-i-aat-a run-pa-neg-ms
b. *ooTunnaatta oot-unn-aatt-a run-pr-neg-ms
c. *ooTumaatta oot-um-aatt-a run-fu-neg-ms
The morphological gap illustrated in (21) and the mismatch and ambiguity illustrated in (20) can be expressed by the following specification. (22) -aat {neg}:
sister ¼ [tense]
To understand how the specifications on -aat in (22) and -a in (17) yield the right results, let us look at the morphological structure of the verb form in (20). (23)
The requirements of -aat in (22) cause vz in (23) to be [tense]. Nevertheless, by the restriction in (17b) on the sister of -a, vy (vz þaat) must be [þtense]. Given the specification of the mother in (17a), vx must also be [þtense]. Although vy is [fut] and we therefore expect vx to be [fut] as well, (14a) yields the complete paradigm of past, present, and future, resulting in the three-way ambiguity of (20). Once again, (14b) is violated. 15.2.4
Tense and Modals
We now turn to the system of modals in Malayalam. Modals like -aam, -aNam, and -aTTe are attached to tenseless verbs, as shown in (24)–(26), while the modal -eene is attached to a verbal host with the past tense su‰x, as in (27).16 (24) n˜aan ooTaam. cf. *ooTiyaam oot-aam oot-i-aam I run-AAM run-pa-AAM ‘I will (promise to) run.’
*ooTunnaam oot-unnu-aam run-pr-AAM
*ooTumaam oot-um-aam run-fu-AAM
Multiple Tenses in Malayalam Verb
(25) awan ooTaNam. oot-anam he run-ANAM ‘He must run.’ (26) awan ooTaTTe. oot-atte he run-ATTE ‘Let him run.’
371
cf. *ooTiyaNam oot-i-anam run-pa-ANAM cf. *ooTiyaTTe oot-i-atte run-pa-ATTE
(27) awan ooTiyeene. cf. *ooTeene oot-i-eene oot-eene he run-pa-EENE run-EENE ‘He would have run (e.g., tomorrow).’
*ooTunnaNam oot-unnu-anam run-pr-ANAM
*ooTumaNam oot-um-anam run-fu-ANAM
*ooTunnaTTe oot-unnu-atte run-pr-ATTE
*ooTumaTTe oot-um-atte run-fu-ATTE
*ooTunneene oot-unnu-eene run-pr-EENE
*ooTumeene oot-um-eene run-fu-EENE
We can describe the facts illustrated by these examples by assuming the lexical entries in (28) and (29).17 (28) -aam / -aNam / -aTT {modal}:
a. mother ¼ [þfin, tense] b. sister ¼ [þinf ]
(29) -eene {modal}:
a. mother ¼ [þfin, þtense] b. sister ¼ [þpast]
e
Examples (24)–(26) illustrate the default match between the syntactic paradigms in f-structure and the morphological forms in c-structure: the modals concatenate with verbs without tense a‰xes. In contrast, (27) illustrates a mismatch. The probability modal -eene does not require tense in f-structure. Hence, the past tense su‰x serves a purely morphological function, leading to a mismatch between the two levels of representation. Notice that the situation in (27) is the mirror image of that in negation, seen in (6). In the negative future form, there is no tense a‰x, and yet it is interpreted as future. The form with the modal -eene contains a tense su‰x but is interpreted as tenseless. Modals have the same distribution as negatives; thus, a modal and a negative cannot co-occur.18 We will therefore incorporate the position of modals in the morphological structure of verbs by expanding (8) as (30). n neg o (30) a. v 0 ! v1 mod b. v1 ! v2 (t-af) The kinds of morphological requirements on tense marking illustrated so far are found in adverbial complementizers as well. Verbal su‰xes such as -aal ‘if ’, koNT ‘while’, and -iTT ‘subsequent to’ require the past tense verb form (ooT-i-aal ‘if x will run/ran/runs’; ooT-i-kkoNT ‘while running’; ooT-i-iTT ‘after running’), while
e
e
e
e
372
Tara Mohanan and K. P. Mohanan
-pooL ‘when’ requires the future tense form (ooTumpooL ‘when running’). Given the grammar we have developed so far, an analysis of the morphological gaps in the complementizers is straightforward. We will not go into the details in this chapter for want of space.19 15.3
Morphological Idioms and Multiple Tense Marking
The phenomenon of morphological idioms in the Malayalam verbal system cannot be discussed without at least some familiarity with the rudiments of the auxiliary system. We therefore turn now to the auxiliary construction (a main verb followed by auxiliary verbs) in Malayalam. Of all the pieces in the verb morphology of the language, the auxiliaries are the most complicated and recalcitrant to analysis. We will not try to provide a complete picture or a detailed analysis of the auxiliary system. What we attempt to do is outline a few examples of auxiliaries that will help the reader make su‰cient sense of the auxiliary construction within which the notion of morphological idioms can be understood, and then proceed to build an approach to morphological idioms on the basis of the examples. 15.3.1
Simple Auxiliaries
Auxiliary verbs in Malayalam are typically grammaticized lexical verbs. For instance, corresponding to the lexical verbs irikk- ‘sit’, aa- ‘be/become’, uL- ‘be/have’, and koL- ‘fit’ are the auxiliary su‰xes -irikk-, -aa-, -uL-, and -koL- illustrated in (31)– (34). (31) ooTiyirunnu oot-i-ir-ikk-tu run-pa-IRIKK-pa ‘used to run’
/ ooTiyirikkunnu (*ooTumirunnu; *ooTunnirunnu) oot-i-ir-ikk-unnu run-pa-IRIKK-pr ‘has run’
(32) ooTukayaawum oot-uka-aa-um run-nf-AA-fu ‘must be running’
/ ooTukayaayi (*ooTiyaayi; *ooTumaayi) oot-uka-aa-i run-nf-AA-pa ‘is about to run’
(33) ooTunnuNT / ooTunnuNTaayirunnu (*ooTiyuNT ; *ooTumuNT ) oot-unnu-unt oot-unnu-unt-aa-ir-ikk-tu run-pr-UL-pr run-pr-UL-pr-AA-IRIKK-pa ‘is/has been running’ ‘was/had been running’ e
e
e
e
(34) ooTikkoLLum / ooTikkoLLaam (*ooTunnukkoLLum; *ooTumkkoLLum) oot-i-kol-um oot-i-kol-aam run-pa-KOL-fu run-pa-KOL-AAM ‘will (manage to) run’ ‘(I) will (¼ undertake to) run’
Multiple Tenses in Malayalam Verb
373
The unacceptable forms in (31)–(34) illustrate certain co-occurrence restrictions on what can appear to the left of the auxiliaries. These restrictions can be stated as in (35). Recall that t refers to the feature of the verb stem that carries finite or nonfinite a‰xes. (35) a. b. c. d.
-irikk-: -aa-: -uL-: -koL-:
[t, [t, [t, [t,
þpast] tense] past, fut] þpast]
Unlike the conditions imposed by the negative, relative, and modal su‰xes that we saw earlier, those in (35) cannot be stated as conditions on sisterhood in a local configuration. Rather, they must be treated as ‘‘adjacency conditions’’ on what can appear to the left of the auxiliary. The skeletal structure in (36) for the form ooTiyirunnu in (31) exemplifies the point. (36)
15.3.2
Complex Auxiliaries
In addition to simple auxiliaries like those in (35), Malayalam also has what can be described as complex auxiliaries. We begin with one involving the form koL-, encountered already in (34). (37) ooTikkoNTirunnu / ooTikkoNTirikkaNam (oot-i-kol-tu-ir-ikk-tu) (oot-i-kol-tu-ir-ikk-anam) run-pa-KOL-pa-IRIKK-pa run-pa-KOL-pa-IRIKK-mod ‘kept running’ ‘must keep running’ At first sight, the forms in (37) might appear to involve the combination of the two auxiliaries -koL- (35d) and -irikk- (35a). However, the sequences exhibit certain peculiarities that cannot be attributed to their component parts. First, unlike the auxiliaries in (31)–(33), -koLþtense cannot occur as the final element when the tense is present or past.
374
Tara Mohanan and K. P. Mohanan
(38) *ooTikkoNTu / *ooTikkoLLunnu (oot-i-kol-tu) (oot-i-kol-unnu) run-pa-KOL-pa run-pa-KOL-pr Second, -koLþtense cannot be followed by negation or certain modals like -eene. (39) *ooTikkoLLilla / *ooTikkoNTeene (oot-i-kol-illa) (oot-i-kol-tu-eene) run-pa-KOL-neg run-pa-KOL-pa-mod Third, the only stem that can occur to the right of -koLþtense in the verb is -irikk-; replacing -irikk- with other stems results in ungrammatical forms. (40) *ooTikkoLLukayaayi / *ooTikkoLLunnuNT (oot-i-kol-uka-aa-i) (oot-i-kol-unnu-unt ) run-pa-KOL-nf-AA-pa run-pa-KOL-pr-UL-pr e
e
One way to explain these restrictions is to assume that the forms illustrated in (37) involve a complex auxiliary, -koNTirikk- (-kol-tu-ir-ikk) that is listed in the lexicon. The semantic nontransparency of the sequence lends additional support to this assumption. The form -aayirikk-, illustrated in (41), presents a similar example within the auxiliary system. (41) ooTukayaayirunnu / ooTaaraayirunnu oot-uka-aa-i-ir-ikk-tu oot-aar-aa-i-ir-ikk-tu run-nf-AA-pa-IRIKK-pa run-nf-AA-pa-IRIKK-pa ‘was in the process of running’ ‘was about to run’ Unlike the other auxiliaries, -aayirikk- shows a wide range of options with respect to what can occur to its left. In (41), it follows nonfinite a‰xes; in (42) below, it follows the modals -aam and -anam; in (43), the negative su‰x; and in (44), the future tense a‰x. (42) ooTaamaayirunnu / ooTaNamaayirikkaam (oot-aam-aa-i-ir-ikk-tu) (oot-anam-aa-i-ir-ikk-aam) run-mod-AA-pa-IRIKK-pa run-mod-AA-pa-IRIKK-mod ‘could have run’ ‘may want to run’ (43) ooTillaayirunnu (oot-illa-aa-i-ir-ikk-tu) run-neg-AA-pa-IRIKK-pa ‘would have not run’
Multiple Tenses in Malayalam Verb
375
(44) ooTumaayirunnu / *ooTiyaayirunnu / *ooTunnaayirunnu oot-um-aa-i-ir-ikk-tu oot-i-aa-i-ir-ikk-tu oot-unnu-aa-i-ir-ikk-tu run-fu-AA-pa-IRIKK-pa run-pa-AA-pa-IRIKK-pa run-pr-AA-pa-IRIKK-pa ‘would have run’ It is tempting to analyze -aayirikk- into the independent auxiliaries -aa- and -irikk-. However, there is evidence that -aayirikk- is a single auxiliary. For example, in none of the forms in (42)–(44) can either -aa- or -irikk- be omitted, as illustrated in (45). (45) *ooTumaayi / *ooTumirunnu oot-um-aa-i oot-um-ir-ikk-tu run-fu-AA-pa run-fu-IRIKK-pa Furthermore, as specified in (35b), -aa- disallows modals, the negative su‰x, and all [þtense] a‰xes to its left, while -aayirikk- allows them (see (42)–(44)). This di¤erence in behavior is unexplained if we treat -aayirikk- as being composed of the independent auxiliaries -aa- and -irikk-. These patterns suggest that -aayirikk-, though morphologically complex, should be treated as a single complex auxiliary. Based on the above discussion, we assume that -kontirikk- and -aayirikk- are complex auxiliaries and state below the adjacency conditions on them, along the lines of (35).20 (46) a. -koNTirikk-: b. -aayirikk-:
[t, þpast] @ [þtense, fut]
The specification in (46a) requires that -koNTirikk- be immediately preceded within the word by the past su‰x. (46b) specifies that -aayirikk- cannot be preceded by the past and present tense su‰xes (44). If these specifications are on the right track, we must accept that c-structure feature requirements can be imposed not only by monomorphemic forms but also by polymorphemic stems. In acknowledging complex auxiliaries as lexical entries, we are allowing units smaller than a word but larger than a morph or morpheme to be listed as nontransparent units at a higher level. We will refer to such units as morphological idioms. Let us now try to incorporate auxiliaries into the word-internal c-structure of verbs. The structure of (47) given in (48) illustrates our assumptions on auxiliary stems. (47) ooTikoNTirikkukayaayirikkaNam ooT-i-koL-tu-ir-ikk-uka-aa-i-ir-ikk-aNam run-pa-KOL-pa-IRIKK-nf-AA-pa-IRIKK-mod ‘must be in the process of continuing to run’
376
Tara Mohanan and K. P. Mohanan
(48)
In order to sanction the structure in (48), we must expand the formation rules postulated in (30) as (49). 91 08 < neg = (49) a. v 0 ! v1 @ mod A : 0 ; 2 v
(t-af) v b. v1 ! v1 v1 The rules in (49), as they are formulated, create the illusion that they appeal to a set of categories derived from a universal inventory, namely, v 0 , v1 , and v2 . Although v 0 expresses the universal substance of ‘‘verb at the word level,’’ the units v1 and v2 below the word level do not have any universal substance. Rather, their substance appears to be language specific. Consequently, the rules in (49) also appear to lack universality. This raises an interesting set of three possibilities for consideration. One is that the rules in (49) are idiosyncratic to Malayalam and are not derivable from any universal statement. This is clearly an undesirable scenario. The second possibility is that the e¤ects of (49) are deducible from a combination of universal constraints and language-specific rankings of these constraints, a scenario consistent with the central claims of OT. The third possibility is that while there is considerable universal substance in (49) (or some appropriate reformulation of it) that can be expressed in an adequate theory, the language-particular and -universal ingredients cannot be factored apart in terms of constraints and their ranking, respectively. While the second possibility is the most desirable one, our hunch is that we will have to settle for the third one. We leave the problem for future research.
Multiple Tenses in Malayalam Verb
15.3.3
377
The Complex Auxiliary aayirikk-
The morpheme aa illustrated in (50) is ambiguous between ‘be’ and ‘become’. (50) ani neetaaw aakunnu. (aa-unnu) Ani leader be/become-pr ‘Ani is/becomes a leader.’ e
In each of the examples in (51) below, the context suppresses one of the two meanings, yielding an unambiguous interpretation, whether in the present or future tense. (51) a. ani toottattil aakunnu / aakum. Ani garden-in be-pr be-fu ‘Ani is/will be in the garden.’ b. witt maram aakunnu / aakum. seed tree become-pr become-fu ‘A seed becomes/will become a tree.’ e
In the past tense, however, the two meanings of aa are manifested by distinct forms, as shown by (52). (52) a. ani neetaaw aayirunnu. (aa-i-ir-ikk-tu) Ani leader be-pa ‘Ani was a leader.’ b. ani neetaaw aayi. (aa-i) Ani leader become-pa ‘Ani became a leader.’ e e
The verbs in (52a) and (52b) are the past tense forms of aa ‘be’ and aa ‘become’, respectively. The contrast between the forms is highlighted by the asymmetries in (53). (53) a. ani toottattil aayirunnu / *aayi. Ani-n garden-l be-pa become-pa ‘Ani was in the garden.’ b. witt maram aayi / * aayirunnu. seed-n tree-n become-pa be-pa ‘The seed became a tree.’ e
Thus, in juxtaposing (50) and (52), we find that the complex form aa-i-irikk-tu is the past tense of aa ‘be’, while aa-i is the past tense of aa ‘become’. Taking the asymmetry in (53) as our cue, suppose we assume two distinct verb stems, aa- and aayirikk-. The examples in (50)–(53) demonstrate that aaþpast unambiguously means ‘be’, while aaþpres and aaþfut are ambiguous between ‘be’ and ‘become’. We have also seen that aayirikkþpast is unambiguously ‘become’. (54) shows that aayirikkþfut is also unambiguous, but it can only mean ‘be’.
378
Tara Mohanan and K. P. Mohanan
(54) a.
ani toottattil aayirikkum. Ani garden-in be-fu ‘Ani will be in the garden.’ b. *witt maram aayirikkum. seed tree become-fu e
Example (55) presents an additional surprise: aayirikkþpres means ‘has become’, and appears to be the present perfect form of aa- ‘become’. (55) a. *ani toottattil aayirikkunnu. Ani garden-in be-pr b. witt maram aayirikkunnu. seed tree become-pr ‘The seed has become a tree.’ e
The generalizations illustrated by (50)–(55) are summarized in the following table.
aa-
A‰x
Meaning
-i/-tu (past)
become
-um
be/become
(future)
-unnu (present) -i/-tu (past) aayirikk-
-um
be
(future)
-unnu (present)
has become
The configuration of form-meaning correspondences in this table suggests that the forms aayirunnu, aayirikkunnu, and aayirikkum are morphological idioms whose meanings cannot be inferred from the meanings of their morphological components.21 The table also points to a di¤erent dimension of mismatch. While the morphologically complex form aayirunnu corresponds to the state ‘was’ in semantics, the morphologically simpler form aayi corresponds to the semantically complex change of state ‘became’. 15.3.4
The Notion of Morphological Idiom
We can now spell out the concept of morphological idiom as a construction with the following properties: (i) It is formed out of a sequence of grammatical morphemes. (ii) The sequence is fixed; its components can be neither rearranged nor substituted by equivalent components. (iii) Its meaning cannot be entirely derived from the
Multiple Tenses in Malayalam Verb
379
meanings of the component parts. (iv) The sequence forms a single lexical category. Properties (i) and (iii) render a morphological idiom part of a periphrastic paradigm (a` la Ackerman and Stump (2004)), while properties (ii) and (iv) make the periphrasis word internal rather than phrasal. The English a‰x sequences -istic and -ical in mechanistic and phonological may be thought of as displaying the rudiments of morphological idiomaticity. What we find in Malayalam seems to be a full-blown version of this phenomenon. Morphological idiomaticity extends beyond what we have called ‘‘complex auxiliaries.’’ The meaning of -um, for instance, is that of futurity (see (3c)). That of aayirunnu is describable as ‘‘a state in the past’’ (see (53a)). This is further shown by (56). (56) ani ooTukayaayirunnu. Ani-n run-nf-AA-pa-IRIKK-pa ‘Ani was (in the state of ) running.’ However, when -um and aayirunnu are put together in that order, they jointly express the meaning of a possible but not actual event, without specific time reference. (57) ani ooTumaayirunnu. Ani run-fu-AA-pa-IRIKK-pa a. ‘Ani would have run (but did/will not).’ b. ‘Ani used to run (but not any more).’ The implied negation in (57) is not deducible from the meanings of the component parts. The semantic noncompositionality of (57) stands out even more starkly when compared with its future tense counterpart in (58). (58) ani ooTumaayirikkum. Ani run-fu-AA-pa-IRIKK-fu ‘Ani might run.’ The meaning of futurity in (58) comes from the future tense su‰x at the end of the form. However, the meaning of likelihood and the absence of negation that result from the replacement of the past tense su‰x in (57) with the future tense su‰x in (58) cannot be inferred compositionally.22 15.4
Concluding Remarks
Our exploration of the morphological structure of verbs in Malayalam began with the idiosyncratic restrictions that verbal a‰xes impose on the tense marking on their hosts and proceeded to unearth three related phenomena, namely, multiple tense markings, mismatches between morphological and syntactic tense, and morphological idiomaticity to which we return, in that order, here.
380
15.4.1
Tara Mohanan and K. P. Mohanan
C-structure, Morphological Restrictions, and Multiple Tense Markings
Multiple tense markings are a result of the general phenomenon of morphological units (stems as well as a‰xes) imposing purely morphological restrictions on units adjacent to them, without any syntactic or semantic motivation. This finding poses a challenge to those varieties of functionalism (for instance, Langacker’s (1987) Cognitive Linguistics) that hold that all structural regularities are motivated by considerations of either meaning or pronunciation. The morphological restrictions we have encountered fall into three broad classes: (i) language-specific structural regularities imposed by the word-internal formation rules of c-structure (49); (ii) sisterhood restrictions imposed by a‰xes on their hosts, such as those stated in (11), (17), (22), (28), and (29); and (iii) adjacency restrictions imposed by auxiliaries on units that cannot be analyzed as their sisters (35) and (46). These three types of regularities jointly define well-formed c-structures internal to a verb. To the best of our knowledge, these word internal c-structure regularities of Malayalam verb morphology are not governed by phonological, prosodic, or syntactic constraints that can be stated in Universal Grammar. It is likely that a subset of what we see here as morphological conditions on morpheme combinations will turn out to be motivated by semantic considerations. However, we doubt that all of the word-internal c-structure of verbs can be made to follow from universal constraints. The analysis of Malayalam verb morphology that we have provided relies on a representation of overt morphological structure in terms of morphs. This is inconsistent with theories of morphology that deny the validity of the notion morph, for instance, Word-based Morphology (Ford and Singh 1983, 1985) or A-morphous Morphology (Anderson 1992). It is hard to see how these theories could capture the substance of our analysis. 15.4.2
Dual Structures and Mismatches
The central claim supported by this study is that of the need for what may be called dual representations in word-internal structure: one motivated by morphophonological considerations and the other by morphosyntactic and morphosemantic considerations. These two dimensions of representations correspond to c-structure and f-structure in LFG, and to PF and ‘‘syntax’’ in Minimalism. The need for the separation of the two sets of considerations has been present in approaches like the Word-and-Paradigm model (Matthews 1972), A-morphous Morphology (Anderson 1992), Lexeme-based Morphology (Arono¤ 1994), Separationist Morphology (Beard 1988, 1995), and so on. But rather than separating the two sets of considerations in terms of the dual parallel representation model in this chapter, these approaches follow the derivational ‘‘Item-Process’’ model of morphology.
Multiple Tenses in Malayalam Verb
381
Though the need for dual structures (c-structure vs. f-structure, PF vs. syntax) has been acknowledged in most syntactic theories, one of its logical conclusions, namely, the extension of dual representations to inflectional features, as far as we know, has not been explored systematically in morphological theory. As an example that motivates the need for such an extension, we have shown that the c-structure feature [fut] associated with the present tense su‰x in Malayalam corresponds to the fstructure feature future in the relative clause construction (18b). To take another example, although the c-structure of a verb may contain multiple tense features, only the verb-final feature has a corresponding f-structure tense feature. Such dual representations of features eliminate the need for null elements (zero morphs) in morphological theory. Dual representations involve the mismatch not only of features but also of compositionality, one of them being a mismatch in the number of c-structure and f-structure units. In (6d), for instance, there are only two c-structure units (ootþilla) that correspond to the three f-structure units: vþfutþneg (12b). The multiple tense markings repeatedly found in the Malayalam verb forms (e.g., (2), (48)) are simply an instance of a mirror image of this mismatch, namely, c-structure units without corresponding f-structure features. 15.4.3
Morphological Idioms
The third phenomenon of theoretical interest in Malayalam verb morphology is that of the idiomaticity of word-internal sequences of grammatical elements. As we have seen, the grammatical elements in the auxiliary system often combine to form complex auxiliaries with relatively inflexible sequences of elements and nontransparency of meanings that are characteristic of phrasal idioms. The idiomaticity extends beyond complex auxiliary stems, even to a sequence of stems and a‰xes that do not form a single constituent (57). In this chapter, we crucially draw on the idea of Dual Representation Morphology to make sense of the morphological idioms in Malayalam. Idiomaticity in inflectional paradigms has aroused considerable interest in the recent literature (Ackerman and Webelhuth 1998; Sadler and Spencer 2001; Spencer 2001, 2002, 2003; and Ackerman and Stump 2004). The idea of periphrastic paradigms expands the notion of ‘‘paradigm’’ from synthetic paradigms such as in go, goes, and went in English to include forms like has gone, had gone, is going, and has been going. Morphological periphrasis as illustrated by Malayalam verb paradigms involves the realization of paradigmatic options across multiple locations within a word, while syntactic periphrasis involves realization across words. The dual representations of language-particular morphological features and universal syntactic features postulated in our analysis of Malayalam verb morphology
382
Tara Mohanan and K. P. Mohanan
are embedded in the multidimensional model of the architecture of grammar of LFG, with at least the following as significant parallel levels of representation: p(honological)-structure, c(onstituent)-structure, and f(unctional)-structure, which subsumes g(rammatical) f(unction)-structure, a(rgument)-structure, and sem(antic)structure. That the morphological idioms in the Malayalam auxiliary system involve a mismatch in the correspondence between form (c-structure) and content (semstructure) is fairly clear. It is also clear that they involve language-particular statements of correspondence. In our analysis, we have taken a relatively conservative approach and adopted the null hypothesis as the default: we have assumed that the mismatch is located in the correspondence between c-structure and the syntax in f-structure and does not extend to sem-structure. We have also assumed for the purposes of this chapter that correspondences between the syntactic and semantic components within f-structure do not exhibit variation across languages. Whether or not these assumptions are tenable remains open to further empirical investigation. Notes The phenomena explored in this chapter have called for our attention for over 20 years, and we hope that we have understood at least some of the intricacies of this complex area. We are extremely grateful to Sharon Inkelas and Kristin Hanson for inviting us to contribute to this volume. We thank Arto Anttila, Chin Seok Koon, and Vivienne Fong for detailed comments, Farrell Ackerman for drawing our attention to the concept of periphrastic paradigms in the recent literature on realizational morphology, and all those who have borne with our discussions of Malayalam verb morphology over the years, including audiences at the Dravidian Linguistics Workshop at SUNY-Stonybrook in 1987, a linguistics colloquium at Stanford in 1988, and the Australian Linguistics Institute in Brisbane in 1998. We are especially indebted to Kristin Hanson and an anonymous reviewer whose comments and suggestions have resulted in a vastly improved chapter. We are honored to have this opportunity to acknowledge the profound influence Paul Kiparsky has had, as guru and mitra, in shaping our life’s paths and our ideas over more than 25 years. 1. The variety of Malayalam described in this chapter is that spoken in central Kerala. There is considerable uniformity across dialects of Malayalam when it comes to morphology. Nevertheless, minor variations in acceptability judgments across dialects, perhaps even among speakers of the same dialect, should not be surprising, particularly in the case of dialects of Malayalam spoken in the southernmost and northernmost regions of Kerala. 2. The following abbreviations are used in the text: fu ¼ Future, mod ¼ Modal, ms ¼ Modification su‰x, nf ¼ Nonfinite, pa ¼ Past, pr ¼ Present, pred ¼ Predicate, adj ¼ Adjective, neg ¼ Negative, pas ¼ Passive, caus ¼ Causative, n ¼ Noun. Forms glossed in upper case italics, namely, KOL, IRIKK, and AA, are grammaticalized verbs. Their nongrammaticalized counterparts have the meanings ‘fit’, ‘sit’, and ‘be/become’, respectively. 3. A legitimate reaction to the apparent oversegmentation in (2) is Is it synchronically valid or is it simply a residue of historical change? Our strategy is to begin with the analysis found in
Multiple Tenses in Malayalam Verb
383
traditional as well as recent Malayalam grammars (Varma 1896; George 1971; Variar 1979; Asher and Kumari 1997), consistent with the intuitions of the language users, and then see how many of the morphological breaks are motivated by empirical considerations. 4. Malayalam has two past tense allomorphs, -i and -tu. The choice between them is idiosyncratic to the individual verb. The allomorph -i has no morphophonological variants; -tu, on the other hand, is variously realized as [tu], [ttu], [nnu], [n˜n˜u], [ndu], and so on, depending on the phonology of the stem. 5. A similar example is cited in Asher and Kumari (1997, 326): e
(i) paThippikkappeTTukoNTirunniTTuNTaayirunna paaTT had been being caused to be learnt - which song-n ‘the song which had been being taught’
The morphological structure of the relevant form in (i) is given in (ii): (ii) path-ipp-ikk-appet-tu-kon-t-irun-n-it-t-unt-aa-irun-n-a learn-vs-caus-pas-pa-KOL-pa-IRIKK-pa-IT-pa-UL-pr-AA-IRIKK-pa-ms ‘caused to be learnt - in the process of having been - which’ vs in this gloss stands for verbalizing su‰x. The segmentation into root and verbal su‰x is ignored in the rest of the paper. While the form in (2) is a verb, that in (i) above is part of a relative clause, made up of a verb and a relativizing su‰x. 6. The interaction of semantics with the morphologically motivated parts of a word is an extremely interesting area of morphosemantics, particularly in morphological idioms, where the meaning of sequences of a‰xes is nontransparent. However, a detailed exploration of this area is beyond the scope of this chapter. 7. A verb stem in Malayalam is often made up of a root and a verbalizing su‰x (rootþvs). The verb stem itself can take causative and passive su‰xes to form a complex verb stem. Since the internal structure of a verb stem does not impinge on the issues central to this chapter, we will not discuss it further. 8. The following example briefly illustrates the use of the copula: (i) kuTTi ootaarunt / ootaarilla. child run-habitual-is / run-habitual-not ‘The child runs (habitually) / doesn’t run (habitually).’ e
In contrast, the su‰x illa in (6) cannot be replaced by unt : *ootiyunt ; *ootunt . In this chapter, we will not be concerned with the illa that acts as the counterpart of uNT . e
e
e
e
9. This gap cannot be explained in terms of, say, a phonological constraint. The coordinating su‰x -um does occur before negation, as in ooTukayumilla (oot-uka-um-illa) ‘will/does not run either’. 10. The situation is similar to the negative paradigm for the English verb be (see Bresnan 2000a for an extensive discussion, and an Optimality Theory (OT) analysis). In English, the first person singular form am cannot take the a‰x -n’t (*amn’t). This gap is filled by aren’t (Aren’t I?). Thus, the form are, which expresses agreement with either plural or second person subject, when combined with n’t, covers first person singular as well, resulting in a mismatch between the morphological form and its syntactic features.
384
Tara Mohanan and K. P. Mohanan
11. For a detailed discussion of correspondence constraints and their expression within the LFG formalism, see Bresnan 2001 and Dalrymple et al. 1995. 12. The account given above can be recast in terms of an OT analysis by the following constraint ranking: *{fut}þ{neg} (¼(11b)) max(neg) max(tense), where max demands the realization of a syntactic feature. We are grateful to Arto Anttila for these observations and the OT formulations. We will not explore in this chapter the question of whether an OT analysis can be extended to all the instances of gaps and mismatches in Malayalam verb morphology. 13. For instance, a phonological inventory includes features like [voice] and [high], traditionally taken to be a proper subset of phonetic features. Unmarked correspondence between the phonetic and phonological levels would imply identical values for the same feature at both levels. No such unmarked correspondences are possible between tones and tone-bearing units. 14. We will represent the c-structure node of this complementizer su‰x as comp in representations. However, as the a‰x also attaches to verbs to yield adjectives, we gloss it in the examples as modification su‰x (ms), following the terminology used in Mohanan and Mohanan (1999). 15. The modification su‰x -a cannot cooccur with illa, the negative su‰x discussed earlier: *ooTiyillaya (oot-i-illa-a: run-pa-neg-ms) and *ooTunnillaya (oot-unnu-illa-a: run-pr-neg-ms). The unacceptability of these forms, we assume, is the result of allomorph selection: illa and -aat- are allomorphs of {neg}. For lack of space, we will not attempt an analysis of the distribution of these allomorphs in this chapter. 16. We use the phonological shapes of the modals in (24)–(27) in place of glosses that appeal to crosslinguistic categories. Accurate glossing would require further exploration of the semantics of the modals, which is beyond the scope of this chapter. 17. This formulation presupposes the assumption that modals are tenseless and finite, while tense morphemes are tensed and finite. We adopt this assumption only for the sake of concreteness; nothing crucial hinges on it. 18. Negation of modals is expressed either by a separate morphological form, as in (i), or by adding a stem to carry the modal, as in (ii), both of which are negations of ooTanam ((25)). As shown by (iii) and (iv), the modal cannot co-occur with the negative su‰xes illa or -aat- on the same stem. (i) ooTaNTa oot-anta run-neg:ANAM ‘must/need not run’ (ii) ooTaatirikkaNam oot-aat-irikk-anam run-neg-IRIKK-ANAM ‘must keep from running’ (iii) *ooTillaNam oot-illa-anam run-neg-ANAM (iv) *ooTaataNam oot-aat-anam run-neg-ANAM
Multiple Tenses in Malayalam Verb
385
19. On the basis of the observation that some constructions allow the su‰xes -i/-tu and -unnu, but not -um, and some constructions require -um (disallowing -i/-tu and -unnu), Babu (1996) argues that -um should be treated as a modal, not a tense marker. Given that there are similar morphological requirements as well as restrictions that involve -i/-tu (e.g., in the complementizer system and the auxiliary system discussed in the next section), we do not find this proposal convincing. In a model of morphology that distinguishes between morphological (c-structure) tense and syntactic (f-structure) tense, the main motivation for Babu’s proposal can be translated as the observation that there are occurrences of -um without corresponding syntactically relevant tense and occurrences of syntactic future without a corresponding -um. 20. The list we provide is far from exhaustive. There are complex auxiliaries in the language that involve nonfinite tense a‰xes, for instance, aar-aa and aar-ul, illustrated below. (i) ani ooTaaRaayi. Ani oot-aar-aa-i Ani-n run-nf-be-pa ‘Ani is about to run.’ (ii) ani ooTaaRuNT . Ani oot-aar-unt Ani-n run-nf-be-pr ‘Ani runs (habitually).’ e
e
Due to lack of space, we will leave an exhaustive list of auxiliaries to another time. 21. Similar nontransparency of semantics is found in the other complex auxiliaries, including koNTirikk. 22. The idea of morphological idiomaticity can be viewed as the word-internal manifestation of the idiomaticity of sequences of grammatical words found, for instance, in John would have come. John has come refers to a present state, resulting from a past event. The addition of would yields the meaning of unrealized potential (as in (57)), not deducible either from would or from has come. See Spencer (2001, 2002, 2003) for a treatment of constructional idioms in Paradigm-based Realizational Morphology.
16
On Pa¯nini 2.4.81 (a¯mah) ˙ ˙
S. D. Joshi and J. A. F. Roodbergen
1. A periphrastic formation using the su‰x a¯m is prescribed by Pa¯nini (P.) 3.1.35 ˙ to form the perfect of secondary verbal bases (causatives, desideratives, etc.) and of some other, simple verbal bases mentioned in the section P. 3.1.35–42. Here P. 3.1.35 introduces the elements a¯m and liti (‘in the perfect’), which are continued ˙ in P. 3.1.3–39. It also contains the exception amantre ‘except in the mantra-portion of the Veda’. P. 3.1.40 prescribes the use of kr- as an auxiliary when lIT follows. ˙ ˙ P. 3.1.41 is a nipa¯tanasu¯tra (‘rule for ready-made irregular forms’). P. 3.1.42 is another nipa¯tanasu¯tra dealing with a¯m and with kr- as an auxiliary used in the aorist ˙ and once in the optative in the language of the Vedas. Possibly the latter two rules were added at a later date (see 5). After P. 3.1.42 the section on the formation of the aorist starts with P. 3.1.43. 2. According to Whitney (1889), sections 1070 and 1073, the periphrastic perfect is almost unknown in the Veda but came gradually into use in the Bra¯hmanas, mainly ˙ in connection with causative bases. This is in agreement with the exception amantre stated by P. 3.1.35. Whitney (ibid.) characterizes the formation as the prefixation of an accusative derivative noun stem in a¯ to the perfect tense of an auxiliary verb, kr˙ or as-, or very rarely bhu¯. In the later language, however, bhu¯- is the usual auxiliary. Whitney (section 1073 c) notes that in the oldest Bra¯hmanas forms with the aorist of ˙ the auxiliary are as numerous as those with the perfect of the auxiliary. In section 1073 e he notes that the noun in a¯m retains its antoda¯tta (‘acute on the final syllable’) accent, whereas, with the exception of a single example from the S´atapathabra¯h˙ mana, the auxiliary is unaccented. Whitney (section 1072 c) also notes that the ˙ connection between the noun and the auxiliary is not so close that other words cannot intervene. The examples given of intervening words are eva, va¯, and the relative pronoun yah. In Whitney’s grammar the form in a¯m and the following auxiliary are ˙ printed as two words. 3. P. 2.4.81 prescribes luk-deletion of the element lI when coming immediately after the su‰x a¯m. Here the form leh is continued from P. 2.4.80. Luk is continued ˙
388
S. D. Joshi and J. A. F. Roodbergen
from P. 2.4.58. LI is the common element in ClI, the aorist marker (P. 3.1.43), and lIT, the marker of the perfect (P. 3.2.115). ˙ 4. The examples quoted by the Ka¯s´ika¯vrtti for the present rule are ¯ıha¯m cakre ‘he ˙ ˙ has striven for’, u¯ha¯m cakre ‘he has inferred’, and ¯ıksa¯m cakre ‘he has looked’. The ˙ ˙ ˙ use of the a¯tmanepada (‘middle’) ending is in accordance with P. 1.3.63. Following tradition, we will for the time being derive and print the forms concerned as two words. The traditional derivation of the first example is as follows: (a) (b) (c) (d)
Dha¯tupa¯tha 1.663 ¯ıhA˙ þ lIT P. 3.2.115 ¯ıh ˙ þ a¯m þ lIT P. 3.1.36 ¯ıh ˙ P. 2.4.81 þ a¯m þ ¯ıh ¯ıha¯m. (e) ¯ıha¯m þ kr P. 3.1.40 ˙ (f ) ¯ıha¯m þ kr þ lIT P. 3.2.115 ˙ ˙ (g) ¯ıha¯m þ kr þ kr þ lIT P. 6.1.8 ˙ ˙ ˙ (h) ¯ıha¯m þ ta P. 3.4.78 (i) ¯ıha¯m þe P. 3.4.81 ( j) ¯ıha¯m þ kar þ kr þ e P. 7.4.66 ˙ (k) ¯ıha¯m þ ka þ kr þ e P. 7.4.60 ˙ (l) ¯ıha¯m þ ca þ kr þ e P. 7.4.62 ˙ (m) ¯ıha¯m þ ca þ kr þ e P. 6.1.77 P. 8.3.23 (n) ¯ıha¯m cakre ¯ıha¯m cakre. ˙ We note that the laka¯ra (‘verbal marker’) lIT is applied twice, once to generate the ˙ form ¯ıha¯m and again to generate the perfect form cakre. It is also clear that the first lIT has to be deleted. Its purpose is merely to allow the addition of a¯m to a verbal ˙ base and to account for the perfect meaning of the form in a¯m. Once that purpose is attained, lIT has to go, and this is what P. 2.4.81 is for. ˙ 5. Whereas the commentary of the Ka¯s´ika¯vrtti on the present rule takes only one ˙ line, that of Patan˜jali takes nearly two pages in the Kielhorn edition of the Ma¯habha¯sya. It includes 13 Va¯rttikas (Vt.). ˙ Vt. I indicates the rule to which it belongs by saying a¯mo ler lope ‘in (the rule dealing with) the deletion of lI immediately following a¯m’. The Vt. then continues ˙ and by saying that in this rule inclusion should be made of (the aorist marker) lUN of (the imperative marker) lOT. Apparently, the Va¯rttikaka¯ra extends the procedure ˙ adopted for the perfect to the aorist and imperative also. The examples quoted by Patan˜jali are ta¯m baijava¯payo vida¯m akran ‘her the Baijava¯pis have known’ and atra ˙
On Pa¯nini 2.4.81 (a¯mah) ˙ ˙
389
bhavanto vida¯m kurvantu ‘your excellency (pl.) must know’. The forms vida¯m akran ˙ (aorist) and vida¯m kurvantu (imperative) are nipa¯tanas (‘ready-made irregular’) forms ˙ mentioned by P. 3.1.41 and 3.1.42, respectively. The derivation of vida¯m akran is as follows: (a) vid(b) vid (c) vid
˙ þ lUN ˙ þ lUN
þ a¯m
Dhp. 2.55 P. 3.2.110 P. 3.1.42
˙ (the aorist marker) should be deleted by P. 2.4.81. To make this At this stage lUN ˙ in the rule. The derivation then continues: possible, the Vt. proposes to include lUN (d) (e) (f ) (g) (h) (i) ( j) (k) (l) (m) (n) (o)
vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m vida¯m akran.
þ
þa þa þa þa þa þa þa þa þa
þ þ þ þ þ þ þ þ þ þ þ
kr˙ kr˙ kr˙ kr˙ kr˙ kr˙ kr˙ kr˙ kr˙ kr˙ kr˙
˙ þ lUN ˙ þ lUN þ jhi þ anti þ ant þ ant þ ant þ ant þ ant þ an
Vt. I on P. 2.4.81 P. 3.1.40 P. 3.1.42 P. 6.4.71 P. 3.4.78 P. 7.1.3 P. 3.4.100 P. 3.1.43 P. 3.1.44 P. 2.4.80 P. 6.1.77 P. 8.2.23
The same derivation would hold for vida¯m kurvantu up to stage (f ): Here lOT (the ˙ imperative) is introduced by P. 3.3.162 and replaced by jhi in stage (g). No augment is added. In stage (h) jhi is replaced by anti. In stage (i) anti is replaced by antu by P. 3.4.86. Then in stage ( j) u is added to kr by P. 3.1.79. In stage (k) guna is applied ˙ ˙ to kr by P. 7.3.84. This gives us vida¯m þ kar þ u þ antu. In stage (l) the a of kar ˙ is replaced by u by P. 6.4.110. Finally, in stage (m) the u added to kr in stage ( j) ˙ is replaced by v by P. 6.1.77. The resulting form is vida¯m kurvantu. ˙ Patan˜jali rejects Vt. I by saying that any su‰x added immediately after a¯m is to be deleted, not only a lI su‰x. Since, however, vida¯m akran and vida¯m kurvantu are ˙ nipa¯tana forms (P. 3.1.41 and 3.1.42), one wonders why a separate statement is needed to justify them. Another question is whether P. 3.1.41 and 3.1.42 were still unknown to Ka¯tya¯yana and Patan˜jali. Vt. II says that a prohibition should be stated against deletion of NaL, the third ˙ and first person sg. perfect ending, after bases ending in a¯m. Patan˜jali provides the examples s´as´a¯ma ‘he has become calm’ and tata¯ma ‘he has su¤ocated’.
390
S. D. Joshi and J. A. F. Roodbergen
The derivation of s´as´a¯ma is as follows: (a) (b) (c) (d) (e) s´am þ (f ) s´a þ (g) s´a þ s´as´a¯ma.
s´ams´ams´ams´ams´ams´ams´am-
þ lIT ˙ þ tiP þ NaL ˙ þa þa þa
Dhp. 4.92 P. 3.2.115 P. 3.4.78 P. 3.4.82 P. 6.1.2 P. 7.4.60 P. 7.2.116
At stage (g) P. 2.4.81 would become applicable because NaL is a substitute for lIT. ˙ ˙ Thus NaL would be deleted. This is not desired. That is why Vt. II was formulated. ˙ Patan˜jali rejects Vt. II by referring to the paribha¯sa¯ (pbh.) arthavadgrahane ˙ ˙ na¯narthakasya grahanam ‘when there is a (possibility of ) reference to a meaningful ˙ element, (then) reference to a nonmeaningful element (is) not (allowed)’ (The Paribha¯sendus´ekhana of Na¯gojibhatta, pbh. 14). The pbh. is, in fact, referred to in Vt. III. ˙ ˙˙ The point is that the meaningful su‰x a¯m used to form the periphrastic perfect is not the same as the meaningless sequence a¯m in s´as´a¯ma. Therefore the mention of a¯m in P. 2.4.81, which is meaningful, does not include the mere phoneme sequence a¯-m, which is meaningless. Vt. III is an iti ced Va¯rttika. It says that, if it is argued that the prohibition against deletion of NaL need not be stated after bases like s´am- on account of the pbh. just ˙ mentioned, then, in any case, we need this prohibition in connection with the verbal base am- ‘to go’. Patan˜jali quotes the example a¯ma ‘he has gone’. Here a¯m is a meaningful sequence. Therefore the pbh. does not apply. Vt. IV is an uktam va¯ Va¯rttika. It says that, alternatively, an answer has already been stated. According to Patan˜jali, reference is to the samnipa¯ta (‘combination’) ˙ pbh., which is quoted in full. See PN, pbh. 85. The pbh. says that a rule occasioned by the combination of two conditioning factors does not become the cause of another rule that destroys the e¤ect of that combination. This can be shown in the derivation of a¯ma, as follows: (a) (b) (c) (d) (e) am þ (f ) am þ (g) a þ
amam am am am a¯m a¯m
þ þ þ þ þ þ
liT ˙ tiP NaL ˙ a a a
Dhp. 1.493 P. 3.2.115 P. 3.4.78 P. 3.4.82 P. 6.1.2 P. 7.2.116 P. 7.4.60
On Pa¯nini 2.4.81 (a¯mah) ˙ ˙
(h) a¯ þ (i) a¯m a¯ma.
a¯m
391
þa þa
P. 7.4.70 P. 6.1.10
Here a¯m is a meaningful response. Therefore the arthavad (‘meaningful’) pbh. does not apply. Consequently, deletion of the ending NaL will be allowed. This is not ˙ desired. To prevent it, reference is made to the samnipa¯ta pbh. The question, there˙ fore, is how does this pbh. prevent the deletion? The point is that the vrddhi applied to the verbal base in stage (f ) has changed am ˙ into a¯m. This vrddhi is conditioned by NaL. The two factors that combine to form ˙ ˙ a¯m are NaL and the ensuing vrddhi. The e¤ect of this combination, namely, a¯m, can˙ ˙ not destroy the ending NaL by luk-deletion. That would amount to upajı¯vyavirodha ˙ ‘the contradiction (of destroying) that upon which one lives’. The discussion then turns to the status of an a¯m-formation like vida¯m. Vt. V says that the luk-deletion prescribed by P. 2.4.81 is an apava¯da ‘exception’ to subsitutions. The substitutions are those prescribed by P. 3.4.78. Therefore, when the question is of applying these substitutions, the luk-deletion will prevail. The consequence is that no finite verb endings can be added to the word ending in a¯m, like vida¯m. But then, what kind of word—if it is a word ( pada) at all—is vida¯m? Vt. VI states an objection to Vt. V. If the luk-deletion prescribed by P. 2.4.81 prevails over the substitutions prescribed by P. 3.4.78, then operations depending on the ˙ -su‰x will not take place. Patan˜jali explains that reference is to the presence of a tiN designation pada, which in the case of verb forms is dependent on the presence of a ˙ su‰x (P. 1.4.14). tiN Vt. VII removes this objection. We can still say that a form like vida¯m is a pada, because it ends in suP ‘a case ending’. But how to explain that vida¯m is an inflected (subanta) word? ˙ su‰x is Vt. VIII says that a laka¯ra (like vid þ a¯m þ liT ) that does not end in a tiN ˙ called krt (by P. 3.1.93). Therefore it is called pra¯tipadika ‘nominal stem’ (by P. ˙ 1.2.46). A case ending can then be added, because the prescription of a case ending is dependent on this latter designation. But then where is the case ending? We do not see it, as Patan˜jali says. Well, we might say that it has been deleted by P. 2.4.82. But this rule applies to avyayas ‘indeclinables’. Then how to argue that vida¯m is an avyaya? The answer is provided by Vt. IX. Vt. IX says that vida¯m is an avyaya because it ends in m. Reference is to P. 1.1.39. Therefore P. 2.4.82 can become applicable for deleting the case ending. So the di‰culty of establishing the word class of vida¯m has been solved. It is an avyaya. Vts. X–XIII deal with accent. By way of introduction to Vt. X, Patan˜jali raises the question of where to put the accent in the expression yat praka¯raya¯m caka¯ra ˙ ‘(the fact) that he has caused to make’ and where not to put it. The expression ya´t
392
S. D. Joshi and J. A. F. Roodbergen
praka¯raya¯m caka¯ra is a subordinate clause introduced by ya´t. Here two rules become ˙ applicable. They are P. 8.1.28, which says that a finite verb is unaccented when immediately preceded by what is not a finite verb, and P. 8.1.30, which says that a finite verb retains its accent when connected with the particle ya´d. Here P. 8.1.30 is the more special rule, so it should prevail. The question is only, which is the tin˙anta (‘finite’) verb meant in the expression praka¯raya¯m caka¯ra? Is it praka¯raya¯m or caka¯ra? ˙ How could praka¯raya¯m be a tin˙anta form at all? That depends on the prakriya¯ (‘derivation’). Two prakriya¯s are possible. Prakriya¯ A goes as follows: (a) (b) (c) (d)
( pra þ) kr ˙ ( pra þ) ka¯r ( pra þ) ka¯r ( pra þ) ka¯r ( pra)ka¯raya¯m.
þ þ þ þ
NiC ˙ ay ay ay
þ þ þ þ
a¯m a¯m a¯m a¯m
þ lIT ˙ þ lIT ˙ þ tiP þ
P. P. P. P.
3.1.26; 3.1.35; 3.2.115 7.2.115; 7.3.84; 6.1.78 3.4.78 2.4.81
Here ( pra) is put within brackets to indicate that it plays no role in the derivation. What matters is that ka¯raya¯m is tin˙anta according to P. 1.1.62. Now the argument continues. Since the finite verb form praka¯raya¯m occurs in a subordinate clause introduced by ya´d, it retains its accent by P. 8.1.30. But which accent? The answer is, antoda¯tta ‘acute on the final syllable’, by P. 3.1.3. This being the case, the particle (gati) pra is unaccented, according to P. 8.1.71. That gives us ya´t praka¯raya¯´m by way of accent. Now we consider prakriya¯ B. If in stage (b) of prakriya¯ A the su‰x lIT is deleted ˙ by P. 2.4.81 before the ending tiP is introduced, then praka¯raya¯m cannot be tin˙anta. In that case, we cannot apply P. 8.1.71. Consequently, pra will retain its accent. This accent is a¯dyuda¯tta ‘acute on the first syllable’ (Phitsu¯tra, 4.13, upasarga¯s´ ca¯bhivar˙ jam). Since ka¯raya¯m also retains its accent (P. 3.1.3), the result is that pra and ka¯raya¯m become separate words having accents of their own. This is not desired. The desired form of the expression quoted is ya´t praka¯raya¯´m, which requires prakriya¯ A. The answer stated by Vt. X says that the accent in the example quoted is assigned on the basis of the original accent of a krdanta stem. Reference is to P. 6.2.139. This ˙ rule presupposed compound formation of a krdanta ‘primary noun’ stem with a ˙ particle (gati) (P. 2.2.18), and it says that in this situation the krdanta stem retains ˙ its original accent. The compound in question is pra þ ka¯raya¯m, formed by P. 2.2.18. Accordingly, ka¯raya¯m, which, as Patan˜jali explains, is the uttarapada ‘final compound-member’, will be antoda¯tta by P. 3.1.3 and P. 6.2.139. Then, since praka¯r˙ , the finite verb form caka¯ra would lose its aya¯m is a krdanta form not ending in tiN ˙ accent by P. 8.1.28. However, this rule is overruled by P. 8.1.30, as stated above. By this latter rule caka¯ra retains its accent. The accent is uda¯tta ‘acute’ of the syllable immediately preceding an ending provided with the anubandha L., like NaL. Thus ˙
On Pa¯nini 2.4.81 (a¯mah) ˙ ˙
393
the proper accentuation of the expression under discussion should be ya´t praka¯raya¯´m ˙ caka¯´ra. Vt. XI says that in this way nigha¯ta ‘loss of accent’ and anigha¯ta ‘non-loss of accent’ can be accounted for. The example quoted by Patan˜jali is caksuska¯mam ˙ ˙ ˙ ya¯jaya¯m caka¯ra ‘he has caused one who was wishing for the faculty of seeing to sac˙ ˙ but caka¯ra, which immediately follows, rifice’. Since ya¯jaya¯m does not end in tiN does, we apply P. 8.1.28. Consequently, we can account for nigha¯ta in the case of caka¯ra. In the sequence caksuska¯mam ya¯jaya¯m both words are atin˙anta ‘not ending ˙ ˙ ˙ ˙ su‰x’. Therefore we in a tiN cannot apply P. 8.1.28. Instead, for ya¯jaya¯m we apply P. 3.1.3, which leads to antoda¯tta accentuation. That accounts for anigha¯ta. Vt. XII notes that there is an undesired possibility of compound formation with the negative ˜ . Reference is to P. 2.2.6. The examples quoted by Patan˜jali are na ka¯rparticle naN aya¯m ‘he has not caused to do’ and na ha¯raya¯m ‘he has not caused to take away’. The solution to this di‰culty is stated by the next Vt. Vt. XIII is again an uktam va¯ Va¯rttika. It says that a relevant statement has al˙ ready been made. According to Patan˜jali, reference is to asa¯marthya ‘lack of semantic connection’. The point is that in the expressions quoted in which caka¯ra is to be supplied, the semantic connection is between na and caka¯ra, and not between na and ka¯raya¯m or ha¯raya¯m. This being so, the question of compound formation, which is conditioned by sa¯marthya ‘semantic connection’ (P. 2.2.1, in the traditional interpretation), does not arise at all. 6. The present rule raises some doubts regarding the grammatical procedure advocated here. Let us return to the example ¯ıha¯m cakre. As shown above, the traditional ˙ derivation requires two main stages, one for ¯ıha¯m and one for cakre. Technically, these two stages can be represented as (i) ¯ıh þ a¯m þ lIT, where lIT is subsequently ˙ ˙ deleted, and (ii) ¯ıha¯m þ kr þ lIT, where lIT appears again. Here (i) is by P. 3.1.35, ˙ ˙ ˙ whereas (ii) is by P. 3.1.40. The question is whether the lIT in (ii) is the same as the ˙ lIT in (i). The traditional derivation denies this. The result is the formation of two ˙ di¤erent words, ¯ıha¯m and a perfect finite verb form like cakre. If there are two words, then basically there are two accents. Against this traditional derivation the following objections can be raised: (1) In the Asta¯dhya¯yı¯ there is no laka¯ra that is not replaced by actual finite verb end˙˙ ings or by a krt substitute. ˙ (2) The form ¯ıha¯m, which apparently conveys a perfect sense—but which?—has no person or number. (3) P. 3.1.40 is conditioned by lIT: no lIT, then no kr-. ˙ ˙ ˙ Now suppose that the lIT of P. 3.1.35 and the lIT of P. 3.1.40 are one and the ˙ ˙ same and that lIT must be there before kr- can be added. In that case the steps ˙ ˙ involved in the derivation of ¯ıha¯m cakre would be as follows: ˙
394
S. D. Joshi and J. A. F. Roodbergen
þ lIT ˙ þ lIT ˙ þ krþ lIT ˙ ˙ If this is what P. 3.1.40 means, three points may be made. In the first place, the derivational procedure is considerably simplified. Second, forms like ¯ıha¯m cakre are ˙ derived as one word, with one accent. In the third place, there is no need for deleting lIT after ¯ıha¯m at all. Or, in other words, P. 2.4.81 is redundant. From this conclusion ˙ another conclusion may be drawn, namely, that P. 2.4.81 was inserted in the Asta¯˙˙ dhya¯yı¯ at some time before Ka¯tya¯yana, presumably in order to accommodate the view of a particular grammarian or his school. (a) ¯ıh (b) ¯ıh (c) ¯ıh
þ a¯m þ a¯m
III
The Lexicon and Change
17
Lexical Irregularity and the Typology of Contrast
Rene´ Kager
17.1
Introduction
This chapter addresses the proper balance between the lexicon and grammar. In particular, it takes a fresh look into lexical phonological irregularity and its implications for the lexicon-grammar interface, integrating insights from three domains: the use of prespecification in blocking alternations (Kiparsky 1993a; Inkelas 1995; Inkelas, Orgun, et al. 1997), phonologically driven allomorphy (Mester 1994; Burzio 1996; Kager 1996; Steriade 1997), and Richness of the Base (Prince and Smolensky 1993; Smolensky 1996; Prince 1998). I propose that irregular alternations involve listed allomorphs whose surface distribution is conditioned by markedness constraints. My allomorphic model strongly reduces the use of morpheme-specific features for lexical irregularity, thus constraining the factorial typology of contrast. Classical rule-based theory (Chomsky and Halle 1968; Zonneveld 1978) used two exception devices. Negative rule features exempt a lexical item marked [R] from undergoing a phonological rule R. Positive rule features license a lexical item marked [þD] to undergo a ‘‘minor rule’’ D. In neither case has the diacritic feature any local articulatory or perceptual definition; it is introduced exclusively to restrict the application of certain phonological rules to designated arbitrary classes of lexical items. Lexical Phonology (LP) (Kiparsky 1982b, 1985) o¤ers a restrictive view of exceptions in phonology. The major observation was made (and accounted for) that exceptions to rules always involve contrastive features, never noncontrastive (allophonic) ones. This observation was explained by the organization of the grammar in LP, distinguishing a lexical and a postlexical component. Contrastive features, by their very nature, are encoded in lexical representations, and hence they can be manipulated or contextually referred to by lexical phonological rules. Since all of a morpheme’s distinctive features (including its exceptional properties with respect to rules) are located in lexical representation, lexical rules are predicted to have (positive or negative) exceptions. In the postlexical component, feeding on the output of
398
Rene´ Kager
the lexicon, direct access to underlying representation (UR) is lost, precluding any reference to exception features by postlexical rules. For reasons unrelated to irregularity, the burden of explanation in phonology has shifted from rules and derivations to constraints and output representations, culminating in Optimality Theory (OT; Prince and Smolensky 1993), a theory of ranked and violable constraints on output forms. The issue naturally arises of how OT deals with lexical irregularity and, moreover, whether OT captures the major generalization that lexical irregularity involves contrastive features. Upon first inspection, three assumptions of standard OT, all of which will be upheld in the analysis that follows, seem to preclude a treatment of lexical irregularity. First, according to Richness of the Base (Prince and Smolensky 1993; Smolensky 1996), no constraints hold on lexical representations. Anything can be an input—it is only the grammar that determines which input forms surface and which are suppressed. Noncontrastive features, for example, can be specified in underlying forms, since the grammar fully neutralizes any such specifications at the surface. Contrast and neutralization are conceived of as interactions of two constraint types. Markedness constraints ban specific feature values from the surface (generally or contextually). These are counterbalanced by faithfulness constraints, enforcing the realization of lexical input feature values. Neutralization results if a markedness constraint with respect to a feature [F] dominates the faithfulness constraint with respect to this feature; contrast results whenever the ranking is the reverse. (1) Neutralization and contrast as interactions of constraints a. Neutralization: Markedness constraint Faithfulness constraint b. Contrast: Faithfulness constraint Markedness constraint On this view, it is not immediately clear how a contrastive specification of a feature di¤ers from an exceptional specification of that feature. Should there be any di¤erence? Second, the prevailing position among OT theorists is that because constraints in Con (the universal constraint set from which a grammar draws) are universal, they cannot refer to individual morphemes. Although it has been argued that certain constraint types are to some extent morpheme specific (e.g., alignment constraints that position an a‰x as prefix or su‰x; McCarthy and Prince 1993a), I assume here that markedness and faithfulness constraints are blind to individual morphemes (pace Hammond 1995; Russell 1995). Conversely, I also assume that individual morphemes cannot be lexically specified, positively or negatively, for being in the scope of some constraint. This type of diacritic has, as far as I know, never been proposed in OT, but it is worth stating the restriction. Third, no morpheme-specific constraint rankings are allowed to encode a morpheme’s phonological exceptionality. Although some OT analyses have proposed
Lexical Irregularity and the Typology of Contrast
399
weakening this assumption, cophonologies (Inkelas and Orgun 1995; Itoˆ and Mester 1995b) are motivated only on the basis of compelling morphological evidence for the organization of the lexicon into strata, on the basis of both the distribution and the shared phonological behavior of classes of morphemes. (See Inkelas, Orgun, et al. 1997 for strong arguments against morpheme-specific cophonologies; Kirchner 1993; Cohn and McCarthy 1998; Pater 1995; Fukazawa, Kitahara, et al. 1998; and others.) In combination, these assumptions block the OT analog of (positive and negative) rule features, disallowing diacritic reference by the grammar to individual morphemes. Given such a set of strong restrictions on the lexicon-phonology interaction, what theory of exceptions has OT to o¤er? I will argue for a maximal separation between lexicon and grammar. The grammar is exclusively responsible for determining the ‘‘contrast space’’ within which alternations in a given language must occur. Any irregularity of morphemes is exclusively captured in the lexicon by lexical specification of the relevant property rather than by diacritics. 17.2
The Factorial Typology of Contrast
To assess the e¤ects of lexical exceptionality on contrast space, we need an optimality-based theory of contrast (Prince and Smolensky 1993; McCarthy and Prince 1995; Smolensky 1996; Kirchner 1997; Prince 1998). This rests upon two assumptions: first, the aforementioned Richness of the Base, according to which no constraints on contrastiveness hold at the input level, and second, a division in Con between markedness and faithfulness constraints, of which only the latter have access to the lexical input while the former are strictly surface based. Markedness constraints (M) are blind to the lexical input, evaluating surface wellformedness only. They can be general (M-General) or contextual (M-Specific). For example, nasality in vowels is evaluated by a pair of markedness constraints in (2), one banning nasal vowels regardless of context (2a) and the other requiring nasality of vowels in a specific context, before a nasal consonant (2b). (2) Examples of general and contextual markedness constraints a. *VNASAL Vowels are not nasal. (general markedness constraint) b. *VORAL N Vowels are nasal before nasal consonants. (contextual markedness constraint) Contextual markedness constraints account for positional neutralization, the phenomenon that contrasts are neutralized in specific positions (while being preserved elsewhere).1
Rene´ Kager
400
Faithfulness constraints (Faith) militate against any divergence between input (lexical) and output (surface) specifications of a segment for some feature F or structural property P. In contrast to markedness constraints, faithfulness constraints have access to lexical representations. An example is (3). (3) IDENT-IO[nasal] Correspondents in input and output have identical values for [nasal]. For the sake of simplicity, I make the assumption that faithfulness constraints are always context free in the sense that their violation never depends on the (prosodic) position of a segment in surface form. This assumption may prove to be too strong, given arguments by Beckman (1998) and others that faithfulness of segments occupying ‘‘strong’’ positions (stressed syllables, onsets) is evaluated by positionally specific constraints, outranking general faithfulness constraints.2 Positional Faithfulness, and its consequences for the factorial typology, will be briefly considered in section 17.7. 17.2.1
The Factorial Typology of Contrast: Nonalternating Cases
Full permutation of rankings of universal constraints should not produce any (types of ) grammars that are crosslinguistically unattested. For example, no language is known to contrast oral and nasal vowels only before nasal consonants while neutralizing nasality in vowels elsewhere. Such a situation should be ruled out by the factorial typology—which is to say that no logically possible reranking should produce it. A factorial typology should predict the clusterings of linguistic properties on a crosslinguistic basis. Reranking the three constraint types (general and contextual markedness and faithfulness) results in a four-way typology (McCarthy and Prince 1995; Prince 1998; Kager 1999). In (4), three constraints pertaining to nasality in vowels are reranked, and predictions are shown for two contexts: a prenasal position ( pan, pa˜n), for which general markedness (*Vnasal ) and contextual markedness (*Voral N) are potentially active, and an elsewhere position ( pa, pa˜), for which only general markedness is active.3 (4) A factorial typology of markedness and faithfulness for [nasal] in vowels Full contrast Faith M-Specific, M-General pan pa˜n pa pa˜ Contextual neutralization M-Specific Faith M-General pa˜n pa pa˜ Total neutralization M-General M-Specific, Faith pan pa Complementary distribution M-Specific M-General Faith pa˜n pa
Lexical Irregularity and the Typology of Contrast
401
A full contrast, in which oral and nasal vowels freely contrast in both contexts, occurs when faithfulness (in this particular case, Ident-IO[nasal]) is undominated, obscuring any e¤ects of markedness constraints. Furthermore, contextual neutralization (nasalizing all prenasal vowels) occurs if contextual markedness (*Voral N) is undominated while contrast is maintained elsewhere (Faith M-General). Additionally, total neutralization (with all vowels oral) stems from undominated general markedness, leaving no room for any role of the lexical input, nor for contextual markedness. Finally, complementary distribution (such that a vowel is nasal if and only if it is in prenasal position, producing allophonic variation) results when both markedness constraints rank above faithfulness, with the contextual constraint taking precedence over the general contraint. The typology correctly excludes languages in which a nasal-oral contrast in vowels occurs prenasally only ( pan, pa˜n) but is neutralized elsewhere ( pa, not *pa˜). This result rests on the assumption that contextual constraints are negative (neutralizing contrasts in specific contexts), not positive (licensing contrasts in specific contexts). Accordingly, positional faithfulness/licensing (Steriade 1995a; Zoll 1996; Beckman 1998) a¤ects the factorial typology in some respects, to be briefly discussed in section 17.7. Degrees of contrastiveness predicted by a factorial typology of contrast can be represented in another way, using Venn diagrams. The area enclosed by the outer oval represents the general context (contrast space), properly including a specific context, represented by the inner oval. Under full contrast (5a), both values of a feature occur in both the specific and general context. Under contextual neutralization (5b), however, only a single value may occur in the specific context, as indicated in the diagram by the exclusion of the lacking specification [nasal] from the inner oval, representing the prenasal context referred to by the constraint *Voral N. (5) a. Full contrast
Faith M-Specific, M-General
b. Contextual neutralization
M-Specific Faith M-General
Both diagrams show situations of partial contrast, with both values licensed in the general context, due to faithfulness dominating general markedness.
Rene´ Kager
402
(6) (Partial) contrast Faithfulness M-General Two diagrams represent the reverse ranking, a complete loss of contrast, in di¤erent ways. First, under total neutralization (7a) toward orality, [nasal] surfaces in both contexts. Second, under complementary distribution (7b), the values [nasal] or [þnasal] occur in complementary contexts. (7) a. Total neutralization
M-General M-Specific, Faith 17.2.2
b. Complementary distribution
M-Specific M-General Faith
The Factorial Typology of Contrast for Alternations
All of a morpheme’s contextual realizations are licensed by the grammar, which is to say that each alternant constitutes a phonotactically licit output form, assuming a language’s phonotactics (defined by the grammar by interactions of markedness and faithfulness constraints). Since alternations are, in this sense, structure preserving (respecting the system of contrasts in the phonology of nonalternating morphemes), we can represent alternations in contrast space by connecting a morpheme’s specification for a feature [F] in the specific context to its specification for [F] in the general context. Abstracting away from nasality in vowels, let us assume a binary feature [F] with values [aF] (the general unmarked value) and [aF] (the general marked value) and a pair of markedness constraints, one general (MG ) and one contextual (MS ): (8) MG : *[aF] MS : *[aF] in some environment E
(general markedness constraint) (contextual markedness constraint)
First we consider the simplest situation, a complete loss of contrast, manifested either as total neutralization (9a) or complementary distribution (9b). Arrows connect the realizations of a morpheme in the special context and the general context, sometimes representing an alternation, as in (9b).
Lexical Irregularity and the Typology of Contrast
(9) a. Total neutralization
M-General M-Specific, Faith
403
b. Complementary distribution
M-Specific M-General Faith
Under total neutralization no alternations occur, and all morphemes are fixed [aF], the context-free unmarked value. In complementary distribution, contextual markedness takes priority over general markedness, making all morphemes alternate between [aF] and [aF]. Next we look into full contrast and contextual neutralization, connecting contextual values of [F] in the following ways. (10) a. Full contrast
Faith M-Specific, M-General
b. Contextual neutralization
M-Specific Faith M-General
Under full contrast (10a), both values freely occur in both contexts, allowing two types of nonalternating morphemes ([aF] and [aF]). Alternations are blocked by undominated faithfulness, obscuring any e¤ects of markedness. In tableau (11), candidates are listed as miniparadigms, pairs of values of [F], one for each context: general ([aF]G ) and specific ([aF]S ). In cells below constraints, violations for individual contexts have been added up.
Rene´ Kager
404
(11) Full contrast: IO-Faith MG , MS /aF/
IO-Faith
*[aF]G
*[aF]S
+ [aF]G @ [aF]S
*
[aF]G @ [aF]S
*!
*
[aF]G @ [aF]S
*!
*
[aF]G @ [aF]S
*!*
**
*
/aF/ [aF]G @ [aF]S
*!*
[aF]G @ [aF]S
*!
*
[aF]G @ [aF]S
*!
*
*
+ [aF]G @ [aF]S
*
**
Lexical inputs to this tableau reflect the assumption that lexical specification is strictly binary, /[aF]/ or /[aF]/. Richness of the Base forces us, however, to consider two more possibilities: underspecification ([qF]), discussed in section 17.4, and listed allomorphy ([aF] @ [aF]), the topic of section 17.6. In contextual neutralization (10b) a subset of morphemes are fixed [aF] in all contexts, while others alternate [aF]S @ [aF]G . Top-ranked contextual markedness excludes nonalternating [aF]; elsewhere, faithfulness prevails over general markedness, producing a contrast. (12) Contextual neutralization: MS IO-Faith MG /aF/ [aF]G @ [aF]S
*[aF]S
*[aF]G
*
*
*
*
**!
**
*!
+ [aF]G @ [aF]S [aF]G @ [aF]S
IO-Faith
*!
[aF]G @ [aF]S /aF/ [aF]G @ [aF]S
*!
[aF]G @ [aF]S [aF]G @ [aF]S + [aF]G @ [aF]S
*!
** *!
*
*
* **
Lexical Irregularity and the Typology of Contrast
405
Summarizing, we find that under the assumption of strictly binary input specification, only two contrastive patterns of alternations are predicted by the factorial typology: full contrast (with nonalternating morphemes of both values while lacking alternating morphemes) and contextual neutralization (alternating morphemes and fixed morphemes of only one value). The factorial typology correctly excludes two logically possible types of alternation in which feature points are connected in the following ways. (13) a. Reverse neutralization
b. Crossed alternation
(13a) shows ‘‘reverse neutralization’’ to a marked value [aF]S , the situation in which morphemes alternate between [aF]G @ [aF]S , their values crossed as compared to the ‘‘natural’’ alternation [aF]G @ [aF]S . Crossed neutralizations are intrinsically suboptimal, as no ranking of our current constraints derives them (Prince 1998). Tableau (14) shows how the violation marks of a pair of ‘‘natural’’ alternants are properly included in those of ‘‘reverse’’ alternants. (14) Intrinsic suboptimality of ‘‘reversed’’ alternation *[aF]S + [aF]G @ [aF]S [aF]G @ [aF]S
*!
IO-Faith
*[aF]G
*
*
*
*
Regardless of lexical value (/[aF]/ or /[aF]/), alternation implies a violation of faithfulness by one of the alternants, canceling out violations in the candidates. One alternant also violates general markedness, again canceling out violations. Any differences between candidates reside only in their violation of specific markedness, a circumstance that favors a ‘‘natural’’ alternation. Basically the same explanation holds for ‘‘crossed alternation’’ (13b), a combination of a regular and a reverse alternation.
Rene´ Kager
406
17.3
Neutrast
A third logically possible type of alternation is incorrectly excluded by the typology. The typology is unable to express a three-way division of morphemes in which there are two fixed-value morphemes plus alternating ones (Inkelas 1995). I will call this pattern a ‘‘neutrast.’’ (15) Neutrast
The neutrast combines ingredients of full contrast and contextual neutralization. As in full contrast, both values occur freely in both contexts, but at the same time alternating morphemes occur, as in contextual neutralization. Assuming a strictly binary input specification, the neutrast cannot be derived under any ranking. This is because (as we have seen above) free occurrence of both values in all contexts implies that the relevant faithfulness constraint (Ident-IO[F]) is undominated, blocking any alternations. For alternation to arise, a contextual markedness constraint must dominate faithfulness. Neutrast is incorrectly excluded by the factorial typology, since at least two types are crosslinguistically attested. First, full contrast combines with a designated set of alternating morphemes, a situation handled in rule theory by minor rules. Second, contextual neutralization may be subject to lexical exceptions, a situation handled by negative exception features in rule theory. Both types of neutrast are represented below. Solid lines indicate ‘‘regular’’ morphemes, while dotted lines indicate (positive or negative) exceptions. (16) a. Neutrast by minor rule
b. Neutrast by exception to neutralization
Lexical Irregularity and the Typology of Contrast
407
I hypothesize that any grammar in which neutrast with respect to [F] occurs necessarily licenses a full contrast of [F], implying a ranking IO-Faith MG . The following generalization thus seems to hold. (17) Generalization Lexical irregularity with respect to [F] may occur only in grammars that maintain a contrast of [aF] @ [aF]. This generalization echoes a theorem of LP (Kiparsky 1982b) that lexical rules may have exceptions, while postlexical rules are exception free. For exceptionality with respect to a feature [F] to occur, [F] must occur in lexical representations, which implies that [F] is contrastive. In OT, due to Richness of the Base, contrast is not captured at the level of lexical representation but at surface level by interactions of constraints. The question then is can OT capture generalization (17)? Earlier I noted that the factorial typology fails to predict neutrast. One way of capturing neutrasts in OT is by diacritic means: morpheme-specific faithfulness constraints or morpheme-specific cophonologies. In the context of this chapter, the drawback of diacritic approaches is that these essentially abandon contrast as a grammar-wide notion, leaving (17) unexplained. If individual morphemes can enforce their lexical specifications by their own faithfulness constraints ranked in specific positions in the hierarchy, the entire notion of contrast becomes meaningless. All phonological features would become contrastive, as no exceptionless neutralization of any feature could be enforced by the grammar. That is, it would become an accident that for a noncontrastive feature [F], every morpheme-specific faithfulness constraint enforcing [F] is dominated by some markedness constraint neutralizing [F]. To enforce grammar-wide noncontrastivess of a feature, language-specific metarankings fixing the rankings of morpheme-specific faithfulness constraints would be needed. Morpheme-specific cophonologies are more restrictive than morphemespecific constraints, but as Inkelas, Orgun, et al. (1997) demonstrate, serious problems still remain. To account for neutrast while explaining generalization (17), we must give up the assumption of strict binarity in lexical specification and allow for input ternarity, as Inkelas (1995) and Inkelas, Orgun, et al. (1997) convincingly argue on the basis of negative exceptions.4 My theory of neutrast builds on insights from Inkelas’s prespecification theory, to be discussed in section 17.4. Section 17.5 will argue, however, that prespecification accounts for a subset of neutrasts only, that is, those involving segmental features but not those involving prosodic properties. To achieve full coverage, I will propose a theory of ‘‘Phonology-Driven Lexical Allomorphy’’ in section 17.6, allowing a set of alternants in a morpheme’s lexical representation and deriving phonology-governed distributions by markedness constraints. Section 17.7 will address the consequences of lexical allomorphy for the factorial typology.
Rene´ Kager
408
17.4
Archiphonemic Underspecification and the Factorial Typology of Contrast
As implied by Inkelas (1995), the assumption of strictly binary lexical specification is incompatible with Richness of the Base. Both stronger and weaker versions of lexical binarity are ruled out. Under a strong version, the lexicon contains only two values for any feature [F], whether [þ/], [þ/0], or [/0]. The weaker version allows all three values in the lexicon while limiting prespecification to two values in any given context (contextual prespecification, Kiparsky 1993a). Under Richness of the Base, the grammar has to produce a mapping for all three logically possible input values of [F] in any context. We now turn to a prespecification account of ternary contrasts in the context of the factorial typology of contrast. 17.4.1
Negative Exceptions to Neutralization
When considering contextual neutralization in the context of a discussion of contrast, we must distinguish degrees of generality: exceptionless neutralization and exceptionsensitive neutralization. That is, given a feature [F], whose values [aF] and [aF] freely occur in a set of environments, EG , there exists a specific environment ES in which either
only the value [aF] occurs (exceptionless neutralization), or the value [aF] occurs residually (exception-sensitive neutralization).
An example of exceptionless contextual neutralization is Dutch coda devoicing (Kager 1999, Booij this volume). Stems ending in obstruents that, by a‰xation, occur in two phonological contexts (onset and coda), fall into two lexical sets: nonalternating stems have voiceless obstruents everywhere (e.g., [pt]@[p.t n] ‘cap(s)’), while alternating stems have two alternants—voiceless coda and voiced onset (e.g., [bt]@[b.d n] ‘bed(s)’). e
e
(18) Dutch devoicing a. Alternating stems [voice]@[þvoice] bed [bt] ‘bed’ bedden [b.d n] ‘beds’ b. Nonalternating stems [voice] pet [pt] ‘cap’ petten [p.t n] ‘caps’ c. No nonalternating stems [þvoice] e
e
No stems occur whose final consonant surfaces as voiced in both contexts. Of course, this gap is not due to a lexical property of stems; it falls out of a general phonotactic pattern of Dutch: coda obstruents are always voiceless. The grammar of Dutch expresses this generalization. Under exceptionless neutralization, a contextual markedness constraint outranks a faithfulness constraint. I assume three markedness constraints to be potentially rele-
Lexical Irregularity and the Typology of Contrast
409
vant to obstruent voicing. Two are contextual markedness constraints; the third bans voiced obstruents across the board. (19) Markedness constraints potentially relevant to obstruent voicing: a. NO-VOICED-CODA Obstruents are voiceless in coda position. b. INTER-V-VOICE Obstruents are voiced intervocalically.5 c. *[son, þvoice] Obstruents are voiceless. The ranking (20) captures complete neutralization of voice in Dutch codas. (20) Exceptionless neutralization of voice in codas: M-Specific Faith No-Voiced-Coda Ident-IO[voice] Inter-V-Voice, *[son, þvoice] This is simply a case of the ranking scheme (10b) with an extra contextual markedness constraint. Note that the relative ranking of *[son, þvoice] and Inter-V-Voice cannot be established. First, for any stem whose lexical input is either [voice] or [þvoice], Ident-IO[voice] obscures any e¤ects of the lower-ranked markedness constraints. Hence, /pt/ surfaces faithfully as [pt @ pt- n], while /bd/ surfaces as [bt @ bd- n], maximally faithful to its input specification to the extent that NoVoiced-Coda allows it. Second, a hypothetical stem unspecified for [voice] runs into a competition between Inter-V-Voice and *[son, þvoice], since any output candidate will violate Ident-IO[voice] to the same extent, twice in this case.6 Assuming a hypothetical stem /md/, either of two output pairs is optimal. e
e
(21) Unestablished constraint ranking for hypothetical unspecified input No-VoicedCoda
Ident-IO [voice]
Inter-VVoice
+ mt @ mt- n
**
*
+ mt @ md- n
**
/md/ e
e
md @ mt- n
*!
**
md @ md- n
*!
**
*[son, þvoice]
* *
*
e
e
Note that these are precisely the two output patterns for lexically specified stems. Since underspecification fails to add a third surface pattern, there is no reason for the learner to posit a three-way lexical contrast. In sum, we have not seen any crucial evidence for underspecification so far.7
Rene´ Kager
410
Turkish coda devoicing, which is discussed by Inkelas (1995), Inkelas and Orgun (1995), and Inkelas, Orgun, et al. (1997), does present such evidence. As compared to Dutch, Turkish has an additional type of stem (22c). (22) Turkish devoicing a. Alternating stems [voice]@[þvoice] kanat ‘wing’ kanad-i kanat-lar ‘wing-pl’ kanad-im b. Nonalternating stems [voice] sanat ‘art’ sanat-i sanat-lar ‘art-pl’ sanat-im c. Nonalternating stems [þvoice] etu¨d ‘etude’ etu¨d-u¨ etu¨d-ler ‘etude-pl’ etu¨d-u¨m
‘wing-acc’ ‘wing-1sg.poss’ ‘art-acc’ ‘art-1sg.poss’ ‘etude-acc’ ‘etude-1sg.poss’
Inkelas (1995), following an earlier proposal by Hayes (1990), argues that this threeway contrast is due to a three-way lexical specification: [þvoice], [voice], [0voice]. In the Turkish grammar, faithfulness to [voice] is high ranking, such that any prespecified value of voice surfaces in all contexts; alternating stems are lexically underspecified, their surface values filled in by markedness constraints, which can be the same as for Dutch, in keeping with the assumption that constraints are drawn from a universal set. The markedness constraints pertaining to [voice] ‘‘apply’’ in a fashion similar to feature-filling rules in LP. All three logically possible lexical values coincide with an attested stem pattern. (23) Three-way contrast in URs Stem type Lexical Unspecified /d/ Specified [voice] /t/ Specified [þvoice] /d/
Output [t @ d] [t] [d]
Example /kanad/ /sanat/ /etu¨d/
This three-way division among stems constitutes the empirical argument for underspecification, which could not be construed for Dutch, where only two patterns surface due to exceptionless contextual neutralization. Inkelas (1995) and Inkelas, Orgun, et al. (1997) argue that the three-way contrast implies undominated faithfulness to the lexical specification of [voice]. Dominated markedness constraints take e¤ect only in the case of an unspecified lexical input, in a feature-filling fashion. Intervocalic voicing and coda devoicing arise from the ranking below. (24) Exception-sensitive neutralization of voice: Faith M-Specific Ident-IO[voice] No-Voiced-Coda, Inter-V-Voice *[son, þvoice]
Lexical Irregularity and the Typology of Contrast
411
The minimal di¤erence with the Dutch ranking (20) is that Ident-IO[voice] dominates all markedness constraints and, crucially, No-Voiced-Coda. The tableaux in (25) show the selection of the most faithful candidate paradigms for inputs specified as [voice] and [þvoice], respectively. (25) Fully specified inputs: Nonalternation enforced by Faith M-Specific /sanat/
Ident-IO [voice]
No-VoicedCoda
+ sanat @ sanat-
Inter-VVoice
*[son, þvoice]
*
sanat @ sanad-
*!
sanad @ sanat-
*!
*
sanad @ sanat-
*!*
*
* *
* **
/etu¨d/ etu¨t @ etu¨t-u¨
*!*
etu¨t @ etu¨d-u¨
*!
etu¨d @ etu¨t-u¨
*!
+ etu¨d @ etu¨d-u¨
* * *
*
*
*
**
Alternating forms derive from an underspecified input, for example, /kanad/. Inkelas, Orgun, et al. (1997) assume that insertion of a value of [voice] in lexically underspecified forms amounts to violation of a faithfulness constraint, Faith, which I will equate with the correspondence constraint Ident-IO[voice]. If we assume surface specifications to be fully specified ([þvoice] or [voice]), all candidates incur the same amount of violation of this constraint.8 (26) Underspecified inputs: Alternation enforced by a pair of markedness constraints /kanad/
Ident-IO [voice]
No-VoicedCoda
kanat @ kanat-
**
+ kanat @ kanad-
**
kanad @ kanat-
**
*!
kanad @ kanad-
**
*!
Inter-VVoice
*[son, þvoice]
*! * *
*
Rene´ Kager
412
With violations of Ident-IO[voice] cancelled between candidate paradigms, previously dormant contextual markedness constraints are activated, favoring devoiced codas and voiced intervocalic obstruents and giving a surface alternation [kanat @ kanad-i]. We witness an ‘‘emergence of the unmarked,’’ a phenomenon well attested in reduplication and other phenomena (McCarthy and Prince 1994, 1995). Judged by the numerical distribution in the Turkish lexicon, stems specified as [voice] or [0voice] are regular, while stems specified as [þvoice] are exceptional. From the prespecification perspective of Inkelas, Orgun, et al. (1997), however, voicing in Turkish is simply a full contrast, enforced in all contexts by a high-ranking faithfulness constraint. The implicit claim is that the skewed distribution of stems in three classes has no grammatical status: the grammar does not refer to individual stems, nor to any lexical diacritic on specific stems. In (27), an instantiation of the neutrast (16b), the exceptional status of stems with prespecified [þvoice] is indicated by dotted lines. This amounts to exceptions to final devoicing (van Oostendorp 2007). (27) Incomplete contextual neutralization
In sum, the grammar (not the lexicon) determines the degree of generality of neutralization. The constraint ranking of exception-sensitive neutralization is identical to that of full contrast: F M. While the grammar determines the contrast space within which all alternations must occur, the lexicon supplies morphemes ([þF], [F], [0F]) in various frequencies. 17.5
Problems with Underspecification
Although archiphonemic underspecification provides an adequate account of ternary alternations that involve a binary segmental feature, it is di‰cult to extend to prosodic alternations. Segment insertion, deletion, and alternations of length do not involve a binary feature but rather some prosodic element (mora, root node, etc.). While the lexical presence and absence of a prosodic element can be viewed as the
Lexical Irregularity and the Typology of Contrast
413
counterparts of a prespecified feature [þF] and [F], respectively, it is not clear what the counterpart of the underspecified value [0F] might be. I will argue below that what seems the most reasonable choice, lexical specification of autosegmental association lines, is essentially a nonalternation diacritic in disguise. Let us focus on a length alternation. An example is trisyllabic shortening (TSS; Kiparsky 1982b), a contextual neutralization that has lexical exceptions.9 (28) Ternary alternation in TSS a. Alternating vowel (‘‘regular undergoer’’) ser[i]ne @ ser[e˘]nity b. Nonalternating short vowel (‘‘vacuous nonundergoer’’) tranqu[I]l @ tranqu[I]lity c. Nonalternating long vowel (‘‘exceptional nonundergoer’’) ob[i]se @ ob[i]sity I make the standard assumption that TSS is an alternation of length, involving monomoraic and bimoraic vowels.10 If we conceive of length as a property whose surface specifications are monomoraic or bimoraic, while lexical specifications add a third (underspecified) value, the question of the counterpart of /0F/ arises. First, if we assume the alternating vowel of serene to be lexically long, then prespecified length is no longer available to encode the exceptional nonshortening of obesity. Second, if we prespecify nonalternating obesity as long, it becomes unclear what to underspecify in serene @ serenity. For example, if serenity is short, with structurefilling lengthening in serene, the contrast with tranquil is lost. Simple moraic preservation does not su‰ce to capture the ternarity. We need extra machinery: a lexical contrast between prelinked and floating moras.11 That is, we might distinguish serene from obese in the following fashion. The second vowel of obese would have two prelinked moras, while that of serene would have one linked mora plus a floating one. Consider the lexical representation of a vowel with a floating mora (29a) and its surface alternants (29b, c). (29) a. m m V
b. m m V
c. m V
(I assume concreteness of representation: surface alternants of underspecified vowels are represented identically to nonalternating long and short vowels.)12 To enforce nonshortening and nonlengthening by mora prespecification, we need a faithfulness constraint militating against changes in input length, Wt-Ident-IO (McCarthy 1995; Borowsky and Harvey 1997; Broselow, Chen, and Hu¤man 1997; Gussenhoven this volume).
Rene´ Kager
414
(30) WT-IDENT-IO If a A Domain( f ), if a is monomoraic, then f (a) is monomoraic. if a is bimoraic, then f (a) is bimoraic.
(no lengthening) (no shortening)
Wt-Ident-IO requires identical quantity of output segments and their input correspondents. This requirement is made in both directions, excluding both the addition of quantity (lengthening) and its loss (shortening). Note that the constraint does not refer to association lines, on the tacit assumption that lexical moras are preassociated. With underspecification of length, however, this assumption cannot hold. If the alternating stems are lexically represented by a floating mora, then nonalternating stems are not faithful to moras but to root-to-mora association lines. Let us examine the four logically possible changes. Two occur in alternations (31a, b), and the other two are excluded nonalternations (31c, d). (31) a. ser[i]ne mm mm ! V
V
c. *ob[e˘]sity mm m ! V
add line
V
b. ser[e˘]nity mm m ! V
delete line, delete m
V
d. *tranq[þI]l m mm ! V
delete m
add line, add m
V
The only way to distinguish (31a, b) from (31c, d) is to render Wt-Ident-IO sensitive to di¤erent violations of moras and association lines separately. A constraint NoDelink, enforcing the autosegmental association of segments to moras, was proposed in McCarthy (2000). Informally, this constraint says that if a segment S1 is linked to a mora m1 in the input, then its correspondent segment S2 must be linked to the corresponding mora m2 in the output. Note that for a violation of No-Delink to occur, the segment and the mora must be present in both input and output. For obesity, however, we crucially need violation of No-Delink for a prelinked mora that is no longer present in the output. A stipulation must then be made that a deletion of a prelinked mora violates No-Delink. This stipulation is entirely unmotivated in the area of autosegmental phenomena for which No-Delink was originally proposed by McCarthy (2000). More generally, it amounts to a serious weakening of Correspondence Theory, in particular its assumption that failed associations are evaluated only for segments that stand in correspondence. This is the first problem. A prespecification analysis of TSS based on this assumption is given in tableau (32). Undominated No-Delink blocks shortening in obesity. Shortening of serenity (that is, deletion of its floating mora) satisfies No-Delink, even though it violates
Lexical Irregularity and the Typology of Contrast
415
Wt-Ident-IO. Shortening itself is conditioned by a metrical constraint EvenTrochee (Prince 1990; Hayes 1995), requiring the trochaic foot to be quantitatively balanced (i.e., short-short). (32) A prespecification analysis of TSS Prelinked length
No-Delink
+ ob[i]sity ob[e˘]sity
EvenTrochee
Wt-Ident-IO
* *!
No-Long-V *
*
Floating length ser[i]nity + ser[e˘]nity
*!
* *
This brings us to the second problem for the floating mora account. Its dual use of faithfulness (to segment-to-mora association lines and moras separately) amounts to diacritic use of association lines. Both constraints, No-Delink and Wt-Ident-IO, have identical patterns of violation except that the former enforces length in nonalternating stems only, while the latter is relevant to all types of stem. No-Delink, under its revised interpretation, is essentially a parochial version of length faithfulness (Wt-Ident-IO) with local relevance for a designated class of stems. It is di‰cult to distinguish this analysis from one involving morpheme-specific (diacritic) ranking of a pair of faithfulness constraints, with nonalternating stems selecting a nondeletion ranking (Faith1 Even-Trochee) and alternating ones selecting a deletion ranking (Even-Trochee Faith2 ). In sum, a prespecification theory of ternary alternation successfully deals with structure-filling feature alternations but runs into problems in dealing with ternary prosodic alternations. Since alternating prosodic elements such as moras cannot be underspecified themselves, indirect underspecification is required, implying faithfulness to lexical moraic association lines, which violates central assumptions of Correspondence Theory and is arguably diacritic. Ternary alternations involving entire segments (deletions or insertions) involve similar (or more serious) complications.13 Nevertheless, I believe that the prespecification theory of neutrasts o¤ers two ideas that are key to analyzing the phenomenon in a more general way. First is the idea that neutrast is essentially a full contrast, in which IO-Faithfulness to a lexical property P dominates all markedness constraints with respect to P. This captures the generalization that neutrast only occurs in grammars that license a contrast with respect to P. Hence, neutrast is not a ‘‘leaking’’ contextual neutralization, with lexical exceptions to a neutralization ranking M-Specific Faith. The second valuable idea is
Rene´ Kager
416
cancellation of faithfulness violations in alternating morphemes in neutrast, allowing markedness constraints to jump into activity (producing an emergence of the unmarked). Cancellation is straightforward in the case of featural alternations but comes at a higher cost in dealing with prosodic alternations, involving length, and presence or absence of entire segments. 17.6
Lexical Allomorphy
We can maintain both central insights into ternary alternation, full contrast ranking and the cancellation of faithfulness in alternants, while avoiding the problems of prespecification of prosodic properties. This can be achieved by allowing multiple alternants in a morpheme’s lexical representation, which are subject to selection by the phonology. Such a theory of Phonology-Driven Lexical Allomorphy will be developed in section 17.6.3. 17.6.1
Allomorphy in Dutch Length Alternations
I begin by introducing the ‘‘minor rule’’ that is the focus of this section. Dutch has an alternation of vowel length known as ‘‘open syllable lengthening,’’ (OSL) which occurs in a number of (mostly nominal) stems (Zonneveld 1978; Booij 1995, this volume), as exemplified in (33c). Nouns in (33a) and (33b) represent the far larger classes of nonalternating stems.14 (33) Ternary alternation in Dutch OSL a. Nonalternating short vowel (many stems) kl[þ]s @ kl[þ]sen ‘class(es)’ p[]t @ p[]ten ‘pot(s)’ h[]g @ h[]gen ‘hedge(s)’ k[I]p @ k[I]pen ‘chicken(s)’ b. Nonalternating long vowel (many stems) b[a]s @ b[a]zen ‘boss(es)’ p[o]t @ p[o]ten ‘paw(s)’ r[e]p @ r[e]pen ‘bar(s)’ c. Alternating short @ long vowel (few stems) gl[þ]s @ gl[a]zen ‘glass(es)’ sl[]t @ sl[o]ten ‘lock(s)’ w[]g @ w[e]gen ‘road(s)’ sch[I]p @ sch[e]pen ‘ship(s)’ In alternating stems (33c), short vowels occur in closed syllables (singulars) and long vowels in open syllables (plurals). In nonalternating stems (33b), long vowels occur in both closed and open syllables. In nonalternating stems (33a), short vowels occur
Lexical Irregularity and the Typology of Contrast
417
in closed syllables (singulars) or before a single intervocalic consonant (plural). Intervocalic consonants after (stressed) short vowels are usually considered ambisyllabic, an aspect to which I will return. As shown in diagram (34), this is a neutrast. Three types of stem occur: those with nonalternating length are represented with solid lines in the diagram, while exceptional alternating stems are represented by a dotted line. (34) Minor rule
Basically this is a full contrast (with short and long vowels occurring freely in both contexts), with a ‘‘leakage’’ of alternating stems. Note that contrast diagrams of exception-to-neutralization (16b, 27) and minor rule (16a, 34) are identical, except for the relative frequency of alternating morphemes. In exception-sensitive neutralization, the nonalternating morphemes are rare, while in minor rules, alternating morphemes are. Since the numerical distribution of morphemes of various alternation types is due to the lexicon, the grammars underlying both types of neutrast will be assumed to be identical. As is typical of minor rules, the direction of the alternation is not directly clear. Alternating stems occur in two contexts, both of which are positively characterized, as either an open or a closed syllable. Upon first inspection, the process may be characterized either as OSL or closed syllable shortening (CSS), both natural processes attested in many of the world’s languages. Both OSL and CSS are interactions of antagonist e¤ects: the maximization of weight in stressed syllables (due to the Stressto-Weight Principle [SWP]) and avoidance of overweight (due to *mmm), which would arise by a long vowel in a closed syllable.15 (35) Constraints triggering length alternations a. SWP If stressed, then heavy. b. NO-TRIMORAIC-SYLLABLES (*mmm) No trimoraic syllables. Nevertheless, the literature on Dutch phonology (Zonneveld 1978; Booij 1995) considers the alternation a lengthening rather than a shortening because of its partly
Rene´ Kager
418
neutralizing character. As indicated in (33c), there are two sources for long [e] in stems participating in the alternation: either // (weg) or /I/ (schip). Because of the neutralization in the direction of the long vowel alternants (/, I/ ! [e]), most analysts assume a minor rule of OSL that has approximately the following format. (36) Dutch OSL as a rewrite rule V ! V / h[back]i h[high]i [þd]
]s
The features in angled brackets express the generalization that front vowels are simultaneously lowered, accounting for the neutralization referred to above. The positive exception feature [þd] spells out the ‘‘handshake’’ between lexicon and grammar: only morphemes marked [þd] undergo the rule. 17.6.2
The Counterpart of Minor Rule in OT
At this point, having discussed a rule-based analysis of OSL, we should ask what the counterpart of the notion of minor rule in OT might be. Two answers can be rejected out of hand, after discussions in earlier sections. First, morpheme-specific rankings (markedness and faithfulness constraints ranked in a morpheme-specific way) have been rejected for the reasons outlined in section 17.3. Essentially, this solution merges the grammar with the lexicon and loses any generalization about the grammar-wide, morpheme-independent roles of contrastive (or redundant) features. Second, the option of morpheme-specific faithfulness constraints is rejected for the same reason. Under this heading, I also rubricize attempts to use prelinked association lines to enforce a distinction between alternating and nonalternating morphemes, as discussed in section 17.5. Third, there is a well-established option of cophonologies, understood as subdividing the grammar into rankings depending on principled lexical strata (Inkelas and Orgun 1995 on Turkish; Itoˆ and Mester 1995b on Japanese). To avoid their diacritic use, however, cophonologies should be used for genuinely morphologically defined strata only. Clearly this is not the case here, since the set of stems participating in OSL does not form a stratum by any independent criteria. Essentially, the rejection of these options entails that we have given up the diacritics approach to exceptions, involving features to establish contact between a set of ranked constraints and a set of designated morphemes. But if the grammar is ‘‘blind’’ to specific morphemes, how do we restrict alternations to specific morphemes? The answer advanced here is that all unpredictable properties of morphemes should be expressed solely in the lexicon, including their alternating or nonalternating status. This brings us to allomorphy. Let us first establish that Dutch OSL has all the properties of a lexically listed alternation (see Lieber 1982; Bybee 1988, 1995; and for Dutch, Booij, 1995, this vol-
Lexical Irregularity and the Typology of Contrast
419
ume). First, the set of alternating stems is idiosyncratic. This is, of course, the essence of a minor rule, an alternation restricted to a designated set of morphemes. In this case, the set of alternating stems fails to reoccur in other alternations in Dutch, which means that the designated class of stems is unpredictable by any independent means, ruling out any stratal account.16 Second, the OSL alternation itself is idiosyncratic because no productive alternations of length in Dutch are accompanied by obligatory tense-lax alternations (see Gussenhoven, this volume). Moreover, OSL is accompanied by alternations in height that are partly predictable (e.g., high [I] only alternates with mid [e]) and partly unpredictable (e.g., leveled height in [þ] @ [a], glas @ glazen, next to alternation in ([þ] @ [e]), stad @ steden). Tenseness and height alternations accompanying OSL in nominal plurals are shown below (partly from Zonneveld 1978, 58; Booij 1995, 87, this volume). (37) Tenseness and height alternations accompanying OSL, with examples of undergoers [I] @ [e] gelid ‘rank’, lid ‘member’, schip ‘ship’, smid ‘blacksmith’ [] @ [e] bevel ‘order’, gebed ‘prayer’, gebrek ‘shortcoming’, spel ‘game’, weg ‘road’ [þ] @ [e] stad ‘city’ [þ] @ [a] bad ‘bath’, bedrag ‘amount’, blad ‘leaf ’, dag ‘day’, dak ‘roof ’, dal ‘valley’, glas ‘glass’, graf ‘grave’, pad ‘path’, staf ‘sta¤ ’, vat ‘barrel’, verdrag ‘treaty’ [] @ [o] gebod ‘command’, god ‘god’, hertog ‘duke’, hof ‘court’, hol ‘hole’, lof ‘benediction’, lot ‘lot’, oorlog ‘war’, slot ‘lock’ The long vowel participating in OSL alternations is obligatorily nonhigh (a fact to which I will return shortly), causing a kind of minor ‘‘neutralization’’ of height in the case of [I, ] @ [e].17 Moreover, no alternations of rounding occur between short and long vowels under OSL. Third, OSL lengthening ‘‘overapplies’’ in certain morphological contexts (Zonneveld 1978), systematically in denominal verbs (smeed ‘to forge’, baad ‘to bathe’, loot ‘to draw lots’), and idiosyncratically in diminutives (glaasje, scheepje). In these morphological contexts, the same stem allomorphs occur as in plurals, but without matching OSL’s canonical open syllable context. Regardless of the issue of how to analyze overapplication, the point is that only stems that have long vowel allomorphs in their plurals can have long allomorphs in diminutives and verbs.18 This ‘‘lexical conservatism’’ (Steriade 1997) is accidental under a derivational analysis of OSL, using minor rules for di¤erent morphological contexts, but not under an allomorphic account (Lieber 1982 o¤ers similar arguments for other languages).19
Rene´ Kager
420
17.6.3
Phonologically Driven Lexical Allomorphy
I now turn to an analysis of OSL alternations by allomorphs. I assume that lexically unpredictable alternations (such as minor rules) are encoded in the lexicon by listed allomorphs, which are selected by markedness constraints. My model of phonologydriven allomorphy builds on earlier work by Prince and Smolensky (1993), Mester (1994), Burzio (1996, to appear), Kager (1996, 1999), Drachman, Kager, et al. (1996), Perlmutter (1996), Steriade (1997), and Hayes (1999b).20 My assumptions are the following. First, a morpheme’s lexical phonological representation may contain one or more allomorphs, each of which is independently available as an input to grammatical mapping. For example, the phonological representation of the stem ‘ship’ in the lexicon is a pair of allomorphs of distinct vowel lengths {/sxIp/1 @ /sxep/2 }, each carrying a di¤erent index for reasons appearing immediately below.21 Second, I assume ‘‘split inputs.’’ Gen supplies a full set of candidate outputs for each individual lexical allomorph. Accordingly, an output candidate is a pair consisting of a candidate output analysis plus the lexical allomorph that serves as its input. The relation between output analysis and its input allomorph is indicated by coindexation, allowing IO-Faithfulness constraints to match each output analysis to its lexical allomorph. An output candidate C can thus be seen as a coindexed pair of an output analysis O and a lexical allomorph I, together (Ii , Oi ). In our particular case, the output candidates [sxep]1 and [sxe.p]2 are phonetically identical, but each is based on a di¤erent lexical allomorph. Third, Eval considers all output candidates based on di¤erent lexical allomorphs in parallel. In the case of output candidates [sxep]1 and [sxe.p]2 , only the former (coindexed with the lexical allomorph /sxIp/1 ) violates faithfulness to length, tenseness, and height; the latter (coindexed with /sxep/2 ) is fully faithful. More generally, a faithfulness constraint evaluating identity for a feature [F] faces structurally identical output candidates, O1 and O2 , each having a di¤erent allomorph I1 and I2 as its input. Since O1 and O2 are structurally identical, both specified as [þF], the candidate is selected whose input allomorph is specified as [þF]. The net result is a kind of cancellation of faithfulness: in the case of two allomorphs with opposite values of [F], at least one output candidate will satisfy the faithfulness constraint. (38) Vacuous optimization of input allomorph {[þF]1 , [F]2 }
Ident-IO[F]
+ [þF]1 [þF]2
*!
Lexical Irregularity and the Typology of Contrast
421
Now consider a case in which the candidate set, at the point of evaluation by IdentIO[F], still contains potentially winning candidates for both values of [F], as in the next tableau. Since for each value of [F] in output candidates an input allomorph can be found that is faithful, candidates of both values (here, [þF]1 and [F]2 ) are passed on for evaluation by markedness. (39) Invisibility of allomorphy to faithfulness {[þF]1 , [F]2 }
Ident-IO[F]
*[þF] *!
[þF]1 [þF]2
*!
[F]1
*!
*
+ [F]2 This explains why allomorphy seems ‘‘invisible’’ to faithfulness constraints, an explanation requiring nothing but standard assumptions about the lexical input (Richness of the Base), Gen (Freedom of Analysis), and Eval (minimal violation). Invisibility to faithfulness implies that phonological distributions of allomorphs (everything else being equal) depend entirely on markedness constraints, precisely as we saw in the discussion of Turkish voice alternations (26). Underspecification and lexical allomorphy share this phonology-driven property. The comparative virtues of the allomorphic model, however, become clear when we turn to prosodic alternations such as Dutch OSL, which resist treatment in terms of prespecification. 17.6.4
Constraining Allomorphy
The main faithfulness constraint relevant to OSL alternations is Wt-Ident-IO (30). Since OSL is an alternation that is part of a neutrast (see (34)), this constraint is undominated. In nonalternating stems, which have only a single lexical phonological representation (with either a short or long vowel), Wt-Ident-IO outranks all (potentially relevant) markedness constraints. (40) Faithfulness respected in single-input stems {/p t/} c
Wt-Ident-IO
+ p t @ p .t n
SWP
*mmm
No-Long-Vowel
*
e c
c
p t @ po.t n
*!
pot @ p .t n
*!
pot @ po.t n
*!*
* *
*
*
*
**
e
c
e c
e
Rene´ Kager
422
{/pot/}
Wt-Ident-IO
SWP
p t @ p .t n
*!*
*
p t @ po.t n
*
pot @ p .t n
*
*mmm
No-Long-Vowel
* *
e c
c
e
c
e c
+ pot @ po.t n
*
*
*
**
e
Stress in plurals based on monosyllabic stems (as in (43)) invariably falls on the stem. Since closed syllables and syllables containing long vowels are bimoraic, SWP is violated only by candidates with monomoraic stressed syllables. For nonalternating stems, violations of SWP are irrelevant, due to high-ranking Wt-Ident-IO. But for stems with length allomorphs, satisfaction of Wt-Ident-IO becomes vacuous— selection of allomorphs comes to depend on markedness constraints only. (41) Alternating stems: Emergence of the unmarked {/sl t/1 @ /slot/2 }
Wt-Ident-IO
c
sl t1 @ sl .t n1
SWP
*mmm
No-Long-Vowel
*!
e c
c
+ sl t1 @ slo.t n2
*
e
c
slot2 @ sl .t n1
*!
e c
slot2 @ slo.t n2
*
*
*!
**
e
The tableau compares only candidate pairs that satisfy Wt-Ident-IO, since, as I argued above, the availability of lexical length allomorphs allows every surface alternant to derive from a lexical counterpart of matching length. In e¤ect, IOFaithfulness to length is bypassed by a judicious choice of lexical allomorphs. This analysis raises a number of important general questions. First, how does allomorphy predict phonologically driven distribution of allomorphs? That is, why is the distribution not the reverse, with the long allomorph in a closed syllable and the short allomorph in an open syllable (*sl[o]t @ *sl[]ten)? This follows from the natural assumption that allomorphs are not subcategorized for phonological context, so that any phonological context sensitivity must be due to markedness constraints (Kager 1996, 1999). In the alternating candidate paradigms in (41), optimal [slt1 @ slo.t n2 ] and suboptimal [slot2 @ sl.t n1 ], the violation marks incurred by the markedness constraints for the optimal paradigm form a proper subset of those of the suboptimal paradigm. The suboptimal paradigm is harmonically bounded: it is never selected, regardless of ranking (Prince 1998). e
e
Lexical Irregularity and the Typology of Contrast
423
Second, can this theory capture any (consistently) nonalternating aspects of allomorphs? For example, OSL is never accompanied by alternations of rounding, as in hypothetical *[slt]1 @ [sla.t n]2 . From a diachronic angle, the explanation is evident: all lexical alternations start out as transparently predictable by markedness constraints (Kiparsky 1982b, Booij, this volume). In the absence of a transparent rounding process, why should lexical rounding alternations ever develop? This diachronic account, however, fails to explain the systematic lack of synchronic rounding alternations under OSL. If allomorphy is driven by contextual markedness constraints, then the lack of allomorphy for a feature [F] may be simply interpreted to reflect the lack of the relevant contextual markedness constraint in UG. That is, lexical allomorphy would be leveled out by general markedness constraints: a kind of ‘‘Stampean occultation.’’ In this particular case, Universal Grammar (UG) may happen to contain no markedness constraints linking roundness to OSL’s canonical prosodic context (syllabification, stress). This account, however, does not generalize. Alternatively, we may assume surface allomorphy to be checked by special constraints maximizing the phonological similarity between members of a paradigm— OO-Faithfulness constraints (Burzio 1996; Benua 1997). For example, the lack of rounding alternations accompanying OSL is captured by a su‰ciently high-ranked identity constraint. e
(42) IDENT-OO[round] Corresponding segments in output forms agree in values of [round]. For the hypothetical set of lexical allomorphs {/slt/1 @ /slat/2 } alternating in rounding, this alternation would not be able to surface, being blocked by the ranking Ident-OO[round] M-Specific [round]. (43) Enforcing paradigm uniformity in allomorphy {/sl t/1 @ /slat/2 }
Ident-OO [round]
c
+ sl t1 @ slot n2
*
e
c
sl t1 @ slat n2
M-Specific [round]
*!
e
c
Presumably, this ranking comes at minimal cost to the learner, assuming all OOFaithfulness constraints to be undominated in the initial state (Hayes 1999b). Without rounding alternations at the surface, the learner has no positive evidence to demote Ident-OO[round], and so it will remain undominated. Third, can we capture obligatory alternating aspects of allomorphs? We find, for example, that allomorphy is structure preserving in the sense that specifications for tense/lax automatically alternate along with length in OSL stems, matching the
Rene´ Kager
424
phonotactics of the Dutch vowel system. Consider, for example, the neutralization of height under OSL (briefly discussed above), which is explained by inviolate phonotactics of Dutch. Recall that short high [I] alternates with long mid [e] (e.g., schip @ schepen), implying a neutralization of height under OSL; short mid [] also alternates with [e] (e.g., weg @ wegen). This neutralization is due to a high-ranked constraint HighV-m: high vowels are short (Gussenhoven, this volume). Diagram (44) shows neutralization of height in relation to length and tenseness. (44) Neutralization of vowel height under OSL lax tense high I i mid ! e
Since the allomorphy is driven by quantitative constraints, hypothetical heightpreserving tenseness alternations such as *sch[I] p @ sch[i] pen simply fall short of meeting the quantitative goal, while height-preserving length alternations such as *sch[I] p @ sch[i] pen would be phonotactically illegitimate. How can we capture this idea by ranked constraints? What we must show is that hypothetical allomorph pairs such as those mentioned above would not be able to surface. To limit the set of possible allomorphs, the grammar must obscure the lexical input, as in any standard case of allophonic variation. First, tableau (45) shows how a height alternation accompanying OSL {schip @ schepen} is grammatically enforced, licensing the lexical paradigm {/sxIp/1 @ /sxep/2 }. After eliminating the topmost paradigm because of its long lax [i], two paradigms are fed into Ident-IO[high], which guarantees faithfulness to the input values of [high] in both surface allomorphs. (45) Grammar licenses alternation of height in input paradigm
*
e
+ sxIp1 @ sxe.p n2
*
e
sxp1 @ sxe.p n2
Ident-OO [high]
Ident-IO [high]
*!
*mmm
sxIp1 @ sxi.p n2
SWP
HighV-m
Wt-Ident-IO {/sxIp/1 @ /sxep/2 }
*!
e
Lexical Irregularity and the Typology of Contrast
425
This o¤ers an argument for the ranking Ident-IO[high] Ident-OO[high]. Any leveling of height (*[sxp1 @ sxe.p n2 ]) is rejected by IO-Faithfulness at the expense of a violation of OO-Faithfulness. (Of course, leveled height in input paradigms will be respected, as w[]g @ w[e.]gen shows.) This ranking correctly predicts that height alternations are impossible for input paradigms that are stable in length. Assuming a hypothetical lexical paradigm {/sxIp/1 @ /sxp/2 }, with leveled length but alternating height, the grammar rejects the height alternation since the distribution of alternants is undetermined by quantity-sensitive constraints. First, any attempt to introduce alternation of length fails on undominated Wt-Ident-IO. Next, the set of remaining candidate paradigms evaluated by Ident-IO[high] is homogeneous in the sense that all have leveled length but alternating height. Each of these paradigms is rooted in a pair of lexical allomorphs that can match their height, bypassing any e¤ects of Ident-IO[high]. The leveling of height in output paradigms is then due to Ident-OO[high]. e
(46) Grammar levels out alternating height in pairs of short allomorphs Ident-OO [high]
Ident-IO [high]
*mmm
SWP
HighV-m
Wt-Ident-IO {/sxIp/1 @ /sxp/2 } + sxIp1 @ sxI.p n1
*
sxIp1 @ sx.p n2
*
*!
sxp2 @ sxI.p n1
*
*!
+ sxp2 @ sx.p n2
*
e
e e
e
sxIp1 @ sxe.p n2
*!
*
e
In tableau (45) we saw that the lexical length allomorphy indirectly licenses a height alternation, which can be said to be parasitic on the length alternation. We now see that if the lexical paradigm is leveled in length, pressure for alternation from quantity-sensitive constraints drops, leading to leveling of height by paradigmatic constraints. More generally, this theory predicts that alternation of a property A may be confined to specific contexts, in which another property B alternates. In Dutch, for example, an alternation of length indirectly licenses an alternation of height. This
Rene´ Kager
426
parasitic licensing of one alternation by another alternation is an interesting result, since it suggests an account of nonderived environment blocking (NDEB) e¤ects (Kiparsky 1993a) in which ‘‘derivedness’’ is due to ‘‘prior application of a phonological rule.’’22 There is an empirical prediction here that NDEB e¤ects are found only in alternations that display the syndrome of properties of neutrasts: contrastiveness and lexical irregularity. We are now in a position to explain the lack of hypothetical OSL alternations such as [I] @ [i], with alternation of length and leveled height. Since any long vowel allomorph with [i] runs into high-ranked HighV-m, either its quantity or height will be obscured. The long allomorph’s quantity is preserved since it satisfies SWP in an open syllable; hence its vowel is lowered to [e]. Minimal violation of Ident-IO[high] predicts that the optimal output paradigm is {[sxIp]1 @ [sxe.p n]2 }. e
(47) Grammar imposes alternation of height in pairs of length allomorphs
e
e
+ sxIp1 @ sxe.p n2 e
sxp1 @ sxe.p n2
Ident-OO [high]
sxIp1 @ sxI.p n1
Ident-IO [high]
sxIp1 @ sxi.p n2
*mmm
SWP
HighV-m
Wt-Ident-IO {/sxIp/1 @ /sxip/2 }
*
*
*! *!
**!
e
The grammar enforces an alternation in height that is not lexically specified. This explains the neutralization of vowel height accompanying OSL, where high and mid short vowels [I, ] both alternate with mid long [e] (e.g., schip @ schepen and weg @ wegen). Finally, let us consider the fate of a hypothetical lexical allomorph pair {/sxIp/1 @ /sxip/2 } leveled in length and height while alternating in [tense]. The explanation for its ill-formedness is analogous to the leveling of height in tableau (46). The grammar rejects any alternation of tenseness that is unaccompanied by a quantity alternation and returns only tenseness-leveled output paradigms: either lax (a pattern attested in kip @ kippen ‘chicken(s)’) or tense (a pattern attested in iep @ iepen ‘elm(s)’).23
Lexical Irregularity and the Typology of Contrast
427
(48) Grammar levels out alternating tenseness in input paradigm Ident-OO [tense]
Ident-IO [tense]
*mmm
SWP
HighV-m
Wt-Ident-IO {/sxIp/1 @ /sxip/2 } + sxIp1 @ sxI.p n1
*
sxIp1 @ sxi.p n2
*
*!
sxip2 @ sxI.p n1
*
*!
+ sxip2 @ sxi.p n2
*
e
e e
e
This concludes the discussion of phonotactic restrictions on allomorphic patterns. In sum, I have argued that the theory of lexical allomorphy advanced here is perfectly capable of capturing generalizations about obligatory alternations, leveling, and neutralization. Quite remarkably, this result was achieved without diacritic means, and without compromising Richness of the Base.24 17.7
The Factorial Typology of Allomorphy
We now have all the ingredients that we need to compute the factorial typology of contrast in a theory of listed allomorphy. Four constraint types will be reranked in the factorial typology. (49) a. Markedness constraints MG *[aF] MS *[aF] / X Y b. Faithfulness constraints IO-Faith Input value of [F] equals output value of [F] OO-Faith No alternations of [F], hence *{[aF] @ [aF]} For reasons of perspicuity I have stated constraints in terms of values of [F]. Recall, however, that lexical allomorphy is crucially motivated by prosodic alternations that are problematic in a (segmental) underspecification model, as shown in sections 17.5 and 17.6. The full factorial typology is below. (50) Factorial typology of allomorphy a. Neutrast: IO-Faith MS MG , OO-Faith [aF]G @ [aF]S [aF]G @ [aF]S [aF]G @ [aF]S
Rene´ Kager
428
b. Full contrast: IO-Faith, OO-Faith MG , MS [aF]G @ [aF]S [aF]G @ [aF]S c. Contextual neutralization: MS IO-Faith MG , OO-Faith [aF]G @ [aF]S [aF]G @ [aF]S d. Total neutralization I: MG , OO-Faith IO-Faith, MS [aF]G @ [aF]S e. Total neutralization II: MS , OO-Faith IO-Faith, MG [aF]G @ [aF]S f. Complementary distribution: MS MG IO-Faith, OO-Faith [aF]G @ [aF]S This is essentially the original four-way typology of section 17.2, supplemented by the two additional patterns discussed here: neutrast as distinguished from full contrast by relative ranking of OO-Faith, and Total Neutralization II, which is a Morpheme Structure Constraint (McCarthy 1998). In Total Neutralization II, OO-identity fixes a single (marked) value [aF] for all morphemes occurring in both contexts, due to domination of a specific markedness constraint over the general constraint. One pattern is incorrectly excluded by the typology. (51) Positional Licensing
In Positional Licensing a contrast occurs in the specific context that is neutralized in the general context. An example is the restriction of long vowels to initial syllables (Steriade 1995a; Beckman 1998). Accordingly, Beckman (1998) proposes the faithfulness constraint format below.25 (52) Ident-IO([F], P) An output segment Standing in Position P has the same value for [F] as its input correspondent. Inclusion of Positional Faithfulness constraints in the typology adds no new patterns. Although the nonalternating pattern of Positional Licensing (51) is generated, it is predicted that there is no neutrast corresponding to this alternating pattern. There are only two possible outcomes for a ternary (allomorphic) input.
Lexical Irregularity and the Typology of Contrast
429
(53) Positional Faithfulness: Ident-IO([F], P) MG Ident-IO([F]) /aF/
Ident-IO([F], P)
MG
Ident-IO([F])
*!
*
*
*!
*
**
**
+ [aF]G @ [aF]S [aF]G @ [aF]S [aF]G @ [aF]S [aF]G @ [aF]S
*!
/aF/ [aF]G @ [aF]S
*!
+ [aF]G @ [aF]S [aF]G @ [aF]S [aF]G @ [aF]S
*!
** *
*
*
*
**!
/aF @ aF/ + [aF]G @ [aF]S [aF]G @ [aF]S
*!
[aF]G @ [aF]S
*!
[aF]G @ [aF]S
**!
This means that neutrasts cannot be phonologically driven by positional faithfulness, a prediction that should be tested against actual cases. 17.8
Conclusions
This chapter has explored lexical irregularity from the perspective of the typology of contrast. First, I observed that phonological irregularity with respect to a feature [F] presumes the occurrence of a contrast [aF] @ [aF] in the general context. This echoes a theorem of LP replicable under allomorphic OT as IO-Faith M-General. Second, ‘‘lexical irregularities’’ are situations of neutrast, a three-way division among morphemes with respect to an alternation. Richness of the Base predicts two kinds of input ternarity: archiphonemic underspecification (Inkelas 1995; Inkelas, Orgun, et al. 1997) and lexical allomorphy, in which the input supplies alternating pairs. Lexical allomorphy may be seen to include underspecification as a special case, although Richness of the Base predicts that both are possible. Hence, allomorphy need not
430
Rene´ Kager
replace underspecification of featural alternations (such as Turkish devoicing), but it crucially does so in prosodic neutrasts (such as Dutch open syllable lengthening). Third, I argued that phonologically conditioned allomorphy is driven by a pair of markedness constraints, with a contextual constraint dominating a general constraint (IO-Faith M-Specific M-General). Fourth, phonology-driven allomorphy has the property of invisibility to IO-Faithfulness constraints, due to the split inputs assumption, the counterpart of cancellation in underspecified inputs. OO-Faithfulness levels out lexical allomorphy, however, if OO-Faithfulness M-Specific. Fifth, and perhaps most important, the factorial typology of contrast is not weakened by adding lexical allomorphy and, presumably, not even by adding Positional Faithfulness. Notes Some ideas presented in this chapter emerged from discussions with Marc Verhijde and Wim Zonneveld at the UiL-OTS in 1996. For various kinds of input to earlier presentations of this chapter, I wish to thank John Alderete, Laura Benua, Geert Booij, Luigi Burzio, Paula Fikkert, Janet Grijzenhout, Harry van der Hulst, John Kingston, John McCarthy, Marc van Oostendorp, Alan Prince, Iggy Roca, Markus Walther, Dieter Wunderlich, Draga Zec, as well as audiences at Heinrich Heine University Du¨sseldorf (linguistics colloquium, December 1998), NELS 30 (Rutgers University, October 1999), University of Essex (linguistics colloquium, December 1999), and University of Postdam (workshop on ‘‘Conflicting Rules,’’ December 1999). Finally, my special thanks go to Sharon Inkelas and Joe Pater for their thorough comments. Needless to say, I am fully responsible for any errors. 1. A theory of positional neutralization should account for crosslinguistically recurrent patterns of neutralization, relating these to universal properties of articulation and perception (Flemming 1995; Steriade 1995a). 2. See Zoll (1996) for an alternative view couched in licensing by positional markedness. 3. A full factorial typology of three constraints contains six logical possibilities, but only four ranking emerge as distinct. This is because both faithfulness and general markedness impose a total ranking on all candidates. Accordingly, these constraints, when undominated, obscure the relative ranking of any constraint they dominate. 4. Kiparsky (1993a) allows some degree of ternarity, in the sense that the lexicon is free to contain segments specified for all three values (plus, minus, or zero) of a feature [F]. However, in any given context, he still adheres to binarity by limiting underlying contrasts to two (for example, plus and zero, or minus and zero). 5. This might be stated more generally as intersonorant voicing, *[voice]/[þson]
[þson].
6. Alternatively, we may assume that feature-filling mappings go unpenalized by featural faithfulness. That is, Ident-IO[F] is violated only if the output changes a specified lexical value of [F], while no violation mark is assigned if the lexical input lacks specification for [F]. 7. Lexicon Optimization leads the learner to posit input [voice] in nonalternating voiceless stems, while for alternating stems, either [þvoice] or [0voice] su‰ces. Inkelas argues that alternating morphemes are lexically underspecified, assuming feature-changing mappings to be more costly in terms of faithfulness constraints than feature-filling ones. See previous note.
Lexical Irregularity and the Typology of Contrast
431
8. Implicitly, I assume Inkelas’s suggestion (p.c.) of ranking Surface-Spec (the requirement that output forms be fully specified) above Ident-IO[voice]. 9. Other exceptionally long vowels before ‘‘shortening’’ su‰xes (from Fudge 1984, 222) occur in apical, vibrative, migratory, vibratory, cyclical, psychical, amenity, denotative, restorative, codify, glorify, nodical, probity. 10. Lahiri and Fikkert (1999) show that, diachronically, length alternations in TSS were due to wholesale borrowing of Romance loans, both su‰xed and unsu‰xed. Lahiri and Fikkert (1999, 229) argue that ‘‘[o]nly later, when these words came to be derivationally related, were quantity alternations observable with TSS operating as a constraint dictated by the prosodic structure of the modern language.’’ This result is fully compatible with my proposal. 11. A contrast between prelinked versus floating segments has been proposed in the literature on ghost segments (Zoll 1996). The alternative is to prespecify higher-level prosodic structures on top of nonalternating segments. For example, Inkelas and Orgun (1995), to account for lexical exceptions to Turkish intervocalic velar deletion, assume prespecification of syllable structure (under extrasyllabicity of stem-final consonants and cyclicity), Inkelas, Orgun, et al. (1997) acknowledge, however, that this analysis cannot be generalized to alternations involving nonperipheral elements. 12. Contrasts between prelinked versus floating segments have been proposed in the literature on ghost segments (Zoll 1996). 13. For example, to account for lexical exceptions to Turkish intervocalic velar deletion, Inkelas and Orgun (1995) assume syllable structure to be prespecified in the input (under extrasyllabicity of stem final consonants and cyclic syllabification). However, Inkelas, Orgun, et al. (1997) acknowledge that this analysis is di‰cult to generalize to alternations involving nonperipheral elements. Another problem for prespecification is posed by exceptions to minimality conditions, for example, exceptional monomoraic words in Turkish (Inkelas and Orgun 1995) and Japanese (Itoˆ 1990). Kiparsky (1993a, 304) assumes prespecification of a monomoraic foot, while canonical bimoraic stems are unspecified for foot structure. It is unclear, however, why prespecified monomoraic feet would block canonical lengthening while monomoraic vowels unspecified for foot structure freely undergo it. Thanks go to Sharon Inkelas for pointing this out. 14. Alternations of voice, exemplified by ‘boss(es)’ and ‘glass(es)’ in (33), are due to final devoicing and thus independent of the length alternatives. 15. If intervocalic consonants after short lax vowels are ambisyllabic, a standard assumption for Dutch phonology (Van der Hulst 1984; Kager 1989; Booij 1995; Gussenhoven this volume), and if closed syllables are heavy (same references), then a third constraint is at play to select a tense long vowel in an open syllable, rejecting a short lax vowel plus ambisyllabic coda. This constraint, No-Coda, militates against closed syllables. See Gussenhoven (this volume) for an analysis of Dutch vowel length and word stress on di¤erent assumptions. 16. Although most OSL stems are monosyllabic nouns, a few polysyllabic nouns occur (oorlog @ oorlogen ‘war(s)’, hertog @ hertogen ‘duke(s)’, both with a conditioning secondary stress on the long vowels), plus a handful of verbs (e.g., gaf @ gaven ‘gave-sg/pl’, kom @ komen ‘come-sg/pl’, kwam @ kwamen ‘came-sg/pl’) and even a single adjective (grof @ grove ‘crude’). 17. In Dutch, alternations of height only occur under OSL alternations.
432
Rene´ Kager
18. See Benua (1997) on overapplication e¤ects due to OO-Faithfulness. 19. Zonneveld (1978) places diminutives and denominal verbs in angled brackets in the OSL rule. However, this analysis fails to explain why long vowel allomorphs of OSL stems need not occur in diminutives, for example, slotje ‘lock-dim.’, godje ‘god-dim.’. I will not go into idiosyncratic length alternations in OSL stems before su‰xes such as -ig (glas @ glazig, but nonalternating gebrek @ gebrekkig) and -elijk (stad @ stedelijk, but god @ goddelijk); see Booij (this volume). 20. Some of these researchers, more ambitiously, attempt to eliminate the notion of underlying representation from phonological theory (Burzio 1996, 2000; Steriade 1997; Hayes 1999). Perhaps the central issue is whether perfectly productive alternations should be dealt with in terms of listed allomorphs. I believe that the issue can be partly resolved, since the grammar constrains the notion of ‘‘possible allomorph’’ in a language, as we will see later. 21. Presumably, nonalternating aspects of allomorphs are collapsed in lexical representations, locating the alternation on the relevant segments, for example, /sx{I1 @ e2 }p/. See Walther (1999). 22. Lubowicz (1998) o¤ers an account of NDEB e¤ects based on constraint conjunction. 23. Whenever the input is leveled for tenseness, the output paradigm reflects this specification, licensing the contrast between lax kip @ kippen and tense iep @ iepen. But whenever [tense] alternates in the input, a leveling takes place toward one of both values, depending on general markedness constraints governing [tense]. 24. A separate issue that I will not discuss here is how to capture any morphologically governed types of allomorphy. Arguably, allomorphs may be subcategorized by morphological contexts, with constraints enforcing morpho-subcategorization overriding markedness constraints. 25. See Zoll (1996) for an alternative theory of Positional Licensing in terms of positive markedness constraints (e.g., ‘‘if a vowel is long, then it stands in s1 ’’).
18
Derived Environment Effects in Colloquial Helsinki Finnish
Arto Anttila
18.1
Introduction
Some phonological rules apply only in derived environments—that is, across a morpheme boundary or if fed by an earlier phonological rule—but are blocked elsewhere. This syndrome, known as Nonderived Environment Blocking (NDEB) (Kiparsky 1993a), is a traditional puzzle in generative phonology. Familiar textbook examples include English Trisyllabic Shortening (Chomsky and Halle 1968), which applies in the morphologically derived divı˘nþity (from divı¯ne) but not in the nonderived nı¯ghtingale. A much-discussed question concerning NDEB is the following: (1) What alternations are subject to NDEB? The answers to (1) have traditionally been given in terms of rule typologies: (2) a. NDEB is a property of nonautomatic neutralization rules (the Revised Alternation Condition, RAC; Kiparsky 1973e). b. NDEB is a property of cyclic rules (the Strict Cycle Condition, SCC; Mascaro´ 1976). c. NDEB is a property of lexical rules (the Elsewhere Condition; Kiparsky 1982d). d. NDEB is a property of structure-building rules operating on underspecified representations (Kiparsky 1993a; for earlier formulations see, e.g., Ringen 1975 and Borowsky 1986). Kiparsky (1993a) showed that rules exist that are both cyclic and lexical, but that nevertheless apply in nonderived environments, contradicting (2b) and (2c). The crucial evidence comes from the optional rule of Vowel Coalescence (henceforth VC) in Colloquial Helsinki Finnish. An example is given in (3): (3) Vowel Coalescence in Colloquial Helsinki Finnish /makea/ ‘sweet’ ! ma´.ke.a@ma´.kee
434
Arto Anttila
As Kiparsky demonstrated, this rule is both cyclic and lexical, yet it applies in nonderived environments. In this chapter, I take a closer look at the Finnish evidence in the light of a corpus of spoken Helsinki Finnish collected by Heikki Paunonen and his associates in the early 1970s and documented in Paunonen 1995. The corpus covers 126 speakers grouped by age, sex, social class, and neighborhood, about 500,000 word forms in all. The raw corpus is available at the University of Helsinki Language Corpus Server at http://www.ling.helsinki.fi/uhlcs. For the purposes of the present study, all the environments relevant to Vowel Coalescence were tagged phonologically and morphologically by the present author, approximately 13,000 vowel sequences in all.1 The main observation is that VC shows blocking e¤ects, but of a peculiar kind. First, blocking is limited to certain phonological environments only. Second, nonderived environments show categorical blocking (absence of alternation), whereas derived environments show quantitative blocking. Put di¤erently: (4) a. Nonderived environments VC is categorically blocked in nonderived environments if the structural change is highly marked; if the structural change is unmarked, VC may apply even in nonderived environments. b. Derived environments VC is quantitatively dispreferred in derived environments if the structural change is highly marked; if the structural change is unmarked, VC is quantitatively preferred. These generalizations do not make immediate sense in rule-based phonology. If NDEB is a property of a class of phonological rules, then one would certainly expect a single rule either to show or not show NDEB. Instead, (4a) states that one and the same rule sometimes does, sometimes does not show NDEB, depending on markedness. (4b) is even more puzzling because a rule’s application probability is usually not taken to be a matter of grammar at all. After considering the Finnish evidence in detail, I come to the conclusion that derived environment behavior cannot be tied to phonological rules of any kind. The reason is that the very same rule may show NDEB in certain contexts, but not in others, depending on markedness. This suggests that reference to global properties of grammar is inevitable, and that rules—that is, specific phonological processes— cannot serve as the locus of explanation. I then construct an analysis of the Finnish facts in terms of Optimality Theory (Prince and Smolensky 1993). More specifically, I propose that NDEB results from root faithfulness (McCarthy and Prince 1995) that is relativized to markedness. A simple partially ordered Optimality-Theoretic grammar is constructed that predicts both the categorical and quantitative blocking pat-
Derived Environment Effects in Colloquial Helsinki Finnish
435
terns observed in the data. Under this view, NDEB turns out to be a categorical special case of quantitative dispreference. I am deeply indebted to Paunonen’s (1995) valuable earlier study of VC based on a slightly di¤erent, but largely coextensive subset of the same corpus. While my interpretation of the facts is quite di¤erent from his, all of Paunonen’s factual observations and most of his generalizations still stand under my reanalysis. 18.2
The Categorical Aspects of VC
Under the most general formulation, VC applies optionally to any heterosyllabic two-vowel sequence, whether underlying or derived, where both vowels are unstressed and the second vowel is [þlow], assimilating the quality of the second vowel to that of the first.2 (5) V1 .V2 ! V1 : where V2 ¼ [þlow] ea ! ee ea¨ ! ee oa ! oo o¨a¨ ! o¨o¨ ua ! uu ya¨ ! yy ia ! ii ia¨ ! ii
ma´.ke.a @ ma´.kee kı´.pe.a¨ @ kı´.pee a´i.no.a @ a´i.noo ´ r.jo¨.-a¨ @ Y ´ r.jo¨-o¨ Y ka´.tu.-a @ ka´.tu-u hy´o¨.ty.-a¨ @ hy´o¨.ty-y la´.si.-a @ la´.si-i ke´n.k-i.-a¨ @ ke´n.ki-i
‘sweet’ ‘sick’ ‘only’ ‘George-par’ ‘street-par’ ‘advantage-par’ ‘glass-par’ ‘shoe-pl-par’
VC is genuinely optional in the sense of being variable within an individual. In the following short dialogue we have four instances of coalescence and four instances of noncoalescence. The same speaker sometimes does, sometimes does not apply VC, apparently unpredictably. (6) OH: Millasii ihmisii siel ka¨y judoomassa? /millas-i-a/, /ihmis-i-a¨/, /judoa-ma-ssa/ ‘What sort of people practice judo there?’ JS: Siel ka¨y iha, nuoria ja vanhojaki. /nuor-i-a/ ‘Some are really young, but there are old people too.’ OH: Miehia¨ naisia? /mieh-i-a¨/, /nais-i-a/ ‘Men? Women?’ JS: Joo miehii ja naisia. /mieh-i-a¨/, /nais-i-a/ ‘Yes, men and women.’
436
Arto Anttila
As the last line shows, both the coalesced and hiatus variants may occur within the same NP; more examples are cited in Paunonen 1995, 106–107. Even more strikingly, both variants may occur within the same word. The word useampia ‘manycomp-pl-par’ contains both /ea/ and /ia/. All four logical possibilities are attested: (7) /usea-mp-i-a/ ‘many-comp-pl-par’ a. useampia (no VC, no VC) b. useempia (VC, no VC) c. useampii (no VC, VC) d. useempii (VC, VC) However, the randomness is merely apparent. We now turn to the ironclad structural regularities, both categorical and quantitative, that govern variation. The categorical regularities are summarized in (8). The table shows that the NDEB e¤ect emerges clearly when the first vowel is high, but not so clearly when it is mid. Here ‘‘þ’’ means VC may apply, ‘‘’’ that VC may not apply; ‘‘I’’ represents a high vowel /i, u, y/, ‘‘E’’ a mid vowel /e, o, o¨/, and ‘‘A’’ a low vowel /a, a¨/. (8) The general picture
Nonderived Derived
Noun Adjective Noun Adjective
IA
EA
þ þ
=þ þ þ þ
(9) The first generalization VC is blocked in nonderived environments if the first vowel is [þhigh] (IA). (10) a. minia¨@*minii mini-a¨@mini-i b. rasia@*rasii lasi-a@lasi-i c. saippua@*saippuu hattu-a@hattu-u d. Po¨ytya¨@*Po¨ytyy lo¨yly-a¨@lo¨yly-y
‘daughter-in-law’ ‘mini-par’ ‘box’ ‘glass-par’ ‘soap’ ‘hat-par’ (a place name) ‘steam-par’
As (10) shows, derived IA-sequences coalesce, nonderived IA-sequences do not. What we see here is classical derived environment behavior. Now, if NDEB is a property of rules and the IA@II alternation exhibits NDEB, then one would expect the EA@EE alternation to do the same, for certainly, under any reasonable rulebased analysis, both alternations should be attributed to one and the same phonological rule. But this is not what we find.
Derived Environment Effects in Colloquial Helsinki Finnish
437
(11) The second generalization VC applies across the board if the first vowel is [high] (EA). (12) a. hopea@hopee ove-a@ove-e b. ainoa@ainoo Aino-a@Aino-o c. pimea¨@pimee nime-a¨@nime-e
‘silver’ ‘door-par’ ‘only’ (a female.name-par) ‘dark’ ‘name-par’
Contrary to all expectations, VC applies to hundreds of nonderived EA-stems, all common native words such as pimea¨ ‘dark’. It is exactly this class of words that crucially proves the independence of cyclicity/lexicality and NDEB (Kiparsky 1993a, 280–282). The fundamental generalization here is that IA-stems show NDEB, EAstems do not. In other words, the derived environment e¤ect is split along a phonological dimension. However, there is a group of EA-stems that do show the expected NDEB e¤ect: recently borrowed nouns.3 (13) An exception to the second generalization VC is blocked in nonderived environments if the first vowel is [high] (EA) and the word is a recently borrowed noun. (14) a. idea@*idee forte-a@forte-e b. Korea@*Koree Palme-a@Palme-e c. komitea@*komitee cumlaude-a@cumlaude-e d. teodikea@*teodikee ukulele-a@ukulele-e
‘idea’ ‘forte’ ‘Korea’ ‘Palme-par’ ‘committee’ ‘cum laude–par’ ‘theodicy’ ‘ukulele-par’
Let us briefly speculate on what could have motivated such an exceptional pattern. We begin from the following observation: of stems ending in EA or IA, nouns canonically end in IA, adjectives in EA. The following statistics are extracted from the electronic version of Nykysuomen sanakirja (Dictionary of Modern Finnish) (Sadeniemi 1973): -IA
(15) Nouns Adjectives
92.6% 2.2%
-EA [651] [6]
7.4% 97.8%
[52] [273]
100% 100%
438
Arto Anttila
Let us assume that, for purely phonological reasons, IA resists coalescence more than EA. Since the vast majority of noun stems end in IA and the vast majority of adjective stems in EA (for historical reasons that are irrelevant here), we have a plausible basis for analogy: the failure of coalescence in IA-sequences has been reanalyzed as failure of coalescence in nouns. Under this scenario, incoming EA-nouns such as idea ‘idea, n’ would simply be imitating the majority of nouns, and those happen to be mostly IA-stems that have independent phonological reasons to resist coalescence. As is typical of analogical change, the regularization does not apply across the board: a small number of native EA-nouns resist the change and only new words succumb to analogy.4 A parallel noun/adjective split can be observed in two other rules that a¤ect stemfinal low vowels: A-deletion (16a), which deletes stem-final [þlow] vowels, and Amutation (16b), which changes them to rounded mid vowels. Both rules apply before the past tense or plural /-i-/: (16) a. a, a¨ ! j b. a, a¨ ! o, o¨
/muna-i-ssa/ ! mun-i-ssa /kana-i-ssa/ ! kano-i-ssa
‘egg-pl-ine’ ‘hen-pl-ine’
In (16), the choice between the two rules is straightforwardly phonological: deletion occurs when the vowel preceding the stem-final one is rounded, mutation elsewhere. But in trisyllabic stems the phonological condition is lost and morphology takes charge: nouns mutate, adjectives delete.5 (17) /kihara/ ‘curl, n’ ! kiharo-i-ssa /kihara/ ‘curly, adj’ ! kihar-i-ssa Also in this case, the morphological condition appears to be a recent analogical development. To the best of my knowledge, it was first observed by G. Karlsson (1978), although weak quantitative reflexes of it are already visible in Nykysuomen sanakirja (Dictionary of Modern Finnish). A possible alternative explanation begins with the observation that virtually all recent borrowings are nouns. Thus, one might suggest that the exceptional pattern has nothing to do with the noun/adjective distinction, but instead all recently borrowed stems follow the same pattern—that is, we have a special loanword phonology. The behavior of recently borrowed EA-final adjectives would decide between the two hypotheses, but to the best of my knowledge they do not exist. This story has the unfortunate drawback of not explaining why recent borrowings should exhibit NDEB, rather than not. After all, both models are abundantly available in the language.6 In sum, the Colloquial Helsinki Vowel Coalescence rule, which is both cyclic and lexical, turns out to exhibit an unusual split NDEB pattern: in IA-stems, the rule is blocked in nonderived environments; in EA-stems it applies generally, except in recently borrowed nouns that show NDEB.
Derived Environment Effects in Colloquial Helsinki Finnish
439
Another way to summarize the data pattern in (8) is to state three implications: (18) Summary a. Coalescence in IA implies coalescence in EA. b. Coalescence in nonderived environments implies coalescence in derived environments. c. Coalescence in nouns implies coalescence in adjectives. These three implications capture the distribution of pluses and minuses in (8). (18b) states the weak form of NDEB observed in the Finnish data: if coalescence applies in nonderived environments in some phonologically and morphologically defined class of stems, it also applies in derived environments in the same class of stems. 18.3
The Quantitative Aspects of VC
The most striking aspect of the Helsinki data is that the categorical blocking pattern turns out to have a quantitative analogue.7 The generalization (19a) is due to Paunonen (1995, 111); the generalizations (19b, c) are mine. (19) a. Coalescence is at least as common in EA as in IA, other things being equal. b. Coalescence is at least as common in derived as in nonderived environments, other things being equal. c. Coalescence is at least as common in adjectives as in nouns, other things being equal. First, consider (19a). The words lasi-a ‘glass-par’ and ove-a ‘door-par’ are both derived nouns. The only di¤erence is phonological: the first word ends in -ia, the second in -ea. VC applies optionally to both, but is systematically more common in the ove-a type (see (21) for actual frequencies). As Paunonen shows, the generalization (19a) is extremely well supported by various types of data. Outside Helsinki, it is reflected in the dialect geography of Finland. In Helsinki, it holds of individual speaker groups, down to the level of idiolects, and also of various morphologically defined word classes, such as partitives and infinitives. In fact, even the behavior of individual words is consistent with the generalization. The word useampia ‘manycomp-pl-par’ behaves as one might expect: (20) /usea-mp-i-a/ ‘many-comp-pl-par’ a. b. c. d.
useampia useempia useempii useampii
(no VC, no VC) (VC, no VC) (VC, VC) (no VC, VC)
Frequency 22 5 3 1
440
Arto Anttila
The unexpected type is (20d) where IA coalesces, but EA does not. There is exactly one such example in the entire corpus. All other examples (thirty out of thirty-one) conform to Paunonen’s generalization.8 The morphological implications (19b) and (19c) also have quantitative reflexes. (19b) emerges in pairs like hopea ‘silver’ and ove-a ‘door-par’. Both are -ea-final nouns. The only di¤erence is morphological: the first word is nonderived, the second derived. VC applies optionally to both, but is systematically more common in the derived ove-a type (see (22) for actual frequencies). Finally, (19c) emerges in pairs like ov-i-a ‘door-pl-par, n’ and uus-i-a ‘new-pl-par, adj’. Both are -ia-final derived words, but the first is a noun, the second an adjective. Again, VC applies optionally to both, but is systematically more common in the uus-i-a type—that is, in adjectives (see (23) for actual frequencies). All these quantitative generalizations are robust. They hold in di¤erent age groups (old, middle-aged, young), social classes, neighborhoods (To¨o¨lo¨, So¨rna¨inen), and sexes. In other words, no matter how we pick a group of people, the same structural hierarchies emerge. For example, consider the distribution of coalescence by age:9 (21) The phonological e¤ect Old N ND EA IA N D EA IA A ND EA IA A D EA IA
3.3% 0% 24.3% 7.8% 49.3% 0% — 9.6%
(1/30) (0/337) (65/267) (147/1,886) (242/491) (0/80) — (146/1,519)
Middle-aged
Young
18.2% 0% 30.5% 11.4% 64.9% 0% — 16.8%
50% 0% 67.2% 43.5% 93.0% 0% — 64.2%
(4/22) (0/263) (60/197) (182/1,597) (334/515) (0/100) — (220/1,308)
(4/8) (0/247) (168/250) (685/1,576) (687/739) (0/81) — (923/1,437)
(22) The derived environment e¤ect Old N EA D ND N IA D ND A EA D ND A IA D ND
24.3% 3.3% 7.8% 0% — 49.3% 9.6% 0%
(65/267) (1/30) (147/1,886) (0/337) — (242/491) (146/1,519) (0/80)
Middle-aged
Young
30.5% 18.2% 11.4% 0% — 64.9% 16.8% 0%
67.2% 50% 43.5% 0% — 93.0% 64.2% 0%
(60/197) (4/22) (182/1,597) (0/263) — (334/515) (220/1,308) (0/100)
(168/250) (4/8) (685/1,576) (0/247) — (687/739) (923/1,437) (0/81)
Derived Environment Effects in Colloquial Helsinki Finnish
441
(23) The part-of-speech e¤ect Old ND EA A N ND IA A N D EA A N D IA A N
49.3% 3.3% 0% 0% — 24.3% 9.6% 7.8%
(242/491) (1/30) (0/80) (0/337) — (65/267) (146/1,519) (147/1,886)
Middle-aged
Young
64.9% 18.2% 0% 0% — 30.5% 16.8% 11.4%
93.0% 50% 0% 0% — 67.2% 64.2% 43.5%
(334/515) (4/22) (0/100) (0/263) — (60/197) (220/1,308) (182/1,597)
(687/739) (4/8) (0/81) (0/247) — (168/250) (923/1,437) (685/1,576)
The same phonological and morphological regularities emerge in each age group. Table (21) shows that -ea coalesces more than -ia, derivedness and part of speech being equal. Table (22) shows that derived words coalesce more than nonderived words, phonology and part of speech being equal. Finally, table (23) shows that adjectives coalesce more than nouns, derivedness and phonology being equal, except among nonderived ia-stems where coalescence is blocked across the board. In other words, the structural hierarchies hold perfectly. Crucially, their surface reflexes are sometimes categorical, sometimes quantitative. How did this pattern arise? Another look at the figures shows that age has a systematic e¤ect on coalescence frequencies: young speakers apply coalescence more than middle-aged apeakers, who in turn apply coalescence more than old speakers. Whether this reflects change in progress or age grading is not obvious. In his discussion of the possible historical scenarios, Paunonen (1995, 122–132) first considers the possibility that Vowel Coalescence reflects a historical change that started from EAwords, this being phonetically the most natural environment for coalescence, and moved on to IA-words along various morphological dimensions, which would explain the asymmetric relation between IA and EA. However, he hastens to point out that regular development is inconsistent with the demographic and dialectal facts. Spoken Helsinki Finnish emerged from a recent confluence of dialects that occurred over a very short period of time. Thus, it is all the more striking that the end result should be structurally so systematic, down to subtle quantitative tendencies. Paunonen (1995, 132) concludes that this can only be attributed to the influence of ‘‘language-internal causal connections’’ (sisa¨inen vaikutusyhteys), or put slightly di¤erently, (universal) grammar. To sum up, we have discovered three structural hierarchies that emerge consistently from the data. Their combined surface e¤ect is sometimes a categorical rule, sometimes a quantitative tendency.
442
Arto Anttila
(24) a. Coalescence is more likely in EA than in IA. The result may be categorical blocking: hopea@hopee ‘silver, n’ versus rasia@*rasii ‘box, n’. b. Coalescence is more likely in derived forms than in nonderived forms. The result may be categorical blocking: mini-a¨@mini-i ‘mini-par, n’ versus minia¨@*minii ‘daughter-in-law, n’. c. Coalescence is more likely in adjectives than in nouns. The result may be categorical blocking: korea@koree ‘beautiful, adj’ versus Korea@*Koree ‘Korea, n’. Table (25) shows the eight possible factor combinations and the observed numbers in the entire corpus. The factors favoring coalescence are underlined (EA, D, A). The coalescence rates are given in two ways: as ranges covering the three age groups and as the average coalescence rate in the entire corpus. (25) Obs1 % ¼ the average coalescence rate in the total corpus Obs2 % ¼ the observed range of variation
N N
EA EA
D ND
N N A A A A
IA IA EA EA IA IA
D ND D ND D ND
VC
Obs1 %
Obs2 %
Number of tokens
applies applies NDEB applies NDEB applies applies applies NDEB
41.0% 18.8% 0% 20.0% 0% — 72.4% 30.2% 0%
24.3–67.2% 3.4–66.7% 0% 7.8–43.5% 0% — 49.5–93.0% 9.6–64.2% 0%
714 48 (native) 12 (borrowing) 5,059 847 0 1,745 4,264 261
Categorical blocking emerges if coalescence is disfavored by at least two factors. To get blocking in EA-stems, we need both morphological criteria to weigh against coalescence and even then the pattern remains lexically restricted. In IA-stems blocking is more widespread; here nonderivedness is by itself su‰cient. In general, nonderivedness is not a su‰cient condition for blocking, but it is a necessary one, whereas being a noun is neither. In this sense, nonderivedness is the stronger of the two morphological conditions in disfavoring the rule’s application. I conclude that the total blocking of an alternation is the categorical limiting case of quantitative dispreference that occurs in phonologically and/or morphologically marked environments.
Derived Environment Effects in Colloquial Helsinki Finnish
18.4
443
An Optimality-Theoretic Analysis
I will now proceed to give an Optimality-Theoretic analysis of the unusual NDEB pattern in Colloquial Helsinki Finnish. This will entail taking a stand on two important questions: (26) a. What is the nature of NDEB? b. What is the grammatical status of quantitative regularities? First, I introduce the general approach by accounting for the purely phonological vowel-height e¤ect (EA versus IA) in terms of Optimality Theory (OT) (Prince and Smolensky 2004). The core of OT consists of two proposals: (i) grammars consist of potentially conflicting constraints, and (ii) conflicts are resolved by ranking the constraints in terms of their relative strength. It is commonly assumed that the ranking relation is a total order—that is, every constraint is ranked with respect to every other constraint. In this chapter, I will relax this requirement by allowing rankings that are genuine par-tial orders. The idea that grammatical constraints may remain mutually unranked goes back to Prince and Smolensky 2004, 61, which mentions ‘‘crucial nonranking’’ as a formal possibility provided by the theory for which no evidence existed at the time of writing (1993). Its empirical relevance was pointed out in Kiparsky’s (1993d) unpublished work on variation in English t,d-deletion and the idea has since been applied to several cases of variation outside English, such as Nagy and Reynolds 1997, Anttila 1997, Anttila and Cho 1998, and Ringen and Heina¨ma¨ki 1999, among others. Second, I turn to the morphological side of the problem. I consider four possible ways of integrating the derived environment e¤ect into the phonological analysis: (i) local conjunction of markedness and faithfulness (Łubowicz 2002); (ii) cophonologies—that is, derived and nonderived lexical items participate in di¤erent rankings; (iii) root faithfulness—in other words, roots are more resilient under markedness pressure than a‰xes (McCarthy and Prince 1995); (iv) prespecification—that is, alternating vowels are underspecified, nonalternating vowels (NDEB) are prespecified (Kiparsky 1993a; Inkelas 2000). I will choose root faithfulness as the most satisfactory of the four options. I then turn to the apparently very similar noun/ adjective e¤ect. While the empirical evidence is less than fully conclusive in this case, I suggest that general considerations favor a cophonology analysis. 18.4.1
The Phonological Effect
I first derive the vowel-height asymmetry, both in its categorical and quantitative manifestations. Dialects of the categorical type include Literary Finnish, where neither EA nor IA coalesces, and the system of old female upper-middle-class To¨o¨lo¨
444
Arto Anttila
residents (Old To¨o¨lo¨), who allow coalescence optionally in EA but never in IA (Paunonen 1995, 112). I start by assuming the following phonological constraints: (27) Constraints a. *EA Avoid /ea, oa, o¨a¨/ hiatus. b. *IA Avoid /ia, ua, ya¨/ hiatus. c. FAITH No coalescence.
(markedness) (markedness) (faithfulness)
I follow Paunonen in assuming that ea ! ee is a phonologically more natural process than ia ! ii. I express this as the ranking *EA *IA.10 This yields a typology of three totally ranked grammars. Two of the corresponding dialects are found in Paunonen 1995, 109 (Literary Finnish, General Ha¨me). While Paunonen’s descriptions of regional dialects are obviously simplifications, they will serve for illustrative purposes here. The third totally ranked grammar describes a potential dialect where coalescence is obligatory across the board: (28) No coalescence (conservative dialect, Literary Finnish) /-ea/
Faith
+ 1a. /-ea/ 1b. /-ee/
*EA *
*!
+ 2a. /-ia/ 2b. /-ii/
*IA
* *!
(29) Coalescence only in mid vowels (General Ha¨me) /-ea/ 1a. /-ea/ + 1b. /-ee/
*EA
Faith
*! *
+ 2a. /-ia/ 2b. /-ii/
*IA
* *!
Derived Environment Effects in Colloquial Helsinki Finnish
445
(30) Only coalescence (advanced dialect, not yet found) /-ea/ 1a. /-ea/
*EA
*IA
*!
+ 1b. /-ee/ 2a. /-ia/ + 2b. /-ii/
Faith
* *! *
This simple three-constraint system generalizes straightforwardly to quantitative regularities if we assume that any partial ordering of constraints is a possible grammar. Let C be the set of constraints and R a binary ranking relation with the following properties: (31) a. Irreflexivity R is irreflexive if and only if for every x in C, R contains no ordered pair hx, xi with identical first and second members. (No constraint can be ranked above or below itself.) b. Asymmetry R is asymmetric if and only if for any ordered pair hx, yi in R the pair h y, xi is not in R. (If x is ranked above y, it cannot be ranked below y.) c. Transitivity R is transitive if and only if for all ordered pairs hx, yi and h y, zi in R, the pair hx, zi is also in R. (If x is ranked above y and y is ranked above z, then x is ranked above z.) The properties (31a–c) define a partial order. In Optimality Theory, grammars are usually assumed to have the additional property of connectedness, which requires that, for every two distinct constraints x and y in C, hx, yi A R or h y, xi A R—that is, every constraint is ranked with respect to every other constraint. This defines a total ordering, or a tableau. If we assume that any partial ordering is a possible grammar, we get three more grammars. This includes the grammars for two additional dialects, Old To¨o¨lo¨ and Western Uusimaa (Paunonen 1995, 109), as well as a grammar labeled ‘‘Most speakers.’’ The resulting typology of six grammars is displayed in diagram (32). The grammars are described as sets of ordered constraint pairs and grouped into natural classes by shared rankings.
446
Arto Anttila
(32)
Most speakers *EA *IA [ea@ee], [ia@ii] Old To¨o¨lo¨ *EA *IA Faith *IA [ea@ee], [ia]
Literary Finnish *EA *IA Faith *EA Faith *IA [ea], [ia]
Western Uusimaa *EA *IA *EA Faith [ee], [ia@ii]
General Ha¨me *EA *IA *EA Faith Faith *IA [ee], [ia]
(Advanced dialect) *EA *IA *EA Faith *IA Faith [ee], [ii]
An important consequence of adopting partially ordered grammars is that it becomes possible for grammars to include other grammars. This is visually represented in diagram (32) in terms of immediate dominance. Consider the grammars labeled ‘‘Most speakers’’ and ‘‘Old To¨o¨lo¨.’’ We will soon see that the first grammar is a rough approximation of the grammar of most speaker groups, whereas the second is the grammar of old female upper-middle-class To¨o¨lo¨ residents. (33) and (34) display these two grammars in two equivalent ways: as ordered pairs of constraints and as totally ordered tableaux. (33) Most speakers a. Ordered pairs {*EA *IA} b. Tableaux Faith ! [ee], [ii]
*EA
*IA
*EA
Faith
*IA
! [ee], [ia]
Faith
*EA
*IA
! [ea], [ia]
(34) Old To¨o¨lo¨ a. Ordered pairs {*EA *IA, Faith *IA}
Derived Environment Effects in Colloquial Helsinki Finnish
447
b. Tableaux *EA
Faith
*IA ! [ee], [ia]
Faith
*EA
*IA ! [ea], [ia]
In terms of totally ordered tableaux, Most speakers properly includes Old To¨o¨lo¨. In terms of ordered pairs of constraints, Old To¨o¨lo¨ properly includes Most speakers. Speaking in the tableau idiom, I will say that the grammar Old To¨o¨lo¨ is a subgrammar of the grammar Most speakers. The fact that grammars may literally include other grammars will be exploited in this chapter in various ways. Now, consider the outputs of these two grammars. Old To¨o¨lo¨ consists of two tableaux: the first predicts coalescence in EA-stems, the second predicts no coalescence. I interpret this empirically as follows: Old To¨o¨lo¨ allows coalescence optionally in EA-stems, but prohibits it in IA-stems. Next, consider Most speakers. This time, we have three tableaux: one predicts coalescence in IA-stems, two predict coalescence in EA-stems, and one predicts no coalescence. The empirical intepretation I will adopt is straightforward: both vowel sequences may optionally coalesce, but EA-stems coalesce more than IA-stems. More specifically, I will assume the following quantitative interpretation of partial ordering (Anttila 1997): (35) Quantitative interpretation a. A candidate is predicted by the grammar i¤ it wins in some tableau. b. If a candidate wins in n tableaux and t is the total number of tableaux, then the candidate’s probability of occurrence is n=t. Given this interpretation, the grammar Most speakers thus predicts a pattern where EA coalesces in two-thirds and IA in one-third of the cases. This is roughly the quantitative profile of Paunonen’s total corpus (Paunonen 1995, 110): EA ! EE 63.1%; IA ! II 29.8%. 18.4.2
The NDEB Effect
I have now introduced the general assumptions of my analysis and generated a simple phonological typology of six dialects, but have not yet addressed the morphological facts, the most important of which is the NDEB e¤ect. At present, the constraints apply indiscriminately to both derived and nonderived forms, predicting many incorrect outputs—for example, optional coalescence in the nonderived rasia@*rasii ‘box’ by Most speakers, something never found in any dialect. A novel Optimality-Theoretic solution to the derived environment problem has been proposed by Łubowicz (2002), who suggests that NDEB e¤ects can be derived from the local conjunction of markedness and faithfulness constraints (Smolensky 1995). The intuitive content of the morphological part of Łubowicz’s proposal can
448
Arto Anttila
be informally summarized as follows: if a stem-final consonant gets grouped into the same syllable with a su‰x-initial vowel (.C-V.), this results in a stem/syllable misalignment that triggers the alternation; within roots no stem/syllable misalignment occurs, hence the alternation is blocked. Thus, according to Łubowicz, the common denominator of morphological derived environment e¤ects is stem/syllable misalignment. However, this does not hold true of Finnish Vowel Coalescence. In cases like o.ve.-a ‘door-par’ the stem and syllable edges are perfectly aligned. We would thus expect coalescence not to take place, but it does.11 Another possibility is to suggest that, just as di¤erent dialects are obviously a‰liated with di¤erent grammars, in the same way di¤erent morphological categories may subscribe to di¤erent cophonologies within an individual’s system.12 (For various uses of cophonologies, see e.g. McCarthy and Prince 1993b; Itoˆ and Mester 1995a, 1995b, 1998; Orgun 1996; Inkelas et al. 1997; Inkelas 1998; Anttila 2002.) Thus, for example, we could say that most speakers use the grammar {*EA *IA} for general purposes, but its subgrammar {*EA *IA, Faith *IA} for nonderived roots, with the notable exception of old female upper-middle-class To¨o¨lo¨ residents who have generalized this subgrammar to all stems, derived and nonderived. The question now arises: Why would nonderived roots be a‰liated with a particular ranking instead of some other ranking? For example, why would nonderived roots not choose the grammar labeled ‘‘Advanced dialect’’, which predicts coalescence everywhere? Indeed, why do we not have derived environment blocking? No principled answer is forthcoming. As an explanation of NDEB, cophonologies seem descriptive and arbitrary.13 A third alternative is to assume that NDEB arises from a specialized faithfulness constraint Faithroot that requires faithfulness to roots. (For various uses of this constraint, see e.g. McCarthy and Prince 1995; Beckman 1998; Alderete 1999.) Under this scenario, coalescence in nonderived forms would result in a violation of Faithroot (rasia/*rasii ‘box’), whereas coalescence in morphologically derived forms would not violate Faithroot (lasia@lasii ‘glass-par’). For a di¤erent faithfulness-driven approach to NDEB, see Burzio 2000. While perhaps not the general solution, Faithroot fares better than its competitors in explaining the Finnish blocking facts. In particular, Faithroot yields the observed asymmetry between derived and nonderived forms, both in its categorical and quantitative manifestations. To derive the categorical NDEB e¤ect in IA-stems, it su‰ces to rank Faithroot above the markedness constraint *IA. This yields the subtler dialect typology in (36). Coalesced forms appear in boldface. (36) Rankings: *EA *IA, Faithroot *IA /ove-a/ 1. Faith *EA Faithroot *IA ove-a 2. Faith Faithroot *EA *IA ove-a
/hopea/ hopea hopea
/lasi-a/ lasi-a lasi-a
/rasia/ rasia rasia
Derived Environment Effects in Colloquial Helsinki Finnish
3. 4. 5. 6. 7. 8.
Faithroot Faith *EA *IA Faithroot *EA Faith *IA Faithroot *EA *IA Faith *EA Faith Faithroot *IA *EA Faithroot Faith *IA *EA Faithroot *IA Faith
ove-a ove-e ove-e ove-e ove-e ove-e
449
hopea hopea hopea hopee hopee hopee
lasi-a lasi-a lasi-i lasi-a lasi-a lasi-i
rasia rasia rasia rasia rasia rasia
Instead of three, five distinct invariant systems are now predicted. The variable systems will be discussed shortly. (37) a. b. c. d. e.
No coalescence (1–3) Coalescence in derived EA-environments (4) Coalescence in all derived environments (5) Coalescence in all EA-environments (6–7) Coalescence everywhere but nonderived IA-environments (8)
Yet another possibility is to derive NDEB from representational assumptions, in particular prespecification. I will put this alternative on hold and take it up after discussing the part-of-speech e¤ect. At this point, I will tentatively accept the root faithfulness hypothesis as the basis for NDEB. 18.4.3
The Part-of-Speech Effect
I now turn to the part-of-speech e¤ect. The basic observation is that nouns coalesce less than adjectives, just as nonderived forms coalesce less than derived forms. This seems to suggest that the two should receive parallel synchronic analyses. One way to do this would be to introduce another specialized faithfulness constraint Faithnoun , parallel to Faithroot , following Smith 1997. Such an analysis could be implemented as follows. We first add the constraint Faithnoun into the existing partial ordering. This results in a system with five constraints (*EA, *IA, Faithroot , Faithnoun , Faith) and two binary rankings (*EA *IA, Faithroot *IA). Within this space, we must now locate the grammar that best matches the Finnish data. Here we will try to model the average frequencies in the entire corpus; modeling the distinct dialects within the corpus is a challenge that will be left for future work. The best match is achieved by the grammar in (38) using mean absolute deviation as a measure of fit.14 The percentages are very good in categories with a large number of tokens, but worse in categories with fewer tokens. (38) The noun-faithfulness analysis Constraints: *EA, *IA, Faithroot , Faithnoun , Faith Rankings: *EA *IA, Faithroot *IA, *EA Faithroot , *EA Faithnoun , Faithroot Faithnoun
450
Arto Anttila
(39) Predictions and observations Pred N
Obs N Pred%
Obs1 % Obs2 %
Number of tokens
ove-anoun hopeanoun ideanoun lasi-anoun rasianoun
571 38 10 1,012 0
293 9 0 1,014 0
80 80 80 20 0
41.0 18.8 0 20.0 0
24.3–67.2% 3.4–66.7% 0 7.8–43.5% 0%
714 48 (native) 12 (borrowing) 5,059 847
toope-aadj makeaadj uusi-aadj kauhiaadj
— 1,396 1,279 0
— 1,263 1,289 0
80 80 30 0
— 72.4 30.2 0
— 49.5–93.0% 9.6–64.2% 0%
0 1,745 4,264 261
However, this is just the first step; the interaction of phonology and morphology gets subtler. Recall that not all nouns are alike: native nouns like hopea ‘silver’ coalesce variably (hopea@hopee), whereas otherwise identical recent borrowings like idea ‘idea’ never coalesce (idea/*idee). Since the grammar in (38) does not distinguish between the two noun classes, it predicts the same coalescence rate (80 percent) in both. Because /ea/-final nouns are not very common, this oversight is negligible from the quantitative point of view, but numbers aside, it is obvious that a generalization is being missed. The question is how to capture increasingly smaller subregularities and ultimately lexical idiosyncrasies in this framework. The most straightforward move would be to introduce a high-ranking constraint Faithrecentlyborrowednoun that would block coalescence in idea ‘idea’, but not in hopea ‘silver’. Two problems arise. First, if one wants to maintain the hypothesis that constraints are universal, this one will probably not qualify. Second, as discussed in section 18.2, the blocking of coalescence in idea and other recently borrowed nouns can be plausibly attributed to analogy with the statistical majority of nouns that typically end in /ia/. But if this is so, then the explanation has nothing to do with universal constraints, but rather the brute fact that Finnish happens to have many more nouns ending in /ia/ than in /ea/ plus the independent principle ‘‘Imitate the majority.’’ An alternative is to postulate cophonologies. Under this approach, di¤erent lexical items, or groups of lexical items, subscribe to di¤erent rankings, hence their phonological di¤erences. Note that this is not equivalent to choosing a ranking randomly if we make the reasonable assumption that lexeme-specific phonologies are subgrammars of the general phonology of the language. Under this scenario, the lexical items idea ‘idea’ and hopea ‘silver’ are similar in that both share the general rankings of
Derived Environment Effects in Colloquial Helsinki Finnish
451
Finnish; they are di¤erent in that idea ‘idea’ imposes further rankings that categorically block coalescence. However, if we posit distinct cophonologies for native and recently borrowed nouns, the obvious question is why not do the same for nouns and adjectives. What makes a cophonology analysis particularly attractive is that it allows us to dispense with the constraint Faithnoun . This simplifies our constraint inventory, which in turn reduces the number of possible grammars. The question is whether the simpler fourconstraint system is still empirically su‰cient—that is, whether we are still able to find a grammar that satisfactorily matches the Finnish data. The grammar shown in (40) seems a reasonably good fit. (40) The cophonology analysis Constraints: *EA, *IA, Faithroot , Faith General rankings: *EA *IA, Faithroot *IA Noun rankings: General rankings W Faithroot *EA Adjective rankings: General rankings W *EA Faithroot Minor regularity: Four native /ea/-final nouns (hopea ‘silver’, ha¨pea¨ ‘shame’, lipea¨ ‘lye’, aukea ‘clearing’) do not subscribe to the noun rankings, but only to the general rankings of Finnish. (41) Predictions and observations Pred N Obs N
Pred%
Obs1 %
Obs2 %
Number of tokens
ove-anoun hopeanoun ideanoun lasi-anoun rasianoun
357 18 0 1,265 0
293 9 0 1,014 0
50.0 37.5 0.0 25.0 0.0
41.0 18.8 0.0 20.0 0.0
24.3–67.2% 3.4–66.7% 0% 7.8–43.5% 0%
714 48 (native) 12 (borrowing) 5,059 847
toope-aadj makeaadj uusi-aadj kauhiaadj
— 1,309 1,066 0
— 1,263 1,289 0
75.0 75.0 25.0 0.0
— 72.4 30.2 0.0
— 49.5–93.0% 9.6–64.2% 0%
0 1,745 4,264 261
In this analysis, the noun and adjective cophonologies are literally embedded within the general phonology of Finnish. This point is made more visually by diagram (42), which displays all the subregularities (partial orderings) hidden within the general Finnish rankings. The eight tableaux are located on the right, numbered from 1 to 8 and annotated with the relevant outputs. All in all, we find thirty-two distinct subregularities of which twelve are invariant, twenty variable.
452
Arto Anttila
(42)
The noun grammar correctly blocks coalescence in recently borrowed nouns like idea ‘idea’ (idea/*idee). The marginal variability in native nouns like hopea ‘silver’
Derived Environment Effects in Colloquial Helsinki Finnish
453
(hopea@hopee) follows if we place them directly under the general phonology of Finnish. The pattern is typical of analogical change: a handful of native roots have not yet caught up with the recent analogical trend, and perhaps never will, whereas incoming roots unhesitatingly subscribe to the noun grammar. Note that the exceptional variability is predicted to be phonologically possible only in EA-roots. In IAroots, the general phonology of Finnish dictates blocking across the board. In general, cophonologies seem the right way of handling the synchronic complications due to lexical di¤usion (Wang 1969, 1977; Labov 1994; Kiparsky 1988a, 1995), a species of analogical change that proceeds lexical item by lexical item, yet crucially in a phonologically constrained manner. In his otherwise insightful discussion, Paunonen did not notice the crucial role that the noun/adjective distinction plays in Vowel Coalescence. This led him to the puzzling observation that coalescence is systematically more common in nonderived EA-nominals (e.g., makea ‘sweet’) than in derived EA-nominals (partitives) (e.g., ove-a ‘door-par’). To make sense of this observation, he proposed the following explanation: (43) The alternation korkea@korkee ‘high’, ainoa@ainoo ‘only’ [nonderived words, A.A.] thus only concerns the shape of a single lexeme, whereas in partitives [derived words, A.A.] the alternation between coalesced and hiatus forms would entail increasing complexity in the morphological rule system. From this point of view, the sound development ea, ea¨ > ee and oa > oo is freer to proceed in ea, oa nominals than in partitives. For this reason, one would expect that in ea, ea¨ nominals the coalescence forms would be more common than in ea, ea¨-final partitives. Indeed, this is the case. (Paunonen 1995, 141) [Translation mine, A.A.] In other words, Paunonen is forced to conclude that sound change somehow avoids creating complexity in morphological rule systems. However, there is an alternative explanation available. While perfectly accurate, Paunonen’s word counts do not distinguish between nouns and adjectives. Recall that virtually all EA-final nonderived forms (e.g., makea ‘sweet’) are adjectives, and virtually all EA-final derived forms (e.g., ove-a ‘door-par’) are nouns. Thus, the tendency of makea ‘sweet’ to coalesce more than ove-a ‘door’ is not because the first is nonderived and the second derived; it is because the first is an adjective and the second a noun. This strong part-of-speech e¤ect is hidden in Paunonen’s word counts, and once teased apart, the puzzle disappears. Indeed, nonderived words are less likely to coalesce than derived words, not the other way around. The general predictions of my analysis are summarized in (44). The following two generalizations hold of all the thirty-two subregularities:
454
Arto Anttila
(44) All else being equal, Vowel Coalescence is a. At least as likely in EA as in IA sequences. b. At least as likely in derived as in nonderived environments. In other words, the grammar correctly ties the likelihood of coalescence to segmental phonology and derivedness. In contrast, no general predictions are made concerning the relative coalescence rates of nouns and adjectives. This is because the noun/adjective distinction is not hardwired in the constraints, but follows from the language-specific assignment of lexical items to cophonologies: the fact that nouns coalesce less than adjectives is contingent on the particular assignment chosen. It would thus not be surprising to find a language where the behavior of nouns and adjectives were the reverse of Finnish—that is, where adjectives are more faithful to their inputs than nouns. It is here that the noun-faithfulness analysis (Smith 1997) and the cophonology analysis make crucially di¤erent predictions. Let us now summarize the main results. We have found that the blocking of Vowel Coalescence can be traced back to three independent factors: vowel height, (non)derivedness, and part of speech. These e¤ects add up and in the extreme case result in categorical blocking. I have proposed that the derived environment e¤ect and the part-of-speech e¤ect are fundamentally di¤erent: the derived environment e¤ect follows from the (presumably universal) phonological constraint Faithroot , whereas the part-of-speech e¤ect is language-specific and synchronically implemented by means of the very di¤erent device of cophonologies. 18.5
A Note on Prespecification
We are now ready to consider yet another approach to blocking e¤ects: prespecification. Various versions of this approach have been proposed by, for example, Ringen (1975), Borowsky (1986), Kiparsky (1993a), Inkelas and Cho (1993), and Inkelas (2000). Kiparsky’s formulation is stated in (45): (45) NDEB is a property of structure-building rules operating on underspecified representations (Kiparsky 1993a). The intuition is simple: if the target of a structure-building rule is underspecified, the rule can apply and the result is alternation. If the target of a structure-building rule is prespecified, the rule cannot apply because it is defeated by the preexisting representation and the result is blocking. In the Finnish case, we could prespecify the nonalternating vowels as [þlow] and leave the alternating vowels underspecified for height as shown in (46): (46) A prespecification analysis a. /hopeA/ hopea@hopee b. /idea/ idea/*idee
/A/ underspecified for height /a/ prespecified [þlow]
Derived Environment Effects in Colloquial Helsinki Finnish
455
As Sharon Inkelas (personal communication) has pointed out to me, this approach is perfectly compatible with Optimality Theory. For a demonstration that the approach extends to blocking e¤ects of various kinds, see Inkelas (2000). Following Inkelas, we could assume that the grammar permits underspecification only if an alternation has been observed and that the speaker assumes that an observed sequence is nonalternating until proven otherwise. It would be straightforward to reformulate the present analysis along these lines. All we need to do is change our assumptions about underlying forms on a lexeme-by-lexeme basis and substitute NoLongVowel (Rosenthall 1994) for Faith; both constraints essentially say ‘‘don’t coalesce.’’ However, a prespecification analysis does not seem warranted, for reasons to which I now turn. First, prespecification does not extend to quantitative regularities in any obvious way. This means that partial ordering would remain a necessary part of the analysis. Since partial ordering by itself captures the relevant di¤erences, such as the di¤erence between hopea@hopee ‘silver’ (variable coalescence) and idea/*idee ‘idea’ (blocking), it is not clear what prespecification would add to the analysis. Second, as pointed out by Inkelas, prespecification provides a uniform approach to all blocking e¤ects. However, as we have seen, not all blocking e¤ects are alike. Blocking of the nonderived environment type is a well-known crosslinguistic fact; to capture it I posited the (potentially universal) constraint Faithroot . Blocking of the noun type is arguably a Finnish-specific fact that results from the analogical influence of the majority of nouns. Furthermore, blocking of the vowel-height type may well have a phonetic explanation. All three blocking phenomena can undoubtedly by described as prespecification, but in reality they have quite di¤erent sources. It seems that by adopting prespecification we would simply end up stipulating that most roots, nouns, and ia-sequences tend to have a prespecified /a/, whereas most su‰xes, adjectives, and ea-sequences tend to have an underspecified /A/. This would describe the facts, but we could just as well have stipulated the opposite. The point is that different types of blocking originate in quite di¤erent ways and at least some types of blocking can be derived from general principles, grammatical or otherwise. Third, following Inkelas, we might propose that underspecification is possible only if an alternation has been observed, otherwise full specification is assumed. However, as we have been observing throughout this chapter, the fact that an alternation is observed in context A, but not in context B, is not a historical accident, but something that follows from the grammar itself, at least in part. Again, while underspecification can describe the facts, it adds little to our understanding of why alternations should occur in certain environments, but be blocked in others. In sum, while prespecification is compatible with Optimality Theory, Vowel Coalescence does not provide evidence for a prespecification analysis of blocking.
456
18.6
Arto Anttila
On the Limitations of This Study
Finally, a comment on the limitations of the present study is due. My quantitative results are preliminary in two respects. First, only -ea and -ia stems have been considered. The reason for choosing these two was that they are the most frequent of the vowel sequences subject to coalescence. However, Paunonen’s study suggests that including -oa, -o¨a¨, -ua, and -ya¨ will not drastically change the general picture. Second, each category contains words of varying lengths. For example, among ia-final derived adjectives we find both suur-i-a ‘big-pl-par’ (three syllables) and semmos-i-a ‘such-pl-par’ (four syllables). The di¤erence is potentially relevant because word length a¤ects secondary stress placement, which in turn a¤ects coalescence. The prediction is that coalescence should be blocked if secondary stress falls on either vowel. In su´u.r-i.-a there is no secondary stress; in se´m.mo.s-ı`-a secondary stress falls on the third syllable, although perhaps only at the word level or postlexically (Kiparsky 2003a), and the matter is further complicated by the fact that secondary stress is itself optional in certain circumstances (Keyser and Kiparsky 1984). It will be interesting to see whether including the secondary stress factor in the analysis results in more accurate quantitative predictions. 18.7
Conclusion
In this chapter, I have examined a peculiar case of nonderived environment blocking: optional Vowel Coalescence (VC) in Colloquial Helsinki Finnish. Its theoretical significance lies in the fact that it runs counter to all the best-known explanations of derived environment behavior. As Kiparsky (1993a) already showed, VC is incompatible with the Strict Cycle Condition (Mascaro´ 1976) and the Elsewhere Condition (Kiparsky 1982d) because it is both cyclic and lexical, yet applies in nonderived environments. We may now further add that VC is also incompatible with the Revised Alternation Condition (Kiparsky 1973e): VC is a nonautomatic optional neutralization rule that is blocked in nonderived environments in certain phonological contexts, but not in others. Finally, although it would be possible to tie derived environment behavior to structure-building rules by using stipulative prespecification to block VC in the desired environments, this would miss the fundamental phonological and morphological generalizations, in particular the quantitative ones. What the evidence ultimately shows is that derived environment behavior cannot be tied to phonological rules of any kind. The reason is that the very same rule may show NDEB in certain environments, but not in others, depending on global grammatical considerations. The old explanations were clearly in the right direction, but time and again it turned out that NDEB was tied to a particular rule type only approximately, if at all. It now seems that the failure to unify NDEB e¤ects was mainly
Derived Environment Effects in Colloquial Helsinki Finnish
457
due to the decision to take rules—that is, specific phonological processes—as the locus of explanation. In this case at least, Optimality Theory leads one to look for explanations in a still better direction: since Optimality Theory does not even recognize rules as theoretical entities, it suggests that the explanation for NDEB must be found from global interactions of markedness and faithfulness. I have argued that the blocking of Vowel Coalescence in Colloquial Helsinki Finnish is a fairly heterogeneous phenomenon that can be traced back to at least three distinct sources (M ¼ some markedness constraint, F ¼ some faithfulness constraint): (47) a. Markedness di¤erences: *ea F *ia (Blocking in IA-sequences, alternation elsewhere) b. Special vs. general faithfulness: Faithroot M Faith (Blocking in roots, alternation elsewhere) c. Cophonologies: Faithroot *ea vs. *ea Faithroot (Blocking in nouns, alternation in adjectives) I have also shown that all three types of blocking may surface both categorically and quantitatively. Both types of e¤ects follow from an Optimality-Theoretic analysis that assumes quantitatively interpreted partially ordered grammars. Notes * Earlier versions of this chapter were presented at a sta¤ seminar at the Centre for Advanced Studies/Department of English Language and Literature, National University of Singapore (August 1999), a Symposium on Variation Theory and Formal Theory at the 28th Conference on New Ways of Analyzing Variation, University of Toronto (October 1999), MIT Phonology Workshop (March 2001), and New York University Linguistics Colloquium (March 2001). I thank Bao Zhiming, Young-mee Yu Cho, Vivienne Fong, Gregory Garretson, Kristin Hanson, Tarja Heinonen, Sharon Inkelas, Brett Kessler, Anna Łubowicz, K. P. Mohanan, Tara Mohanan, Carol Neidle, Anthi Revithiadou, Cathie Ringen, Caro Struijke, Cheryl Zoll, and an anonymous reviewer for comments and/or discussion. I thank Kotimaisten kielten tutkimuskeskus (Research Institute for the Languages of Finland) for the permission to use the electronic version of Nykysuomen sanakirja (Dictionary of Modern Finnish) (Sadeniemi 1973) and the Department of General Linguistics, University of Helsinki, for giving me access to Heikki Paunonen’s Spoken Helsinki Finnish Corpus. The research for this chapter was in part supported by a National University of Singapore Faculty of Arts and Social Sciences Postdoctoral Fellowship. All errors are mine. 1. The tagging had to be done manually because of morphological ambiguities that are common in speech due to reduction phenomena. The following decisions are worth mentioning: (i) Everything was tagged, including the interviewer. (ii) Unclear cases, including irresolvable ambiguities, were marked by a special tag and excluded. (iii) Ambiguities that were frequently irresolvable include oikee, either from oikea ‘right’ or oikein ‘really’, where only the first is an instance of Vowel Coalescence, and part-of-speech ambiguities like suomalaisia, either
458
Arto Anttila
‘Finn-pl-par, n’ or ‘Finnish-pl-par, adj’. Such cases are often, but not always, disambiguated by the context. (iv) Sequences of two vowels where either vowel bears primary stress were not tagged (e.g., te´.at.te.ri ‘theater’, ra´u.ta.tie¼a`.se.ma ‘railway station’) because they do not meet the structural description of the rule (see below). 2. The following abbreviations are used in the glosses: adj ¼ adjective, comp ¼ comparative, ine ¼ inessive, n ¼ noun, par ¼ partitive, pl ¼ plural. 3. Of the following list, only idea and komitea come from Paunonen’s corpus. The rest are based on my own judgments. 4. The reviewer asks whether EA-coalescence is also disfavored in nouns that contain EAsequences but end in IA-sequences. There are no such examples in Paunonen’s corpus. Nykysuomen sanakirja (Dictionary of Modern Finnish) contains two relevant examples: genealogia ‘genealogy’ and oseanografia ‘oceanography’. According to my intuitions, coalescence is out of the question in both cases (*geneelogia, *oseenografia), confirming the generalization that coalescence is blocked across the board in nonderived recently borrowed nouns in just the way to be proposed below. 5. For a more detailed account of the phonology/morphology interactions involved, see Anttila 2002. 6. A parallel case is the Polish rule of First Velar Palatalization (Rubach 1984), which palatalizes velars before front vocoids and is blocked in nonderived environments. An interesting observation, attributed to Christina Bethin by Łubowicz (2002, 256) is that all these NDEB cases are borrowings: [ke]f ’ir ‘kefir’, [ke]lner ‘waiter’, [k’i]s´el ‘jelly’, [ge]nc’jana ‘gentian’, a[ge]nt ‘agent’, [g’i]ps ‘plaster’, [x’i]g’jen´istka ‘hygienist’, [x’i]stor’ja ‘history’, [xe]m’ik ‘chemist’. In addition, all are nouns. 7. At this point, I have quantitative data only on the vowel sequences /ea/ and /ia/, which are the two most common mid-low and high-low combinations. 8. One may wonder how (20d) is possible at all. The explanation hinges on secondary stress. Recall that coalescence requires both vowels to be unstressed. In this case, u´.se.a`m.pii is a possible stress pattern and coalescence is blocked because of secondary stress on the third syllable. The role of secondary stress in Vowel Coalescence has not escaped the attention of Finnish historical phonologists; see for example Rapola 1966, 419. Unfortunately, at this point I do not yet have serious quantitative data on the secondary stress factor. I hope to return to this topic on another occasion. 9. The absence of derived -ea-final adjectives can be considered an accidental lexical gap. A possible example of an /e/-final adjective is /toope/ ‘stupid’, in the partitive singular /toope-a/. While Paunonen’s corpus contains no examples of this word, according to my intuitions this word clearly coalesces: toope-e ‘stupid-par’. 10. The constraints *EA and *IA are deliberately noncommittal as to the exact nature of this asymmetry, which is found in other languages as well (see, e.g., Haas 1988). Casali (1996, 48– 50) suggests that vowel sequences where the two vowels are not su‰ciently distinct from each other are disfavored. 11. An analogous example is Basque final-vowel raising (Hualde 1989), discussed by Łubowicz 2002 and Inkelas 2000. 12. I use the term cophonology to refer to any subgrammar of a given language. Of two cophonologies, one may or may not be a subgrammar of the other. There are obvious similarities
Derived Environment Effects in Colloquial Helsinki Finnish
459
between cophonologies, the levels of Lexical Morphology and Phonology (Kiparsky 1982d; Mohanan 1986; Kiparsky forthcoming), and lexical strata in languages like Japanese, as analyzed in Itoˆ and Mester 1995a, 1995b, 1998. 13. As the reviewer points out, no advocate of cophonologies has ever used them to explain nonderived environment blocking; however, see Yu 2000. The reviewer also asks what it would look like if a rule or constraint applied only in nonderived environments, and suggests it would look like a morpheme structure condition, and those abound. However, it is easy to imagine a dialect of Finnish where coalescence would apply optionally in roots (rasia@rasii ‘box’), but not in derived forms (lasi-a/*lasii ‘glass-par’). No such dialects exist and we must explain why. 14. This grammar was found by exhaustive search using a simple recursive algorithm. Within the space defined by the five constraints and the two binary rankings, the algorithm found 579 nonidentical partial orderings of which 40 were total orderings.
19
A Historical Perspective on Nonderived Environment Blocking: The Case of Korean Palatalization
Young-mee Yu Cho
19.1
Introduction
This chapter addresses the well-known problem of Nonderived Environment Blocking (NDEB) e¤ects from the perspective of Korean Palatalization, a phonological rule showing NDEB e¤ects that has been cited at every stage of the long evolution of theoretical perspectives on nonderived environment blocking. Korean Palatalization is an obligatory neutralization rule that applies only in environments derived by virtue of a combination of morphemes.1 This chapter examines Korean Palatalization both diachronically and synchronically, ultimately arguing that the development of the derived environment e¤ect results from a change in the ranking of lexical faithfulness constraints. The chapter thus confirms recent e¤orts of Anttila, building on earlier work by Kiparsky, to show that Nonderived Environment Blocking cannot be a rigid principle of grammar but instead is an language-specific result of the way individual rules are stated (Kiparsky 1993a) or constraints are ranked (Anttila, chap. 18, this volume) in a particular grammar. 19.2
Nonderived Environment Blocking
From the very beginning of Generative Phonology, a number of rules have been observed to apply only across morpheme boundaries, to skip morpheme-internal targets, or to apply only when another phonological rule has created a new environment that meets the structural description of the rule in question. This syndrome is widely known as Nonderived Environment Blocking (NDEB) (Kiparsky 1993a). NDEB has been a source of many theoretical discussions over the past thirty years, as attempts have been made, in various frameworks, to define the nature and the scope of rules subject to NDEB. The earliest proposal is Kiparsky’s (1968b) Alternation Condition, which stipulates that ‘‘neutralization processes cannot apply to all occurrences of a morpheme’’ (Kiparsky 1982b, 148). The Alternation Condition has the e¤ect of limiting the abstractness of underlying phonological representations by
462
Young-mee Yu Cho
prohibiting morphophonemically invariant morphemes from undergoing any rules of phonological neutralization. As a consequence, rules of absolute neutralization are indirectly prohibited in the grammar. Kiparsky (1973b, 1973e) subsequently reconceptualized the Alternation Condition from a passive constraint on underlying representations to an active condition on rule application, formalized as the Revised Alternation Condition (RAC). A statement of the RAC, taken from Iverson and Wheeler 1988, 326, is given below: (1) Revised Alternation Condition Obligatory neutralization rules apply only in derived environments. An environment X is derived with respect to a rule Y if X satisfies the structural description of Y crucially by virtue of a combination of morphemes or the application of a structure-changing rule. The RAC underwent a further necessary transformation in the context of Lexical Morphology and Phonology (LMP; Kiparsky 1982d), a theory in which phonological and morphological operations are interleaved in lexical strata. The RAC surfaced in that theory as a global condition on cyclic rule application called the Strict Cycle Condition (SCC), reflecting its parallels with strict cycle proposals by Chomsky (1965) and Kean (1974). A formulation of the SCC (Kiparsky 1982d, 41) is given below: (2) Strict Cycle Condition (SCC) a. Cylic rules apply only to derived representations. b. Def: A representation f is derived w.r.t. rule R in cycle j i¤ f meets the structural analysis of R by virtue of a combination of morphemes introduced in cycle j or the application of a phonological rule in cycle j. The SCC di¤ers from the earlier RAC in targeting only cyclic rules in derived representations (Mascaro´ 1976; Kiparsky 1982d). As a result, nonderived sequences escape application of cyclic rules, but do undergo noncyclic rules. Kiparsky (1982d) showed that the SCC can be derived from the Elsewhere Condition; on the assumption that nonderived forms are associated lexically with identity rules, those identity rules are more specific than, and therefore block by Elsewhere disjunctive ordering, structure-changing phonological rules of the grammar. Ten years later, however, it had become clear to Kiparsky that NDEB e¤ects and cyclicity are independent. Kiparsky (1993a) cites examples of cyclic rules that apply even in nonderived environments and noncyclic rules that exhibit NDEB. He even discusses one case in which the same rule, Finnish Vowel Coalescence, sometimes shows NDEB e¤ects and sometimes does not (see Anttila and Cho 1999 and Anttila chap. 18, this volume for the demonstration that phonological markedness is responsible for the variable NDEB e¤ects in this case). Given this lack of correlation between NDEB e¤ects and rule type, Kiparsky (1993a) abandoned the attempt to
A Historical Perspective on Nonderived Environment Blocking
463
trace NDEB e¤ects to a general constraint on phonological rules themselves, such as the SCC, the RAC, or the Alternation Condition. Instead, Kiparsky proposed to derived NDEB e¤ects from the postulation of a class of purely structure-filling rules that can apply only to underspecified representations (see Poser 1982 for the roots of this idea). Feature-filling rules will fail to apply to segments that are prespecified for the feature in question. Insofar as morpheme-internal segments always appear in the same phonological context, they also will fail to alternate; even if lexically underspecified, they undergo the same cyclic structure-filling rules in every context. Only those lexically underspecified segments that surface in a variety of morphological and phonological contexts can alternate. This proposal, which eliminates the need for disjunctive definitions of ‘‘derived environment,’’ embodies the view that the explanation for NDEB e¤ects should be sought not in a general condition on rules but elsewhere in the grammar. Several proposals along similar lines have been formulated within Optimality Theory (OT; Prince and Smolensky 1993; McCarthy and Prince 1993b). In particular, both Burzio (2000) and Anttila (chap. 18, this volume) propose accounts of particular cases of NDEB that rely on an OT formulation of the original idea behind Kiparsky’s (1982d) Elsewhere Condition, namely the notion of ‘‘distinct from a lexical item.’’ In this chapter, I follow the lead of Burzio and Anttila, proposing that NDEB should be understood as a lexeme faithfulness e¤ect. The proposal is couched within Stratal Optimality Theory, Kiparsky’s (1997a; 2002b) implementation of LMP in the general framework of OT. Stratal OT assumes three universal phonological strata Cyclic (¼ word-internal), Word, and Postlexical. The phonology of the three strata varies according to the ranking of faithfulness with respect to the other constraints in the constraint hierarchy. This conception of the lexicon, shared by the pre-OT version of LMP (Kiparsky 1982d) as well as the OT-based Sign-Based Morphology theory of the phonology-morphology interface developed in Orgun 1996, captures the crosslinguistic generalizations that cyclic and word-stratum a‰xations are distinct and that morphologically bound stems can be cyclic domains while prosodically incomplete elements (a‰xes and roots) are not cyclic domains. 19.3
FAITH-LEX and NDEB
Working within the general framework of Stratal OT, I propose that NDEB e¤ects follow from the specialized faithfulness constraint family, Faith-LEX, which requires faithfulness to an input lexeme. To illustrate briefly how the new proposal works, let us analyze one of the textbook examples of NDEB, namely English Trisyllabic Shortening (TSS), discussed extensively in Chomsky and Halle 1968, Siegel 1974, Allen 1978, Kiparsky 1982d, and many other works. As analyzed in Kiparsky 1982d, English Stratum 1 a‰xation includes a nominalizing su‰x -ity, which triggers TSS in
464
Young-mee Yu Cho
morphological derivatives like divinity (divineþity). TSS fails to apply to forms derived at Stratum 2 (blindingly) or to nonderived forms (nightingale). In Stratal OT, Stratum 1 corresponds to the Cyclic stratum; the generalization that needs capturing is that Cyclic stratum a‰xes, but not Word stratum a‰xes, trigger TSS. Faith-LEX is the constraint family responsible for capturing this generalization. Faith-LEX requires input-output identity for inputs that instantiate a listed lexical item. In Stratal OT, inputs come in three types—Root, at the Cyclic Stratum; Stem, in the Word stratum; and the Prosodic Word, at the Phrasal Stratum. Recall that faithfulness constraints, including Faith-LEX, are ranked independently at each stratum; for terminological clarity, I will refer to the instantiation of Faith-LEX at the Cyclic stratum as Faith-ROOT; at the Word stratum, as Faith-STEM; and at the Postlexical stratum, as Faith-PWD. Faith-LEX constraints coexist with the fully general input-output faithfulness constraint family (Faith), which covers all inputs, not just those that happen to correspond to lexically listed lexemes. NDEB e¤ects result from the ranking Faith-LEX Phono-constraint Faith, where ‘‘Phono-constraint’’ represents a markedness constraint. Ranking Phonoconstraint above Faith means that violations of Phono-constraint are repaired. Ranking Faith-LEX above Phono-constraint, however, protects those inputs to the stratum in question that happen to correspond to a listed lexical item. This analysis is conceptually parallel to that in Kiparsky 1982d, in which identity rules for listed lexical items apply disjunctively with, and block, general structure-changing phonological rules. To illustrate, let us return to the case of TSS in English. As shown in (3), the ranking Faith-ROOT TSS Faith protects una‰xed roots such as nightingale from undergoing TSS at the Cyclic stratum.2 By contrast, a form derived at the Cyclic stratum, like divin-ity, is not protected. Faith-ROOT has no jurisdiction over /divineþity/ since there is no lexically listed item of the form [divine-ity]. The optimal output for the input /divineþity/ is thus the one that satisfies TSS, the markedness constraint underlying Trisyllabic Shortening, at the expense of low-ranked Faith:3 (3) English Trisyllabic Shortening (TSS): Cyclic stratum /nightingale/
Faith-ROOT
+ a. n[ai]ghtingale b. n[i]ghtingale
TSS
Faith
* *!
*
/divineþity/ a. div[ai]nity + b. div[i]nity
*! *
A Historical Perspective on Nonderived Environment Blocking
465
In the Word and Postlexical strata, in contrast to the Cyclic stratum, TSS is ranked lower than both kinds of faithfulness. As a result, words derived at the Word stratum and phrasal combinations are immune from undergoing TSS. In (4), blindingþly is a new morphological concatenation at the Word stratum. Because blindingly is not lexically listed, the outputs are not subject to Faith-STEM. However, because TSS is ranked below general Faith, TSS has not e¤ect; the [ai] of blindingly is preserved. (4) English Trisyllabic Shortening (TSS): Word stratum /blindingþly/
Faith-STEM
Faith
+ a. bl[ai]ndingly b. bl[i]ndingly
TSS *
*!
Within this framework, NDEB is not restricted to cyclic rules, consistent with the dissociation argued for in Kiparsky 1993. As Kiparsky (forthcoming, 67) notes, ‘‘All di¤erences between the levels [Cyclic, Word, and Phrasal] involve the ranking of markedness constraints with respect to the antagonistic Faithfulness constraints. Each reranking is a minimal one, such as might constitute a dialectal isogloss, or which (if involving the postlexical phonology) could correspond to a sound change.’’ The NDEB-generating ranking, Faith-LEX Markedness Faith, is possible at any stratum, consistent with the fact, observed in Kiparsky 1993a, that NDEB e¤ects have been attested for Word and even Postlexical phonology (see e.g. Kiparsky’s discussions of Finnish Assibilation, Sanskrit Ruki, and Slave Continuant Voicing). In section 19.8 we will see examples of NDEB e¤ects at these strata in Catalan and Sanskrit. 19.4
Korean Palatalization
In this section I turn to my main case study, Korean Palatalization, tracing its evolution over the past 500 years and using the tools developed in the preceding sections to analyze its synchronic manifestation. Korean Palatalization is the phenomenon where dental stops (/t/ and /t h /) are realized as postalveolar a¤ricates ([c] and [c h ]) when followed by an /i/ or /y/ across morpheme boundaries.4 I will generally use ‘‘TI’’ from here on to refer to the set of sequences eligible for palatalization, and ‘‘CI’’ to refer to their palatalized counterparts.
466
Young-mee Yu Cho
The consonant inventory of Korean is given in (5): (5)
Plain Aspirated Tense
Labial
Dental
m
n L5 t th t’ s s’
p ph p’
Alveopalatal
Velar Ð
c ch c’
k kh k’
Korean Palatalization exhibits classical NDEB symptoms, applying only across morpheme boundaries. The examples in (6) show that Palatalization applies across derivational and inflectional su‰xes in both nominal and verbal morphology (see Cho and Sells 1995 for a summary of Korean morphology).6 Since Palatalization is a featural assimilation process, the di¤erence between /i/ and /y/ is of no significance; the only di¤erence is that the vowel occupies the syllabic peak while the glide does not: (6) Korean Palatalization a. Across derivational su‰xes mat-i ! maci ‘eldest-nml ¼ eldest one’ h tot-i ! h toci ‘sun rise-nml ¼ sunrise’ kat h -i ! kac h i ‘be_like-nml ¼ together’ kut-i ! kuci ‘be_firm-adv ¼ firmly’ ‘adhere-caus ¼ to a‰x’ put h -i- ! puc h i mut-hi- ! muc h i ‘bury-pass ¼ to be buried’ b. Across inflectional su‰xes ‘field-nom ¼ field’ pat h -i ! pac h i ky t h -i ! ky c h i ‘side-nom ¼ side’ pat h -ita ! pac h ita ‘field-cop ¼ to be the field’ e
e
For arguments that the endings in (6) are a‰xes, rather than clitics or postpositions as is sometimes assumed, see Cho and Sells 1995. The characteristic NDEB behavior of Korean Palatalization is illustrated by the examples in (7), in which morpheme-internal TI sequences (underlined) are immune to Palatalization: (7) ti titi-ta t hi n t h i-namu
‘where’ ‘to tread’ ‘blemish’ ‘zelkova tree’
e
A Historical Perspective on Nonderived Environment Blocking
467
The domain of Palatalization is morphologically restricted beyond simply showing NDEB e¤ects; it applies only within the prosodic word. It is well known in Korean phonology (Kang 1992; Han 1994) that the root plus any su‰xes forms a domain called the prosodic word; prefixes constitute prosodic words of their own for purposes of syllabification and other prosodically oriented processes. Prefixed words are thus parallel to compounds and word sequences in consisting of two prosodic words. There are several existing proposals regarding the formal derivation of prosodic word structure; I will set this issue aside for the present discussion.7 The examples in (8) illustrate the failure of Palatalization to apply across prosodic domain junctures. Instead, either nothing happens and the TI sequences are preserved, or optional /n/-insertion (Cho 1998) occurs. Palatalization is not an option: (8) Lack of Palatalization across prosodic domain juncture a. Prefix þ root i. hot h -ipul ! hotipul, honnipul ‘single-blanket ¼ single-layer blanket’ ii. t t-yaÐmal ! t tyaÐmal, t nnyaÐmal ‘outside socks ¼ outer socks’ b. Compounds pat h þ ilaÐ ! patilaÐ, pannilaÐ ‘field þ ridge ¼ ridge of a field’ c. Word boundaries kot il na ! kotil na, konnil na ‘soon get_up ¼ get up soon’ e
e
e
e
e
e
The minimal pair in (9) illustrates the crucial role played by prosodic domains. When the noun /pat h / ‘field’ is combined with the comitative su‰x /-ilaÐ/, Palatalization applies, but when /pat h / is compounded with the homophonous word /ilaÐ/ ‘ridge’, Palatalization does not apply. Instead, we find either neutralization of the syllable-final /t h / to [t], or optional /n/- insertion: (9) Minimal pair /pat h -ilaÐ/ ! [pac h iraÐ] /pat h þ ilaÐ/ ! [pat][iraÐ], [pan][niraÐ]
‘field-com’ ‘ridge of a field’
Morpheme-internally, plain TI and palatalized CI sequences are fully contrastive in Contemporary Korean: (10) TI and CI in nonderived environments TI CI tik t ‘the letter t’ ci c ‘the letter c’ titi‘to step on’ cici‘to burn’ t h i t ‘the letter th’ c h i c h ‘the letter ch’
468
Young-mee Yu Cho
‘where’ ‘grass’ ‘joint’
cil pcanc h i maciki e e
ti canti mati
‘to feel dizzy’ ‘feast’ ‘patch of field’
e
It is only across morpheme boundaries within a prosodic domain that the contrast is not permitted. The NDEB e¤ects associated with Korean Palatalization have been well known for a long time. Ahn (1985) attributes the NDEB properties of Palatalization to the SCC; Iverson and Wheeler (1988) argue for reviving the RAC (see section 19.7 for discussion). Kiparsky (1993) uses prespecification and underspecification to distinguish morpheme-internal TI sequences, which do not undergo Palatalization, from heteromorphemic TI sequences, which do. Oh (1995) puts forward a related prosodic account in which the stem-final /t/ is distinct from stem-medial /t/ due to its extraprosodicity. In her analysis, Palatalization targets only unsyllabified segments. Oh’s approach shares common traits with Lubowicz’s (1998) local conjunction approach of treating a stem-final segment as misaligned, although, as Inkelas (2000) points out, there are several problems associated with focusing exclusively on the stem-final segment as the target of NDEB, since in some languages, derived environment alternations a¤ect stem-initial segments. 19.5
An Analysis
This chapter presents a new characterization of the NDEB e¤ects exhibited by Korean Palatalization. One advantage of the proposed account over its predecessors is its ability to extend to the diachronic development as well as the synchronic state of Palatalization. I begin with the synchronic facts, which I will analyze using the framework laid out in section 19.3. The markedness constraint responsible for palatalization is Pal, which penalizes dental stops before high front vocoids; ranked above Faith-[ant], Pal results in palatalization. We may assume a further markedness constraint, A¤ricate, which requires [þanterior] segments to be [þdelayed release]. The e¤ects of Pal are inhibited by the higher-ranking Faith-LEX, which protects lexical sequences from palatalization: (11) a. PAL TI sequences are prohibited. b. AFFRICATE [þanterior] segments are [þdelayed release]. c. FAITH There should be correspondence between the input and the output.
A Historical Perspective on Nonderived Environment Blocking
469
i. IDENT-[ant] Corresponding segments are identical in [anterior]. ii. IDENT-[del-rel] Corresponding segments are identical in [delayed release]. d. FAITH-LEX There should be identity between a lexeme and the corresponding output. (categorically assessed) e. Ranking Faith-LEX Pal Affricate Faith The interaction of these constraints is illustrated in (12), which shows how Palatalization applies in the su‰xed form /mat-i/ ‘eldest-nom’ but is blocked from applying to the monomorphemic word /mati/ ‘joint’. (For simplicity, Affricate is not shown; the two Ident constraints are collapsed under Faith.) The analysis is parallel to that given above for English TSS: (12) NDEB in Korean Palatalization: Cyclic stratum /mati/
Faith-ROOT
+ a. mati b. maci
Pal
Faith
* *!
*
/mat-i/ a. mati + b. maci
*! *
The underived form, /mati/, is subject to Faith-ROOT, thus preserving the unpalatalized /ti/ sequence; by contrast, the derived form, /mat-i/, is not protected by Faith-ROOT and must obey Pal. The question as to why only su‰xation creates a derived environment for Palatalization in Korean can be answered straightforwardly in the Faith-LEX analysis. Prefixation and compounding create Word boundaries; at the Word stratum, Pal is ranked lower than Faith, and, as a result, Palatalization is not applicable. 19.6
History of Palatalization
Careful examination of the rise and the development of Palatalization reveals an interesting fact: Palatalization did not initially exhibit NDEB. In fact, the morphological restriction came about only when a second sound change introduced phonological opacity into the grammar.
470
Young-mee Yu Cho
I begin with a look at the consonantal inventory of Middle Korean of the fifteenth and sixteenth centuries. As is well established in the history of Korean phonology, Middle Korean had no palatal a¤ricates (Huh 1964; K.-M. Lee 1972; Kim-Renaud 1974). The letters that represent Modern Korean postalveolar a¤ricates (( , , ) were dental a¤ricates, as described in phonological treatises such as Hwunmincengumenhay (1459) and Sasengthongkopemlyey (1517) (K.-M. Lee 1972, 65). These treatises clearly specified that the letters represented sounds with the same place of articulation as such dentals as [t, s, n], although they were pronounced as a¤ricates. Another piece of evidence, introduced by K.-M. Lee, lies in loan phonology. Mongolian loanwords into Middle Korean that started with the palatal [‰] were transcribed using both a dental a¤ricate and the glide [y] to denote palatality: (13) Mongolian loanwords into Korean (thirteenth century) Mongolian Korean Gloss gal‰an kantsyam l ‘horse with white forehead and cheeks’ ‰e’erde tsy ltam l ‘red horse’ v
v
e
By contrast, alveopalatal consonants in loans into Contemporary Korean are consistently adopted as palatal a¤ricates without the help of the glide [y], as illustrated in (14). The di¤erence is clearly attributable to the change in place of articulation of these a¤ricates. (14) English loanwords into Korean (twentieth century) English Korean jersey c ci juice cus jean cin chocolate c h ok h oles charming c h amiÐ e
Another piece of evidence that Middle Korean lacked palatal a¤ricates comes from conservative dialects that still lack them. The Northwestern dialects of Korean retain the original dental a¤ricates. Interestingly enough, these dialects also lack Palatalization: (15) Standard Korean caÐma cicin maci (/mat-i/) pac h i (/pat h -i/)
Northern dialects tsaÐma tsitsin mati pat h i
‘rainy spell’ ‘earthquake’ ‘eldest’ ‘field-nom’
Finally, even in the standard dialect, the dental place of articulation is preserved in a¤ricates preceding the high back unrounded vowel /ı/, as illustrated in (16).
A Historical Perspective on Nonderived Environment Blocking
471
(16) /ts/ before [i] ([c] elsewhere) /yoc m/ yots m ‘recently’ /-c’ m/ ts’ m ‘approximately’ ts h Ð ‘floor’ /c h Ð/ Between the fifteenth and late seventeenth centuries, two sound changes occurred: backing of dental a¤ricates in all contexts and Palatalization of dental stops in the high front vowel context. (17) Sound changes in Middle and Late Middle Korean a. Backing of dental a¤ricates ts ! c b. Palatalization of dentals t, t h ! c, c h / i, y The relative chronology of these two sound changes is not firmly established, but it is often assumed that (17a) preceded (17b), although it is tempting to think that backing of dental a¤ricates to a palatal location might have been initiated in a palatal context and then generalized to the other environments. Yu (1824) reported the rise of /ci/ in di¤erent contexts in the seventeenth century; there is, however, no clear evidence as to whether Palatalization of dental a¤ricates was unconditional or whether it was initially triggered by limited assimilation. What is known is that both changes were complete by the end of the eighteenth century; all dental a¤ricates had changed to alveopalatals before high front vocoids, both morpheme-internally and across morpheme boundaries. Examples are shown in (18): (18) Palatalization of dentals before /i/ (fifteenth > eighteenth-century Korean) a. Morpheme-internally tik h i- > cik h i‘to keep’ moti- > moci‘to be cruel’ tyoh- > cyoh-, coh‘to be good’ ty psi > cy psi, c psi ‘plate’ b. Across a morpheme boundary k t-hi > k c h i ‘like, as’ ‘friend-nom’ p t-i > p ci e
e
e v
e
e v
To model this sound changes formally, let us first characterize the grammar of Middle Korean of the fifteenth and sixteenth centuries, when Palatalization was dormant. The anterior value in the underlying representation of TI sequences surfaced intact in the output, as guaranteed by a ranking in which the Pal constraint ranks below Ident-[ant]:
472
Young-mee Yu Cho
(19) Stage 1: Fifteenth–sixteenth centuries (no Palatalization) /ti/
Ident-[ant]
+ a. ti b. ci
Pal *
*!
/tsi/ a. ci
*!
+ b. tsi
*
/tþi/ + a. ti b. ci
* *!
The second stage came about sometime in the seventeenth century, when Palatalization began to apply optionally. The strict ranking between Ident-[ant] and Pal did not hold any longer. With the undoing of this ranking, an input TI sequence could surface either as TI or as CI, regardless of whether the environment was morphologically derived. We may assume that by this time /ts/ had changed to /c/. (20) Stage 2: Seventeenth century (optional Palatalization) /ti/
Ident-[ant]
+ a. ti + b. ci
Pal *
*
/ci/ a. ti
*!
*
+ b. ci /tþi/ + a. ti + b. ci
* *
The literature notes that Palatalization was an optional rule for some time (K.-M. Lee 1980; Choi 1992), supporting an analysis of language change in which
A Historical Perspective on Nonderived Environment Blocking
473
ranked constraints become unranked (Anttila and Cho 1998). Importantly, both morpheme-internal and morpheme-final /t/’s were subject to optional Palatalization. The next stage is characterized by obligatory Palatalization. At this stage Pal is ranked above Faith (represented by Ident-[ant], as shown in (21)): (21) Stage 3: Early nineteenth century (obligatory Palatalization) /ti/ a. ti
Pal
Ident-[ant]
*!
+ b. ci
*
/ci/ a. ti
*!
*
+ b. ci /tþi/ a. ti
*!
+ b. ci
*
There were apparently some di¤erences in the way Palatalization targeted di¤erent strata of the lexicon. As illustrated in (22), obligatory Palatalization was established much faster in the Native Korean stratum than in the Sino-Korean stratum, where variation lasted for some time longer; this scenario supports Itoˆ and Mester’s (1995b, 1999) approach to lexical stratification solely in terms of constraint ranking. Palatalization ultimately spread from the core of the lexicon to all lexical strata (Cho 2001): (22) Palatalization in di¤erent strata (K.-M. Lee 1980; Choi 1992) (eighteenth century) a. Native Korean vocabulary ti-ta ! ci-ta ‘to lose’ ‘collapse’ munh ti-ta ! munh ci-ta tut-ti ! tut-ci ‘to hear’ b. Sino-Korean vocabulary tyosy n @ cos n ‘Chosun’ koktyo @ kokco ‘melody’ tyoty k @ toc k ‘thief ’ e
e
e
e
e
e
When Palatalization was completed, morpheme-final etymological T was realized either as T or C, depending on whether the following su‰x started with /i/.
474
Young-mee Yu Cho
Morpheme-internally, however, Palatalization had the e¤ect of absolute neutralization. Just as children learning Modern English have no reason to posit [x] in nightingale or a [kn] cluster in knife, post-Palatalization Korean learners had no reason to posit an underlying di¤erence between morpheme-internal TI and CI sequences, since both would always surface as CI. (23) represents Stage 4, at which inputs simplified due to the completion of the Palatalization sound change. (Note that certain OT assumptions about the Richness of the Base (Smolensky 1996) would render this stage vacuous.) (23) Stage 4: Early nineteenth century (reanalysis of /TI/ as /CI/) /ci/
Pal
Ident-[ant]
*!
*
a. ti + b. ci /tþi/ a. ti
*!
+ b. ci
*
With Palatalization firmly established by the end of the eighteenth century, the nineteenth century witnessed another sound change that interacted with Palatalization by providing it with a potential new target. In the nineteenth century the diphthong / ‘i/ began to lose its diphthongal quality and to surface as [i], as shown in (24). (24) Nineteenth century: Monophthongization of /i ‘i/ to [i] mun ‘i > muni ‘pattern’ pon ‘i > poni ‘inside skin of a nut’ > ky nti‘to endure’ ky nt ‘imut ‘i> muti‘to be dull’ t’ ‘i > t’i ‘belt’ t ‘i > ti ‘where’ n t h ‘i-namu > n t h i-namu ‘zelkova tree’ e
e
e
e
Stage 5 is modeled in (25) as a constraint reranking, assuming a constraint *Diph( ‘i) to ban the relevant diphthongal sequence. After some period of fluctuation (again obtained by the unranking of the two constraints), Stage 5 is captured as the ranking of *Diph( ‘i) above the faithfulness constraint preserving input / /. (25) Stage 5: Nineteenth century (monophthongization of /i ‘i/ to [i]) a. *DIPH(i ‘i) The tautosyllabic sequence / ‘i/ is prohibited.
A Historical Perspective on Nonderived Environment Blocking
475
b. MAX(i ‘) Do not delete the input / ‘/. c. Ranking *Diph( ‘i) Max( ‘) The tableau in (26) shows how the new ranking simplifies input diphthongs: (26)
/p ‘i/ a. p ‘i
*Diph( ‘i)
Max( ‘)
*!
+ b. pi
*
Diphthongs preceded by a dental onset present a more challenging case. Given the ranking established for Palatalization in (23), the predicted interaction between Monophthongization and Palatalization is one of feeding: the surface [i]s created by Monophthongization should trigger palatalization of a preceding dental. This predicted interaction is illustrated in (27), where [ci] emerges as the winning output for input /ti/. Given that *Diph( ‘i) must outrank Max( ‘), and given that Pal must outrank Ident-[ant], there is no way for the candidate [ti] to win. (27) *Diph(i ‘i) Max(i ‘), Pal Ident-[ant] /t ‘i/
Max( ‘)
Pal
b. ti
*
*!
+ c. ci
*
a. t ‘i
*Diph( ‘i)
Ident-[ant]
*!
*
The prediction that [ci] is the correct outcome for input /t ‘i/ does not, however, correspond to what actually happened in Korean; as shown in (28), the result of Monophthongization following T was a surface TI sequence, opaque from the perspective of Palatalization. The interesting result is that Palatalization applies in morphologically derived environments but not in phonologically derived environments. In the derivational framework, this represents counterfeeding: the structural context of one rule (Palatalization) is potentially satisfied due to the application of a prior rule (Monophthongization), but cannot apply due to rule ordering. (28) Counterfeeding relationship between Monophthongization and Palatalization /t ‘i/ — Palatalization ti Monophthongization [ti] (opaque surface form)
476
Young-mee Yu Cho
Numerous approaches to opacity have been proposed; here we resort to an analysis utilizing sympathy (McCarthy 1999). In sympathy theory, a selector (‘‘¶’’) constraint identifies one ‘‘sympathy’’ candidate to which all others are required, by a sympathy constraint, to be faithful in some respect. The selector constraint in this case is ¶Max-V, which is satisfied by output candidates that preserve input diphthongs. In (29), the input is /t ‘i/. The sympathy (‘‘{’’) candidate is the most optimal (according to the rest of the constraint ranking) of those candidates preserving the input diphthong, namely [t ‘i] and [c ‘i]. Of these, [t ‘i] best satisfies faithfulness; they tie on Pal, since neither contains a Palatalization environment. Thus [t ‘i] is the sympathy candidate. The sympathy constraint, {Ident-([ant]), requires output candidates to be identical in consonant anteriority to the sympathy candidate ([t ‘i]). The candidates satisfying the sympathy constraint are [ti] and [t ‘i]; of the two, [ti] is optimal since it best satisfies *Diph( ‘i), and [ti] is therefore judged the winning candidate. Because Pal and Ident-[ant] are both ranked below the sympathy constraint, neither plays a prominent role in selecting the winner in this tableau. (29) Stage 5: Opaque interaction of Palatalization and Monophthongization /t ‘i/ { a. t ‘i
*Diph( ‘i)
¶Max( ‘)
{Ident-[ant]
Pal
Ident-[ant]
*!
+ b. ti
*
c. ci
*
* *!
*
In Contemporary Korean there is no longer any trace of / ‘i/ in most syllables originally containing it; Monophthongization has occurred in all environments except in word-initial onsetless syllables, as illustrated for Standard Korean in (30). (30) Tautosyllabic etymological /i ‘i/ across dialects Standard Southwestern Southeastern sa ‘i-sa isa hok ‘i-hok ihok ‘i-mi imi mi
‘doctor’ ‘doubt’ ‘meaning’
e e e
Some dialects, mainly the Southern ones, have simplified even these forms to yield monophthongal forms. Assuming that the first half of the [ ‘i] diphthong constitutes an onset to the second half, these dialectal di¤erences can be modeled by two di¤erent rankings: Onset *Diph( ‘i) versus *Diph( ‘i) Onset. When Monophthongization had run its course, there was no longer any reason to assume an underlying / ‘i/ diphthong anywhere except, of course, in initial onsetless syllables. The residue of historical Monophthongization is thus the source of one of the three types of synchronic exceptions to Palatalization. The first type, which we
A Historical Perspective on Nonderived Environment Blocking
477
have just seen, are the Native Korean words that etymologically contained the sequences /t ‘i/ or /t h ‘i/, as shown in (31). (31) Morpheme-initial /ti, t h i/ from ti < t ‘i titi-ta < t ‘it ‘i-ta < t h ‘i t hi h n t i-namu < n t h ‘i-namu mati < mat ‘i canti < cant ‘i
/ti ‘i, t h i ‘i/ ‘where’ ‘to tread’ ‘blemish’ ‘zelkova tree’ ‘joint’ ‘grass’
e
e
All the tautomorphemic nonpalatalizing TI sequences mentioned thus far in the chapter belong to this class of words. While for the forms in (31) there is no concrete synchronic evidence of an underlying diphthong, there is another set of data for which the sympathy analysis is still plausible synchronically. When the two vowels, / / and /i/, are concatenated across a morpheme boundary, it is possible to pronounce these as two syllables (i.i) or coalesce two vowels into one syllable (i); in this one context, there is still synchronic evidence for Monophthongization as an active process. Synchronic Monophthongization gives rise to the second type of counterexample to Palatalization. As shown in (32), a /t -i/ sequence that Monophthongizes to /t-i/ does not Palatalize: (32) Synchronic variation /t’ -i-ta/ t’ ita @ t’ita /t h -i-ta/ t h ita @ t h ita
‘to catch the eye’ ‘to be open’
The sympathy analysis of the diachronic counterfeeding relationship between Monophthongization and Palatalization in (29) (or its equivalent) is also required in a synchronic grammar of Korean to account for the failure of Palatalization in the monophthongal variants of the words in (32). The third set of exceptions to Palatalization consists of foreign loans introduced in the twentieth century; some examples are presented in (33). Anterior consonants in the source language (here, English) are systematically realized as dentals in Korean, regardless of context, resulting in unpalatalized sequences such as [ti, t h i]: (33) Lack of Palatalization in foreign vocabulary ticit h al ‘digital’ t hi ‘tea’ t h ina ‘Tina’ mot h ip ‘motive’ pati ‘body’ t h ipi ‘TV’
478
Young-mee Yu Cho
The two sets of surface exceptions to Palatalization for which there is no concrete evidence of an underlying diphthong are responsible for the NDEB e¤ect for which Palatalization is renowned, illustrated again in (34). (34) Stage 6: Contemporary Korean /maci/ a. mati
Faith-ROOT
Pal
Faith
*!
*
*
+ b. maci /mati/ + a. mati b. maci
* *!
*
/mat-i/ a. mati + b. maci
*! *
In lieu of the tautomorphemic example mati ‘joint’ in (32), we can easily substitute an example from the loan vocabulary such as t h i (‘tea’) with the same result, with no special ranking provision for recent loans. In both cases, a root containing an internal TI sequence is subject to Faith-ROOT in the Cyclic Stratum. We have seen clear historical reasons for the morpheme-internal nature of synchronic exceptions to Palatalization in Korean: morpheme-internal /ti, t h i/ sequences are immune to Palatalization (because they were not eligible to undergo the acrossthe-board Palatalization sound change), while morphologically derived environments regularly undergo it.8 Although careful investigations are needed, other NDEB cases seem to be analogous to Korean Palatalization in their provenance. Chamorro Vowel Lowering, another derived-environment rule, is reported to apply to ‘‘virtually all native Chamorro words’’ except in a handful of native words and a large class of loans (Chung 1983). This is the exact configuration in the Korean lexicon. The Korean pattern reflects the historical facts that all dentals at one point in history changed to palatals before [i] and that those synchronically available TI sequences are products of another independent process or recent borrowing. The idea that words showing NDEB are ‘‘marked’’ or ‘‘marginal’’ is also supported by other languages. Exceptions to Finnish Vowel Coalescence (Anttila, chap. 18, this volume) and Polish First Velar Palatalization (Lubowicz 1998) consist exclusively of recently borrowed words. These languages have a more restricted set of
A Historical Perspective on Nonderived Environment Blocking
479
exceptions than do Korean and Chamorro, in that the original sound change, undergone by all native forms, was never obscured by a later process in the way that Korean Monophthongization rendered Palatalization opaque. Kiparsky (1982, 77, b) makes a similar observation that NDEB exceptions are restricted about Sanskrit, noting, of the retroflexing ‘‘Ruki’’ rule in Sanskrit, that ‘‘/»/ is the highly normal case in the Ruki environment even in nonderived environments. Words like /bisa/ ‘lotus’ and /kusuma/ ‘flower’ are marginal in the Sanskrit lexicon. In contrast, words such as /vi»a/ ‘position’ and /kar»/ ‘drag’ are quite common.’’ Korean Palatalization confirms the hypothesis that NDEB is not an inalienable property of a particular phonological alternation per se but is derivable from constraint interaction. Korean Palatalization was at one point an across-the-board process that admitted no exceptions. Even when the subsequent introduction of exceptions turned Palatalization into a ‘‘derived-environment’’ rule, CI sequences still remained unmarked relative to TI sequences, just as [i»] in Sanskrit was unmarked, relative to [is], even in nonderived environments. NDEB e¤ects are not intrinsically necessary; they result from the introduction of an external source of opaque exceptional forms, either through additional language change or through borrowing.9 19.7
Postlexical Palatalization
In addition to the lexical Palatalization of dental stops discussed up to this point, Korean also has an automatic postlexical palatalization process. Before a high front vowel, all coronals are palatalized phonetically, resulting in prepalatalized segments (35a). As shown in (35a) versus (35b), the postlexical Palatalization process is, predictably, insensitive to morphological boundaries. The palatalization of /s, n, l/ (35c) is also generally attributed to the postlexical manifestation of the lexical Palatalization rule: (35) Postlexical Palatalization t¸i a. ti t hi t¸ h i t’i t¸’i b. os-i osˇi nun-i nunˇi kot¸il na kot il na c. si s˘i su¨psˇu¨ps’inun sˇinun k’ini k’inˇi p’alli pa··i ariarˇi-
‘where’ ‘blemish’ ‘belt’ ‘clothes (nom)’ ‘eye (nom)’ ‘soon get up’ ‘poem’ ‘to be easy’ ‘germ’ ‘meal’ ‘fast’ ‘to hurt’
e
e
e
e
480
Young-mee Yu Cho
Iverson and Wheeler (1988) argue explicitly that lexical and postlexical palatalization result from one and the same rule. When the result of applying the rule lexically would be neutralizing, the RAC limits the rule’s application to derived environments. By contrast, when the result of applying Palatalization is a purely allophonic alternation, the RAC plays no role, thus accounting for the across-the-board application of Palatalization in the postlexical stratum. There is, however, a fundamental di¤erence between lexical and postlexical Palatalization that imperils a single-rule account. When dentals undergo Palatalization at the Cyclic stratum, they become alveopalatal a¤ricates ([c], [c h ]); when they undergo postlexical Palatalization, they become prepalatal, as illustrated in (35a). The di¤erence between the two Palatalization processes can be accounted for by ranking faithfulness di¤erently in the Cyclic and postlexical strata. In the Cyclic stratum, as we have seen, Pal outranks both Ident-[ant] and Ident-[del-rel], giving rise to a dental ! alveopalatal alternation. In the postlexical stratum, however, dentals become prepalatal. Their value for [anterior] changes, but their value for [delayed release] does not. We can attribute this to a reranking of the markedness constraint Affricate and the faithfulness constraint Ident-[del-rel]. The rankings at the two strata are illustrated in (36): (36) a. Cyclic stratum: Dentals become alveopalatal a¤ricates before high front vocoids
Pal Ident-[ant] Ident-[del-rel] Affricate b. Postlexical stratum: Dentals become prepalatal before high front vocoids Pal Ident-[ant], Affricate Ident-[del-rel] Affricate An illustration of allophonic palatalization at the postlexical stratum is provided below: (37)
/ti/ a. ti
Pal
Ident-[ant]
Ident-[del-rel]
Affricate
*!
+ b. t j i
*
c. ci
*
* *!
All sequences in Korean obey Pal at the postlexical stratum, where Faith-LEX is ranked too low to interfere. This analysis relates the two palatalization processes, attributing both to Pal, but captures their di¤erences through di¤erent rankings of the constraints a¤ecting the optimal means of satisfying Pal.
A Historical Perspective on Nonderived Environment Blocking
481
One more fact remains to be discussed before we leave the section on Korean Palatalization. Oh (1995) claims that Palatalization depends on morphological information only, and that application of a phonological rule does not create a derived environment, as evidenced by the fact that Glide Formation in /te- / ! [ty ] ‘to burn’ does not trigger Palatalization of [t] even though the phonological environment is met. e
e
(38) Lack of Palatalization in a phonologically derived environment /te- / ! ty (Vowel Raising) ! *cy (Lack of Palatalization) e
e
e
This argument presupposes that the /e/ of /te/ is being directly converted to /y/, potentially feeding Palatalization. However, in fact Glide Formation (GF) only applies to /i/; the relationship between /e/ and /y/ in (38) is mediated by lexical variation between /e/ and /i/ in this particular morpheme. As shown by the data below, GF is possible only for morphemes that show this variation; for morphemes that do not show the variation, GF is not possible. (39) Variations in Glide Formation and Vowel Raising a. Lexical variation in vowel height: GF possible /te-/ @ /ti-/ ! /ty/ ‘to burn’ /pe-/ @ /pi-/ ! /py/ ‘to cut’ b. Lack of lexical variation in vowel height: No GF /he- / *hi *hy ‘to count’ /k nne- / *k nni *k nny ‘to hand over’ /e- / *i *y ‘to scrap out’ e
e
e e
e
e e e
e
e
e e
GF merely reflects di¤erent syllabic organization between [i] and [y]; its application should not be expected to a¤ect the applicability of Palatalization. The reason that Palatalization does not apply to the morphemes in (39a) is that the relevant TI sequences are morpheme-internal, protected by Faith-LEX. 19.8
Sanskrit RUKI
In this section, I discuss a slightly more complex case of NDEB that involves two kinds of derived environments—one morphologically derived and the other phonologically derived. Because of evidence that these two notions of derivedness do not always pattern together, some formulations NDEB have attempted to state the principle behind NDEB e¤ects in purely morphological terms. Most notably, Hammond (1992) proposes to derive NDEB e¤ects from a constraint on the acquisition of structurechanging morphologically conditioned rules. Hammond assumes that phonological rules cannot create a derived environment.
482
Young-mee Yu Cho
In the present framework, however, there is no reason to arbitrarily limit the scope of NDEB by distinguishing phonologically derived environments from morphologically derived environments since both kinds can be derived from the same ranking schema. Although Korean Palatalization only presents positive evidence of being conditioned by morphological derivedness, the basic ranking schema used for Korean NDEB e¤ects extends in principle to both phonologically and morphologically derived environments. Faith-LEX takes care of morphologically derived environments by giving special status to a lexeme. Phonologically derived environments are handled by introducing another markedness constraint (M1 , in (40b)), which, if ranked high enough, potentially interacts with the process in question (40b). (40) a. Morphologically derived Faith-LEX M Faith b. Phonologically and morphologically derived M1 Faith-LEX M2 Faith The scenario in (40b) is clearly attested in Sanskrit. According to the well-known Ruki rule, /s/ retroflexes following any of the four segments /r, u, k, i/: (41) Sanskrit RUKI rule (Kiparsky 1982d) s ! » / {r, u, k, i} The rule is blocked in morphologically simple cases (42a) but applies in a number of morphologically derived contexts (42b): (42) a. Morphologically underived contexts barse ‘tip’ kusuma kisalaya ‘sprout’ pis b. Morphologically derived cases da-da-si ‘you give’ bi-bhar-»i kram-sya-ti ‘he will go’ vak-»ya-ti sena-su ‘armies’ agni-»u
‘flower’ ‘move’ ‘you carry’ ‘he will say’ ‘fires’
These familiar e¤ects are easily modeled using the framework developed in this chapter: at the stratum in which the a‰xes in (42b) are attached, the ranking Faith-LEX Ruki Faith obtains; ‘‘Ruki’’ represents the markedness constraints responsible for /s/ retroflexion. Ruki is triggered not only in morphologically derived environments, however, but also in phonologically derived environments. Ablaut is a morphophonological process that results in vowel change (43a) and vowel loss (43b), among other e¤ects, in morphological Ablaut environments. As shown in (43), the application of Ablaut can feed the Ruki rule:
A Historical Perspective on Nonderived Environment Blocking
483
(43) Zero Grade Ablaut and RUKI (Kiparsky 1982d, 76) a. /a/ ! /i/ /sas-ta/ ! si»a ‘taught’ b. /a/-loss /ja-ghas-anti/ ! jak»anti ‘eat (reduplicated 3. pl.)’ c. /vas/ ! u» ‘shine, dwell’ /va-vas-us/ ! u»us ‘shone, dwelt’ d. /-vas/ ! -u» (gen. sg.) /vid-vas-as-/ ! vidu»as ‘knowing’ In an analysis like Hammond’s, which draws inherent distinctions between ‘‘phonologically derived’’ and ‘‘morphologically derived’’ environments, it matters a great deal whether Ablaut in Sanskrit is a phonological process or not. For our purposes, this question need not be raised at all. Ablaut simply involves markedness constraints that figure in the computation of optimal outputs. The following tableaux illustrate the interaction of Ablaut and the Ruki rule. I assume a constraint, or constraint family, ‘‘Ablaut,’’ containing the markedness constraint(s) responsible for triggering the range of e¤ects shown in (43). Since Ablaut applies morpheme-internally, Ablaut must outrank Faith-ROOT, yielding the following overall ranking: (44) Ablaut Faith-ROOT Ruki Faith In Ablaut environments, satisfaction of Ablaut neutralizes Faith-ROOT by disrupting lexeme-output identity. Once Faith-ROOT is violated, there is nothing to prevent satisfaction of the Ruki constraints. In the tableau below, (45a) is a case where the context for Ablaut is not met; Faith-ROOT is satisfied at the expense of the lowerranked Ruki constraint. By contrast, (45b) is an Ablaut environment. Satisfying Ablaut means violating Faith-ROOT; the candidate satisfying both Ablaut and Ruki is therefore the winner: NDEB in Ruki (Ablaut Faith-ROOT Ruki Faith)
(45) a.
/bi-bhar-si/[Ablaut]
Ablaut
bibharsi
Faith-ROOT
Ruki *!
+ bibhar»i b.
Faith
*
/ja-ghas-anti/[þAblaut] jagha»anti jaksanti + jak»anti
*! *!
* **
484
Young-mee Yu Cho
The fact that both morpheme concatenation (42b) and the application of Ablaut (43) create a ‘‘derived environment’’ for Ruki is given a unified account in terms of constraint ranking. Both are environments in which Faith-ROOT is violated, removing the impediment to satisfaction of the lower-ranked markedness constraint, in this case, Ruki. Turning briefly to a diachronic perspective, we note that Ruki was a word stratum process in Rigvedic Sanskrit but became restricted to being a cyclic process in Classical Sanskrit (Kiparsky 1997a). (46) summarizes the change in terms of Faithfulness promotion, a very common trajectory of historical change. (46) a. Rigvedic Ruki applies across a‰x, compounding, and clitic boundaries, but not phrasally across word boundaries. Cyclic stratum: Faith-ROOT Ruki Faith Word stratum: Faith-STEM Ruki Faith Phrasal stratum: Faith-PWD, Faith Ruki (no Ruki) b. Classical Ruki applies across a‰x boundaries but not across compound, clitic, or word boundaries. Cyclic stratum: Faith-ROOT Ruki Faith Word stratum: Faith-STEM Faith Ruki (no Ruki) Phrasal stratum: Faith-PWD, Faith Ruki (no Ruki) The present study can also naturally be extended to the so-called SCC e¤ects in phrasal phonology (Rice 1990). For instance, Sanskrit and Catalan final voicing, triggered by a sonorant, applies between words in both compounds and phrases, but it does not apply within words (47a, b). (47) The ‘‘SCC’’ e¤ects in voicing assimilation a. Catalan (Mascaro´ 1987) to[t] [r]ic ! to[dr]ic ‘all rich person’ se[t] [ma]ans ! se[dm]ans ‘seven hands’ (cf. temple [-pl-] ‘temple’, submari [-pm-] ‘submarine’) b. Sanskrit (Whitney 1889) sa[t] [a]ha ! sa[da]ha ‘good day’ ta[t] [n]amas ! ta[dn]amas ‘that homage’ (cf. putra [-tr] ‘son’ marutþi ! maru[ti] ‘wind-locative’) c. Faith-PWD Voicing Assim Faith The SCC makes word-internal structure unavailable at the phrasal stratum just as the NDEB makes root-internal structure unavailable at the cyclic stratum. The ranking in (47c) accounts for the integrity of an isolated prosodic word, in contrast to the variation in phrasal concatenation.
A Historical Perspective on Nonderived Environment Blocking
19.9
485
Conclusion
In the chapter, I have shown that the development of Korean Palatalization follows the change in constraint ranking shown in (48). History reveals no inherent relationship between Palatalization and NDEB. NDEB e¤ects developed well after the onset of Palatalization, modeled here by a reranking of Faith-LEX above the triggering Pal constraint: (48) Historical development of Korean Palatalization a. Across-the-board Palatalization: Pal Faith-ROOT, Faith b. NDEB: Faith-ROOT Pal Faith Two additional constraint-ranking schemas, involving markedness and faithfulness constraints, have also been validated by data from Korean and Sanskrit: (49) NDEB rankings a. Faith-LEX M Faith b. M1 Faith-LEX M2 Faith
(M satisfied only in morphologically derived environments) (M2 satisfied in both morphologically and phonologically derived environments)
The factorial typology based on the four-constraint system in (49b) predicts three patterns: (i) application in all environments (M Faith-LEX, Faith), (ii) application only in derived environments either phonologically or morphologically (M1 Faith-LEX M2 Faith), and (iii) application only in morphologically derived environments (Faith-LEX M Faith). Application only in phonologically derived environments cannot be described in this system, which I believe to be a desirable consequence. Although there are a number of cases in which only morphologically derived environments are considered derived, there are few or no cases where only phonologically derived environments, but not morpheme combinations, count in NDEB.10 Notes Earlier versions of this chapter were presented at colloquia in 2001–2002 at University of Massachusetts, Rutgers University, and Cornell University. Over the course of working on this topic I have benefited from discussions with Arto Anttila, John Alderete, Gregory Iverson, John McCarthy, Joe Pater, Alan Prince, John Whitman, and Draga Zec. I am particularly grateful to the editors of this book. Abbreviations: adv ¼ adverbializer, caus ¼ causative, com ¼ comitative, cop ¼ copula, gen. sg ¼ genitive singular, nml ¼ nominalizer, nom ¼ nominative, pass ¼ passive, 3. pl ¼ third-person plural.
486
Young-mee Yu Cho
1. Korean Palatalization can be decomposed into two independent processes, lexical a¤rication and postlexical palatalization. Section 19.7 deals with the di¤erences between the two. I retain the term Palatalization because of its wide usage. 2. The idea of Faith-ROOT constraints was introduced in McCarthy and Prince 1995, and has been taken up in many subsequent works in OT. 3. Whether TSS represents a group of ranked constraints or one monolithic constraint will not be discussed here. 4. Kim (1999) argues that the a¤ricates assumed to be alveopalatal or palatal are, in fact, alveolars and that Palatalization is an a¤rication process. Clearly there seem to be a range of variations in place of articulation for these a¤ricates. However, even if it turns out that the process in question is not of change of place but that of manner, the overall point made in this chapter will still remain valid. 5. There is only one liquid phoneme in the language, represented here by ‘‘L.’’ It surfaces as /l/ in the coda position or as part of a geminate and as [r] in onset position. See Cho 1997 for a detailed discussion on the distribution of this liquid. 6. Korean has a productive phrasal process of obstruent voicing between two sonorants that render [mati] and [maci] as [madi] and [maji]. I ignore this process for the sake of simplicity of representation here. 7. In particular, Kang (1992) proposes an end-based account in which left-edge alignment creates proper morphological bracketing. 8. I also predict that loanwords that end with a dental would undergo Palatalization when they are combined with an i-initial su‰x. This is in line with the loss of exceptionality under derivation demonstrated in several other languages—for example, certain words in Catalan do not undergo vowel reduction when una‰xed but do undergo reduction when a su‰x is attached (Kiparsky 1993a; Inkelas 2000). One potential example in Korean is the English word pot. All borrowed nouns ending in a dental stop are optionally realized as [s]-final before the vowel-initial nominative, as in [pas-i]. The alternative is to apply vowel epenthesis, as in [p h at h -ka], which results in the selection of the /ka/ nominative su‰x. 9. A remaining outstanding question is: What caused the change from across-the-board application to an NDEB process? In OT terms, the question can be rephrased as ‘‘What caused the rise of Faith-ROOT?’’ 10. See, however, McCarthy 2003b, which presents a system for generating precisely this kind of NDEB e¤ect.
20
Lexical Storage and Phonological Change
Geert Booij
20.1
Introduction
Empirical investigations of the division of labor between storage and computation in language behavior can be executed in a number of ways. An obvious one is that of psycholinguistic experimentation. Another way, which will be focused on in this paper, is language change. Language change is a psycholinguistic laboratory of nature, a window on how speakers produce and interpret language. For example, if a language loses a phonological rule while the e¤ects of that rule are preserved in a number of words, a possible explanation is that the outputs of that rule must have been stored at the stage when the phonological rule was still active and thus survived after the loss of that rule. The question that I will address in this paper is how far phonological change provides evidence for the kinds of phonological information about lexical items stored in lexical memory. That is, whereas Kiparsky in his early work on phonological change since 1965 (compiled in Kiparsky 1982a) focused on phonological change as evidence for the structure of the grammar, I will take a di¤erent, extragrammatical perspective and ask what we can learn from phonological change about lexical memory. Asking this question is also in line with Kiparsky’s conclusion at the end of his book Explanation in Phonology that linguistic change does not provide a window on the structure of the grammar as directly as was hypothesized in Kiparsky 1968c: ‘‘Before we can exploit historical evidence for synchronic purposes we need a firm theory of the intervening factors’’ (Kiparsky 1982a, 234). Lexical storage is certainly one of these intervening factors that deserve more detailed investigation. A necessary preliminary remark is that the issue of storage versus computation with respect to a specific regularity of a language is not a matter of ‘‘either . . . or.’’ The conclusion that a particular linguistic form must be stored in the lexicon does not preclude the existence of a rule that accounts for most or all of the properties of that form. This position has not been a standard one in Generative Grammar, which has always been strongly influenced by the Bloomfieldian view of the lexicon as the basic list of irregularities. For instance, Kenstowicz (1994, 60) motivates the claim that predictable information is not stored lexically as follows.
488
Geert Booij
(1) ‘‘Generative grammar’s answer to this question is based on the hypothesis that the human capacity for language is designed in such a way as to minimize the amount of information that must be stored in the speaker’s mental lexicon.’’ However, a few pages later in the same book, Kenstowicz (1994, 69–70) points out that this point of view is no longer self-evident: ‘‘with the advent of neural science and more accurate estimates of the capacity of the human brain, this ‘‘economy of storage’’ argument is not compelling in and of itself.’’ What I suspect is that the storage argument given in (1) has never been a serious consideration. Rather, the aim has always been to give an elegant and formally as simple as possible analysis of the distribution and alternation patterns of a language. In the area of Generative Morphology, linguists have always been aware of the possible and even necessary simultaneity of rules and stored outputs. Many complex words, once formed, must be lexically stored because they have unpredictable formal and/or semantic properties. Nevertheless, the rules that created these complex words may still be productive. Jackendo¤ (1975) therefore advocated a view of morphological rules in which these rules function as redundancy rules with respect to existing complex lexical items, but this does not exclude their creative use for the coining of new complex words. That is, predictable information can be lexically stored, and redundancy rules tell us which part of that information is predictable, that is, does not count as independent information. Similar views were defended by Arono¤ (1976) and Booij (1977a, 1977b). In sum, we should avoid what has been called the ‘‘rule/ list fallacy’’ by Langacker (1987, 29), namely, the reasoning in which lists and rules are mutually exclusive. In fact, the relation between lexical storage and morphological rules is even stronger and of a more principled nature: a morphological rule does not exist without a set of listed words instantiating that rule. The native speaker first acquires complex words. It is on the basis of recurring patterns in sets of similar complex words with a systematic pairing of form and meaning that the speaker may conclude the existence of a morphological rule, which then may result in extension of the set of words of that particular form.1 This view of morphology presupposes that words (and idioms), not morphemes, are the units of lexical storage. The role of morphemes in the analysis of word structure is of a secondary nature: they play a role in establishing the relations between words but are not the primary building blocks of complex words.2 As mentioned above, in Generative Grammar the prevailing tendency has been to reduce the lexical storage of allomorphy as much as possible. Allomorphy is accounted for in terms of one underlying form for each morpheme and a set of rules for the computation of the actual surface allomorphs. In addition, predictable phonetic details are omitted from lexical representations: it is only phonemic distinctions,
Lexical Storage and Phonological Change
489
that is, contrastive phonetic properties, that are encoded in lexical representations. The adage is what can be computed, should not be stored.3 This relates to the wellknown abstractness controversy: the more abstract our phonology is, the more we can derive di¤erent surface forms from the same underlying representation. Linguists have tried to tackle this issue by looking at phonological change (Kiparsky 1968c), whereas psycholinguists have been trying to solve this problem by means of psycholinguistic evidence (cf. Lahiri and Marslen-Wilson 1992). In the beginning of the 1970s, Bybee and Vennemann argued in favor of a concrete view of phonological representations, partially on the basis of facts of language change (Hooper 1974b; Vennemann 1974). Generative phonologists, however, di¤er in the degree to which they strive for redundancy-free representations, as shown by the debate on underspecification. For instance, in Dutch the velarity of a nasal is contrastive word finally but not before a velar obstruent. Should we therefore omit the specification for place in the lexical representation of velar nasals before velar stops? This is not obvious, and Anderson (1985, 136f ) has warned us that the idea that linguistic representations should be redundancy free is by no means self-evident, and probably wrong. The principle of ‘‘Lexicon Optimization’’ advocated by phonologists working within Optimality Theory (Archangeli and Langendoen 1997) also implies that redundant, nondistinctive phonological information will be stored in lexical representations: the faithfulness condition implies that inputs should di¤er minimally from the corresponding outputs (see Archangeli and Langendoen 1997, 201¤ ).4 Although there are di¤erences in the degree of abstractness that generative phonologists allow for, most generative phonologists assume that at least the e¤ect of automatic phonological rules should not be encoded in the lexical representations of morphemes and words, because they are always computable on the basis of purely phonological information. However, we should not take this position for granted given the storage capacity of human memory. Even if one is willing to accept the storage of the e¤ects of phonological rules, the information stored may still be abstract in the sense that it is phonemes that are stored and not the actual details of the phonetic realization of these phonemes. This is what most generative phonologists assume, even those who advocate a concrete kind of phonology: the phonetic details, such as the acoustic parameters for a particular vowel, are not specified as such in lexical representations but accounted for by the set of language-specific rules of phonetic implementation. However, this position has recently been attacked, for instance in Flemming (1995) and Bybee (2000b). This issue will be returned to in section 20.5. In sum, there are three kinds of phonetically relevant information with respect to which the lexical representation is a point of discussion: (i) predictable phonetic properties that have a contrastive function in some contexts, such as the place of
490
Geert Booij
articulation of nasals; (ii) the e¤ects of phonological processes that create alternations; and (iii) details of phonetic realization. The point of departure for our discussion is that we know that the vastness of the lexical memory allows us to store much of what can be computed, even though it is completely regular, since it will speed up processing. In the area of morphology, there is solid evidence for the storage of frequent plural nouns in a number of languages (Baayen et al. 1997). The same applies to the syntactic level. There are thousands of noun phrases of the type adjective þ noun, such as yellow pages, red tape, green card, black hole, hard disk, little toe, that are listed in our lexical memory because they function as names for things (see Jackendo¤ 1997, chapter 7); yet this does not imply that we do not have a productive rule for the construction of such noun phrases. Given these general considerations about the possibility of storage, and given that there is no conflict between storage and regularity/productivity, it is important to ask ourselves how we can find out about the realities of storage without a priori positions as to what is memorized in the form of lexical representations and what in the form of rules. In this chapter, I will use facts of phonological change as one possible type of evidence for storage and provide evidence for surface forms of words being stored. Another general issue that should be addressed before going into the details of phonological change is that it is often taken for granted that the notions ‘‘lexical representation’’ in the sense of ‘‘the form of a word that is the starting point of morphological operations’’ and ‘‘lexical representation’’ in the sense of ‘‘stored information about a lexical item’’ coincide. This is not a logical necessity, however. Take a simple case like the following pair of nouns from Dutch. (2) hoed ‘hat’ [hut], plural hoeden [hud n] e
The standard assumption is to assume /hud/ as the lexical representation for hoed and derive the phonetic form of the singular from this underlying form by means of the rule of Auslautverha¨rtung. However, it is also logically possible that we store [hut] as the singular form and in addition the basic form /hud/, based on the alternation [hut]/[hud n]. It is even possible that we do not store an underlying form like /hud/ at all and only compute it from the stored phonetic form if we need to do so because we want to apply a morphological operation to that word. In other words, we then reverse the classical analysis: phonetic forms are stored, and underlying forms are computed. This option will be discussed in more detail below. e
20.2
Phonemicization of Allophonic Properties
A number of Germanic languages have seen the loss of the stop in word-final clusters of velar nasal þ velar stop. For instance, Middle Dutch coninc [ko:nIÐk] ‘king’ be-
Lexical Storage and Phonological Change
491
came koning [ko:nIÐ] in Modern Dutch. Thus, the property of the nasal consonant that it is velar before a velar stop lost its allophonic nature and became phonemicized: words ending in [Ð] now contrast with words ending in [m] or [n]. Does this phonological change tell us something about the lexical storage of the place of articulation of the nasal before the following stop got lost and thus about the storage of redundant properties? In a classical rule-based analysis of phonological change, the answer is negative: the nasal can be underspecified at the underlying level (usually graphically represented as N). Two rules apply: the rule that spreads place of articulation from the stop to the nasal and, subsequently, a rule of cluster simplification that deletes the stop (the addition of this rule is the change in the grammar of the adult). The surface form of coninc is therefore [ko:nIÐ]. The child will then store this surface form as the underlying form, thus giving phonemic status to the velar nasal. This account of phonological change makes a sharp distinction between the two generations involved in a change: the adult generation and the next generation that has to acquire the language system. It is doubtful, however, whether such a sharp distinction is justified. It is quite clear that adult speakers also continuously subject the outputs of their language system to reanalysis, that is, there is a continuous inspection of output forms (see Bynon 1977; Hopper and Traugott 1993). The possibility of inspection and reanalysis presupposes that these output forms are stored: have a certain degree of permanence in memory. Moreover, this also enables us to explain phonological change. In the present example, the loss of a final obstruent, the speaker can conclude that the final obstruent can be omitted without the loss of distinction between words, because the place of articulation of the nasal will su‰ce to distinguish a word like zing [zIÐ] ‘to sing’ from zin [zIn] ‘sense’. Specification of place of articulation of the nasal in the word-final consonant cluster is also necessary for theory-internal reasons if we take a constraint-based approach to phonology instead of the traditional rule-based approach. The rulebased analysis makes crucial use of rule ordering of a particular type, nonbleeding order, and of the assumption that a phonological change is to be seen as a rule that is always added at the end of the grammar. If we take an Optimality-theoretical constraint-based approach, it is immediately clear that the velarity of the nasal should be part of the lexical representation even before the loss of the final obstruent: an underspecified representation will induce an extra violation of faithfulness compared to a fully specified lexical representation, which goes against the principle of Lexicon Optimization. The surfacing of the velar nasal is then straightforward: the change involved is that the constraint *CC-Velar (a constraint belonging to the family of constraints on consonant clustering) that forbids velar consonant clusters is ranked higher than Max-IO (a faithfulness constraint that requires identity of input and output).
492
(3)
Geert Booij
ko:nIÐk
*CC-Velar
ko:nIÐk
*!
Max-IO
+ ko:nIÐ
*
If an underspecified lexical representation (forbidden by the principle of Lexicon Optimization, at least if no alternation is involved, see Inkelas 1995) had been assumed for the velar nasal, we would predict the underspecified nasal to surface as coronal [n] since this is the default value for nasal consonants. (I assume that default values are expressed by markedness constraints such as Coronal: the default place of articulation of consonants is coronal, and the place of articulation must be specified, the constraint of Full Specification). (4)
/ko:nINk/
Full Specification
ko:nIÐk ko:nIN
*!
ko:nIÐ
*CC-Velar
Coronal
*!
*
Max-IO
*
*
*!
*
+ ko:nIn
*
In conclusion, given a constraint-based analysis, the only way the velar nasal will surface after the diachronic process of velar cluster simplification is by specifying the place of articulation of this nasal consonant at the level of lexical representation. We should note, however, that this kind of evidence for lexical storage of predictable properties is theory dependent, since, as we saw above, a rule-based analysis did not imply lexical storage of the place of articulation of the nasal. 20.3
Phonologization and Lexicalization: Vowel Lengthening in Dutch
Another relevant case of phonologization of allophonic properties is that of vowel lengthening in open syllables in Early Middle Dutch. This process a¤ected both simplex nouns and complex nouns and even wordþclitic combinations. (5) a. Simplex nouns name ‘name’ smake ‘taste’ stave ‘sta¤ ’
n[a:]me sm[a:]ke st[a:]ve
Lexical Storage and Phonological Change
b. Singular sch[I]p ‘ship’ h[ ]l ‘hole’ oorl[ ]g ‘war’ d[þ]g ‘day’ c. Word þ clitic saetic /sþt Ik/
493
Plural sch[e:]p-en h[o:]l-en oorl[o:]gen d[a:]g-en [sa:tIk] ‘sat I’
c
c
The singular-plural pairs in (5b) have been preserved in modern Dutch (see Booij 1995; Kager, this volume). This process of vowel lengthening is a manifestation of a much more general tendency in Germanic languages toward a requirement that stressed syllables must be heavy, that is, they cannot end in a short vowel but must contain minimally either a long vowel or a short vowel followed by a consonant (also known as Prokosch’s Law).5 In the first stage of this process in Early Middle Dutch (see Van Loon 1986, 89), the lengthening of the vowel was an allophonic process. In the course of time (probably in the 12th century), however, the lengthened vowel was phonemicized, that is, it became an underlying long vowel. How do we know this? The first evidence concerns short vowels followed by geminate consonants. These were protected from lengthening because the first half of the geminate closes the syllable. In the 12th century, degemination took place. Yet, the short vowels that thus ended up in open syllables did not lengthen anymore, and so we find many words with short vowels followed by only one consonant, such as wikke [VIk ] ‘vetch’. That is, the following development took place. e
(6) VCV > V:CV (predictably long vowel) > V:CV (phonemic long vowel) VCCV > VCCV (predictably short vowel) > VCV (phonemic short vowel) These changes are illustrated by the following minimal pair (Van Loon 1986, 89). Modern Dutch week [we:k] wikke [wIk ] e
Early Middle Dutch weke [we:k ] wikke [wIkk ] e
e
(7) Early Germanic wika [wIka] ‘week’ wikkia [wIkkia] ‘vetch’
The classical generative interpretation of this kind of phonemicization is as follows (Kiparsky 1968b): The initial change is the addition of a rule of degemination to the phonological system of the adult speakers. This rule of degemination made the alternation between long and short vowels in the output forms opaque; vowel lengthening is ordered before degemination, and this is a nonbleeding order. The next generation of language users therefore interpreted the length contrast as a phonemic contrast, and hence the length contrast became part of the lexical representation. In addition, the process of vowel lengthening in open syllables disappeared.
494
Geert Booij
A second kind of evidence for the phonemicization of these initially allophonic contrasts for simplex nouns is provided by the later process of schwa apocope (thirteenth century). This phonological change did not a¤ect the length of the vowel in the simplex nouns ending in schwa: after apocope the vowels of words such as naam ‘name’ (
Lexical Storage and Phonological Change
495
phonetically similar forms as instantiations of the same morpheme and for the computation of the phonetic forms of new words. For those words for which there was no alternation involved, such as weke versus wikke given in (7), the assumption of storage in memory of the output forms by adult speakers is also necessary. Only if we assume that predictable length di¤erences are stored can we explain why the rule of degemination was added to the adult grammar: the addition of degemination has been made possible by the length di¤erence, because the latter could now serve to keep words phonetically distinct. The language user is able to survey and inspect the phonetic output forms of words in memory and can thus conclude that the distinction between single and geminate consonants is redundant, given a concomitant distinction in vowel length. The emergence of degemination can therefore be seen as the result of storage in speakers’ memory of output forms. Similarly, the possibility of schwa apocope may also be seen as the result of storage of the phonetic output forms: since the vowel length distinction was kept in memory, the word-final schwa could disappear. This then made it necessary to consider vowel length as information that is part of the lexical representation. Given this analysis, we avoid constructing a complete gap between two generations of language speakers and are not forced to assume that language change is caused only by the imperfect learning of the next generation that has to acquire the language on the basis of outputs. Language change is also an e¤ect of the adult language user who is able to analyze the phonetic forms of words that are accessible in memory. The consequence of the loss of transparency of a phonological process (rule opacity) is that it is no longer possible to compute or store an underlying form for a word that di¤ers from its phonetic form (unless an alternation is involved). A distance between phonetic form and underlying form is only possible in the case of surface-true transparent processes. In the case at hand this means that the plural forms of these nouns, with their long vowels, must be stored for new generations as well since their phonetic forms cannot be computed. This is also the conclusion of Kager (chap. 17, this volume). Since in Dutch the singular noun is identical to the stem that is the basis for morphological operations, it is the short vowel that appears in such cases. Derived word gebr[]kkig ‘handicapped’ h[ ]lletje ‘diminutive’ sch[I]pper ‘skipper’ g[ ]ddelijk ‘divine’ sp[]lletje ‘diminutive’ w[]ggetje ‘diminutive’ c
Plural noun gebr[e:]k-en h[o:]l-en sch[e:]p-en g[o:]d-en sp[e:]l-en w[e:]g-en
c
(8) Singular noun gebr[]k ‘handicap’ h[ ]l ‘hole’ sch[I]p ‘ship’ g[ ]d ‘god’ sp[]l ‘game’ w[]g ‘road’ c
c
496
Geert Booij
In a number of cases the alternation that existed between singular and plural nouns disappeared after the loss of the rule of vowel lengthening (so-called analogical leveling or paradigmatic leveling). This applies to, for instance, the following Dutch nouns, which have a short vowel in their stem-final syllable in both singular and plural forms. Plural noun bissch[ ]pp-en l[]kk-en gem[a]kk-en str[a]¤-en c
(9) Singular noun bissch[ ]p ‘bishop’ l[]k ‘leak’ gem[a]k ‘ease’ str[a]f ‘punishment’ c
The traditional interpretation of these facts is that since the plural forms of these words were not stored, the regular forms with short vowels will show up after the loss of the rule of vowel lengthening. However, since the position is taken here that plural forms can be stored in their output form, the following interpretation is called for: the original plural forms of these words with long vowels did have a lexical representation of their own, but their frequencies, and thus their resting level of activation, was not high enough to block the formation of a regular plural form without vowel lengthening, that is, on the basis of the underlying form of the singular noun. This is also the point of view defended in Wetzels (1981, 95–97), as to analogical leveling: analogical leveling is nothing but the application of productive rules that is not blocked by the existence of stored forms. 20.4
Auslautverha¨rtung and Lexical Storage
So far, our findings with respect to the lexicalization of vowel lengthening are in line with the conclusion of Wetzels mentioned above. However, he argued that paradigmatic leveling takes place only if rules are no longer automatic phonological rules and hence require storage of allomorphs. Only if rules are no longer phonologically conditioned will the di¤erent allomorphs of a morpheme be lexically stored, which then may give rise to leveling by using the ‘‘wrong’’ allomorph. Since, according to Wetzels, the outputs of automatic phonological rules are not stored, they will not lead to paradigmatic leveling. For instance, the alternations created by the automatic rule of devoicing of obstruents in coda position have not, with a few exceptions to be discussed below, been leveled in Dutch. Therefore, we will now have a more detailed look at coda devoicing. It is an uncontroversial assumption within mainstream Generative Phonology that the e¤ects of automatic neutralization rules are not encoded in lexical representations. For instance, Kenstowicz and Kisseberth (1979, 49) argue on the basis of the Russian rule of word-final devoicing of obstruents that it would be wrong to list both
Lexical Storage and Phonological Change
497
the allomorph with a voiced obstruent and the one with a voiceless obstruent in the lexicon and to assume a selection rule for the allomorphs stated as follows. (10) ‘‘If a morpheme has alternants that di¤er with respect to the voicing of a final obstruent, select the alternant with a final voiceless obstruent when the morpheme appears at the end of a word; otherwise, select the alternant with a voiced obstruent.’’ The authors then add that the ‘‘basic criticism is that this sort of analysis fails to adequately characterize the rule-governed nature of the voicing alternation in Russian’’ (p. 49). That is, such an analysis implies that two rules have to be assumed for Russian: a phonological rule of devoicing and a morphological rule of allomorph selection. Language change again provides a window on this issue: if we do not lexically store the outputs of a neutralization rule (with the e¤ect that each output is a possible underlying form), we predict that paradigmatic leveling will not take place. I will therefore have a look at a process similar to Russian devoicing: coda devoicing of obstruents in Dutch. Final devoicing in Dutch is a productive generalization: voiced obstruents cannot occur in codas. If there are two allomorphs for a lexical item with an alternation between a voiced and a voiceless obstruent, it is generally assumed that the underlying form that is stored is the form that ends in a voiced obstruent. If this voiced obstruent ends up in coda position, it is devoiced by the rule of final devoicing. (11) rib [rIp] / ribben [rIb n] ‘rib, sg/pl’ hoed [hut] / hoeden [hud n] ‘hat, sg/pl’ slaaf [sla:f ] / slaven [sla:v n] ‘slave, sg/pl’ kiez [kis] / kiezen [kiz n] ‘molar, sg/pl’ vlieg [vlix] / vliegen [vli' n] ‘fly, sg/pl’ e
e e
e
e
Interestingly, in some cases the alternation is lost. For instance, although we have the alternation hand/handen ‘hand, sg/pl’ [hþnt]/[hþnd n], there is an idiomatic phrase bijdehand [bid hant] ‘lit. at the hand, bright’ that functions as an adjective. One of its inflected forms in prenominal position is bijdehante [bid hant ], that is, there is no alternation between voiced and voiceless obstruent anymore. In Van Loey (1964, 54) the rise of forms such as bijdehante is interpreted as a case of paradigmatic leveling (analogy), which would suggest that the allomorph [hþnt] of hand is stored. However, the problem for this account is that it does not explain why paradigmatic leveling almost never takes place with respect to e¤ects of final devoicing. Another interpretation is therefore called for. The language learner only computes an underlying form that di¤ers from the phonetic one if (i) the two words involved are related, and (ii) the two surface forms are relatable by means of e
e
e
e
498
Geert Booij
a transparent rule. In the case of bijdehand, with its inflected form bijdehante, this adjective has no transparent formal relation to the word hand with its plural form handen, and thus it will be stored with a final /t/. In other words, we can only get underlying forms that di¤er from the surface forms if there is a transparent morphological relation between the words in which the di¤erent allomorphs occur (see Wetzels 1981). Thus, the form bijdehante is not a counterexample to Wetzels’s 1981 claim that automatic phonological rules do not lead to leveling. Another interesting case is the word stad [stþd] ‘town’ with the irregular plural form steden [ste:d n]. The plural noun not only exhibits vowel lengthening but also vowel quality change, from /a:/ to /e:/. Consequently, the underlying form of stad is apparently computed as /stþt/, as shown by new coinings, such as the verb statten ‘to do shopping in town’ derived through conversion from the noun stad ‘city’ and the inhabitant name Lelystatter derived from the toponym Lelystad. A similar phenomenon can be seen in Afrikaans for a number of nouns that have plural forms with lengthened vowels in Dutch. Apparently, the distance between smid ‘blacksmith’ and smeden (plural), for instance, is too large for speakers of Afrikaans, and thus they do not conclude a common stem /smId/. Thus, the lexical form of smid will be /smIt/ rather than /smId/, and consequently the regular plural smitten has arisen. e
(12) Dutch Afrikaans smid [smIt] / smeden [sme:d n] smit [smIt] / smitten [smIt n] ‘smith, sg/pl’ god [' t] / gotten [' t n] ‘god, sg/pl’ god [' t] / goden ['o:d n] lid [lIt] / leden [le:d n] lit [lIt] / litten [lIt n] ‘member, sg/pl’ rat [rþt] / ratten [rþt n] ‘wheel, sg/pl’ rad [rþt] / raderen [ra:d r n] e
e
ec
c
e
e
c
e
e
ee
The point here is that precisely the nouns with the unproductive vowel length alternation exhibit paradigmatic leveling with respect to the still transparent rule of coda devoicing. In the case of the words with the vowel-length alternation discussed above, there is also internal evidence for the storage of the e¤ect of coda devoicing in the lexical entry, because vowel length plays a role in the distribution of fricative consonants. The generalization is that /v/ and /z/ occur only after long vowels, the ‘v/z-constraint’.6 Interestingly, we find the following alternations. graven ['ra:v n] hoven [ho:v n] glazen ['la:z n] staven [sta:v n] verloven [vrlo:v n] e
e e e
e
(13) Nouns graf ['rþf ] hof [h f ] glas ['lþs] staf [stþf ] verlof [vrl f ]
‘grave, sg/pl’ ‘court, sg/pl’ ‘glass, sg/pl’ ‘sta¤, sg/pl’ ‘permission, sg/pl’
c
c
Lexical Storage and Phonological Change
499
Verbs, past tense, SG/PL las [lþs] lazen [la:z n] genazen [' na:z n] genas [' nþs] gaf ['þf ] gaven ['a:v n] vergaf [vr'þf ] vergaven [vr'a:v n] Noun-verb pairs draf [drþf ] ‘trot’ draven [dra:v n] lof [l f ] ‘praise’ loven [lo:v n]
‘read’ ‘cured’ ‘gave’ ‘forgave’ ‘to trot’ ‘to praise’
e e e
e
e
e
e
e
c
In all these cases, short vowels are followed by a voiceless fricative, whereas long vowels are followed by a voiced fricative. In other words, all allomorphs obey the constraint on the distribution of /v/ and /z/. In these cases, the voicelessness of the final obstruent is also part of the underlying form, although this information is predictable and although these obstruents alternate with voiced ones. We know this on the basis of the forms of complex words derived from such nouns: these words are derived either from a form with short vowel þ voiceless obstruent or from a form with long vowel þ voiced obstruent; the combination short vowel þ voiced obstruent does not occur. (14) Denominal word formation — h[ f ]elijk ‘polite’ best[þf ]ing ‘sta¤ ’ l[ f ]elijk ‘praiseworthy’
gl[a:]zenier ‘stained glass artist’ h[o:]veling ‘courtier’ st[a:]ven ‘to prove’ l[o:]ven ‘to praise’
*glazzenier *hovveling *bestavving *lovvelijk
c
c
On the other hand, if the noun ends in a stop or a velar fricative, this restriction does not apply, and the morphology can apply to underlying forms with a voiced stop. Plural goden baden wegen
Derived word g[ ]ddelijk ‘divine’ b[þ]dderen ‘to bathe’ w[]ggetje ‘small road’ c
(15) Singular god bad weg
The v/z-constraint thus restricts the distance between phonetic form and underlying form. Van Loey (1964, 54–55) also mentions a number of cases in which leveling in favor of the allomorph with the voiceless obstruent took place, where no additional alternations are involved that make the two allomorphs di¤erent in other respects as well. Examples from 17th-century Dutch are (16) Middle Dutch gewaet ‘dress’ / gewaden ‘pl’ > 17th-century Dutch gewaeten Middle Dutch cieraet ‘ornament’ / cieraden ‘pl’ > 17th-century Dutch cieraeten In these cases the allomorphs with final [d] have been restored in Modern Dutch.
500
Geert Booij
There are also cases in present-day Dutch of this kind of leveling. An example is the adjective boud ‘bold’. Originally, the inflected form of this adjective was boude, but most speakers of Dutch say boute, which shows that the alternation [d/t] has been leveled out. What these examples show is that leveling in favor of the allomorph with the voiceless obstruent is possible. This kind of leveling can be due to imperfect learning: the language user did not yet compute the correct underlying form of the noun on the basis of the alternation between the singular and plural forms, although the alternation is transparent. Therefore, the underlying form is identical to the phonetic form of the singular noun. We may thus hypothesize that we store phonetic forms and that underlying forms are computed on the bases of stored phonetic forms (a similar position is taken in Leben and Robinson 1977 and Leben 1979). Underlying forms will di¤er from phonetic forms only insofar as that di¤erence follows from transparent phonological rules. If the distance between two forms is too big either semantically (the case of bijdehand ) or because the formal di¤erences are too big (stad-steden and the Afrikaans cases), the underlying forms will not di¤er from the surface forms. A structural constraint such as the Dutch v/z-constraint will also block the computation of an underlying form that is di¤erent from the surface form of a word, even though there is an automatic alternation involved. 20.5
Lexical Diffusion
In the early stages of generative phonology, phonological change was seen as the addition of a rule at the end of the grammar. That is, the lexical representations do not change, but the corresponding surface forms are a¤ected by the application of the added rule. It is only the new generation that will have di¤erent lexical representations with the e¤ect of the added rule lexically encoded, except when it is a rule that creates alternations. In the latter case, lexical representations will not necessarily change (Kiparsky 1968b). A problem for this view is that phonological changes in progress often a¤ect lexical representations but are also simultaneously the source of phonological alternations. For instance, Dutch is subject to a process of /d/-weakening in which intervocalically, /d/ is replaced with the glide /j/ before a following, su‰x-initial schwa (Booij 1995, 90). This causes alternations of the following type. (17) goed ‘good’, inflected form goed-e, phonetic form ['ud ] or ['uj ] rood ‘red’, inflected form rod-e, phonetic form [ro:d ] or [ro:j ] e
e
e
e
Application of this allomorphy-creating rule of /d/-weakening is lexically governed. For instance, it does not apply to the plural form of hoed-en ‘hats’ [hud n]; the phonetic form [huj n] is impossible. Weakening also applies optionally to the derived adjective goed-ig ‘good-natured’ ['ud x] or ['uj x] but obligatorily to the derived noun e
e
e
e
Lexical Storage and Phonological Change
501
goeierd ['uj rt] ‘good-natured person’; the form ['ud rt] is impossible. But other adjectives with stem-final /d/ do not always allow weakening. The adjective wreed ‘cruel’, for instance, does not have the inflected form [wre:j ] for wrede; the only form possible is [wre:d ]. Similar observations apply to processes such as /d/-deletion in Dutch (Booij 1995, 90). The lexically governed nature of these alternations implies that they must be stored lexically, even in the case of morphologically completely regular inflected forms of adjectives such as goed-e. The facts of /d/-weakening support Kiparsky’s (1988a, 1995) claim that it is lexical, that is, neutralizing rules that exhibit lexical di¤usion, since the distinction between /d/ and /j/ is phonemic in Dutch. What is essential from the perspective of this chapter is that they support the view that information provided by rule and information that is lexically stored are not mutually exclusive. These allomorphy facts imply that the recognition system of language users must have a certain robustness because they have no problems in relating goed to goeie and goeierd notwithstanding the phonological di¤erences in the lexical representation of the shared part of these words, the lexical morpheme goed. That is, apparently the phonetic forms of a morpheme do not have to be computed by rule from a common underlying form to be recognizable as allomorphs of that morpheme (contra Lahiri and Marslen-Wilson 1992). In the case of /d/-deletion an intervocalic /d/ is deleted and replaced with a predictable hiatus-filling glide. For instance, the inflected adjective oud-e ‘old’ can be pronounced as [ ud ] or [ uV ]. Both forms must be stored because it cannot be predicted which form allows for /d/-deletion. The phonetic di¤erence even correlates with a semantic di¤erence in the case of the adjective þ noun phrases oude hoer [ ud hu:r] ‘old whore’ (the literal meaning) versus ouwe hoer [ uV hu:r] ‘talkative person’. Labov (1981, 1994) proposed to distinguish two types of phonological change: change that is phonetically gradual and a¤ects all relevant words and change that is phonetically abrupt, replaces a phoneme with another one, and is lexically gradual, that is, exhibits lexical di¤usion. Kiparsky (1988a) argued that the distinction between phonetically gradual and phonetically abrupt changes coincides with the distinction between postlexical and lexical phonological rules. The rules of /d/weakening and /d/-deletion can indeed be considered lexical rules since they are neutralizing. As expected, they have exceptions and thus exhibit lexical di¤usion. Lexical di¤usion always creates surface opacity for rules since the speaker will find forms that have not undergone the rule. Therefore, as stated above, opacity will lead to lexical storage in the sense that for each phonetic form of such words a distinct lexical entry has to be created. This in its turn explains why semantic distinctions may correlate with phonological di¤erences, as in the pair oude hoer/ouwe hoer discussed above. e
e
e
e
e c
e c
e
c
e
c
502
Geert Booij
What about the e¤ects of postlexical rules? The usual assumption (see Kiparsky 1988a, 399) is that the e¤ects of such rules are not encoded in lexical representations. Bybee (2000a) has questioned these assumptions. She argues that a change may be both phonetically and lexically gradual. This implies that the e¤ects of such gradual changes must be lexically stored in the lexical representations of individual words. Vowel reduction in Dutch (Booij 1982, 1995; Van Bergem 1995) is a potentially interesting phenomenon in this respect because it has both lexical and postlexical properties. The basic generalization is that at the phonetic level all unstressed vowels are a¤ected by reduction in the sense that their actual realization can be quite far away from the target values of the acoustic parameters of these vowels (Van Bergem 1995). In addition, there is a lexical process of vowel reduction of vowels in unstressed (nonword-final) syllables that has already a¤ected many words whose full vowels have been replaced with schwa, as have the italicized vowels of the following words. (18) televı´sie ‘television’ a´lgebra ‘algebra’ serena´de ‘serenade’ recla´me ‘publicity’ beto´n ‘concrete’ repetı´tie ‘rehearsal, test’ confere´ntie ‘conference’ Morover, there is also a large number of words that exhibit vowel reduction in more casual speech only, that is, there is still alternation. [ba:na:n] / [b na:n] [po:lisi] / [p lisi] [minyt] / [m nyt] e
e e
(19) bana´an ‘banana’ polı´tie ‘police’ minu´ut ‘minute’
These facts of vowel reduction nicely fit into Kiparsky’s two-stage theory of phonological change: ‘‘the phonetic variation inherent in speech, which is blind in the neogrammarian sense, is selectively integrated into the linguistic system and passed on to successive generations of speakers through language acquisition’’ (Kiparsky 1995, 642). Vowel reduction, originally a purely phonetic proces motivated by ease of articulation, could become a lexical rule subject to lexical di¤usion because the schwa is a phoneme. As a lexical rule, it can be seen as a process in which the place of articulation features of vowels in unstressed syllables are removed from the lexical representations of words; subsequently, a default rule will fill in the relevant features of the schwa (see Kiparsky 1995, 642–647 for this interpretation of lexical di¤usion). Simultaneously, vowel reduction is a postlexical rule applying to vowels in syllables without lexical stress and a process of phonetic implementation for vowels (even those with lexical stress) that are not stressed in a particular utterance.
Lexical Storage and Phonological Change
503
The alternation between full vowel and schwa is also visible in related pairs of words that di¤er in whether the relevant vowel is stressed or not; if the vowel is unstressed it can or must reduce. The words in (20a) exhibit optional reduction of the vowel in italics, whereas in those in (20b) the vowel in italics can be realized only as schwa. (20) a. perso´on ‘person’ perce´nt ‘percent’ pasto´or ‘priest’ b. profe´et ‘prophet’ juwe´el ‘jewel’ geˆne ‘embarrassment’
pe`rsone´el ‘sta¤ ’ pe`rcenta´ge ‘percentage’ pa`stora´al ‘pastoral’ pro`fete´er ‘to prophesy’ ju`welı´er ‘jeweler’ gena´nt ‘embarrassing’
That is, the words in (20a) are subject to the postlexical rule of vowel reduction, whereas the words in (20b) exhibit lexical di¤usion e¤ects. This implies that a morpheme like juweel will have two di¤erent lexical representations: /jyVe:l/ when it is an independent word and /jyV l/ when occuring in juwelier. This does not cause any computational problem since words are the units of storage, and thus the two allomorphs will automatically appear in the right context. The only problem that the language user has to solve here is the recognition problem: how can both forms be recognized as forms of the same morpheme (a prerequisite for the semantic analysis)? But this is part and parcel of commanding a language since allomorphy that is not reducible to one underlying form is a widespread phenomenon, as we saw above (see also Booij 1997a, 1997b, 1998). Van Bergem (1995) rightly qualifies this process of vowel reduction as a sound change in progress that came into being due to interpretation by the native speaker of acoustic vowel reduction as a process of replacement of full vowel with schwa and exhibits lexical di¤usion, since in some words the unstressed vowel can be realized only as a schwa, as we saw above. Furthermore, this kind of reduction is boosted by high frequency: in high-frequency words, unstressed vowels are reduced more easily and more frequently than in low-frequency words. This is to be expected since vowel reduction reduces lexical contrasts and thus impedes word recognition. High frequency, on the other hand, boosts recognition and can thus compensate for the negative e¤ects of vowel reduction. The question then arises of whether the postlexical rule of optional vowel reduction in unstressed syllables is phonetically gradual or should be interpreted as the replacement in lexical representations of the full vowel with schwa by means of a phonological rule, just as in the case of the lexical rule of vowel reduction. If the first position is taken, this may have the further implication that the degree of reduction of the vowel is encoded in lexical representation. This position is argued for in Bybee (2000a,b) for some other cases of reduction and implies that details of the phonetic e
504
Geert Booij
realization of phonemes are lexically specified. Can we say something about this on the basis of Dutch vowel reduction? The data in Van Bergem (1995, 121) show that there is a positive correlation between frequency and acoustic parameters: in high-frequency words the formant values for the unstressed vowels are farther away from the target values than in low-frequency words. Moreover, this correlates with how native speakers perceive these vowels: ‘‘the average number of schwa responses increases when the spectral distance between the test vowel and its target increases’’ (Van Bergem 1995, 125). These results are explained if the growing distance between the acoustic parameters of unstressed vowels in words of high frequency and those of the nonreduced correlates is lexically encoded: each time a vowel is reduced, its formant values move away from those of the unreduced vowel. This is the kind of explanation advocated in Bybee (2000a). The theoretical implication of this step is that lexical representations are not redundancy free as far as phonetic details are concerned, a position also defended by Ohala and Ohala (1995). This position as to the lexical specification of vowel reduction e¤ects is supported by the observation that data concerning optional vowel reduction can be obtained by means of introspection: native speakers of Dutch systematically know that certain words are more susceptible to vowel reduction than others. For instance, they know that the /i/ in minuut ‘minute’ is susceptible to reduction, whereas the /i/ in pilo´ot ‘pilot’ is not. This suggests that such information is stored in memory. The conclusion to be drawn is that phonological rules may be productive and automatic, and yet at least some of their outputs are lexically stored because the process exhibits lexical di¤usion. In addition, it seems that phonetic details concerning the pronunciation of vowels can be stored, given the facts concerning the gradual erosion of unstressed vowels. Note, however, that this kind of stored information does not necessarily lead to a proliferation of distinct underlying forms (in the sense of ‘‘bases for morphological operations’’) for a morpheme. This is a crucial di¤erence between lexical and postlexical rules: it is only the e¤ects of lexical rules that may lead to more than one underlying form for a particular morpheme. 20.6
Conclusions
In this chapter, we saw that the standard view in Generative Phonology of the balance between storage and computation has to be reconsidered. There is a wealth of evidence for the position that predictable information is stored in the lexicon. First, recent theoretical developments in phonology imply that predictable information about morphemes must nevertheless be stored in the lexicon. Second, data concerning phonological change show that computable information concerning the phonetic realization of morphemes nevertheless has to be stored lexically. I also proposed that
Lexical Storage and Phonological Change
505
we should take a radical step with respect to the relation between underlying form and phonetic form: it is not the phonetic form that is computed by the speaker but rather the underlying form. Like storage in general, storage of phonetic forms of words will speed up processing; it is only when we coin a new word that computation of the underlying form of the base word is necessary. These conclusions do not refute the position that the human language faculty has a dual structure: a lexicon with stored representations and rules. The native speaker does need rules for the perception and production of novel forms. What, however, these conclusions do refute is the position that computation and storage of information with respect to the same process or regularity are mutually exclusive. Notes I would like to thank Joan Bybee, Sharon Inkelas, Jaap van Marle, and Leo Wetzels for their comments on an earlier draft of this paper. 1. A more radical position not taken in this paper is that of connectionism, in which the distinction between rules and representations is denied (see Pinker 1999 for discussion of this issue). 2. This is the standard view in the Dutch (and more generally, the European) morphological tradition (see Schultink 1962; Matthews 1974; Booij 1977; Van Marle 1985). It is also a cornerstone of Bybee’s morphological work (see Bybee 1988). 3. Early exceptions are Leben and Robinson (1977), Tiersma (1978), and Leben (1979), who argued that rules such as Trisyllabic Laxing in English are to be seen as parsing rules with primarily a redundancy rule function. 4. An additional argument for redundancy-free lexical representations has been that by omitting redundant properties and assigning these properties at the end of the phonological derivation by means of default rules, the inertness of such properties in phonological derivations follows. However, it has become clear that there is a substantial number of cases in which default or unmarked features such as the feature [coronal] for consonants do play a role in phonological processes, that is, are not inert (see McCarthy and Taub 1992; Booij 1993; Hall 1994). Note, however, that it may be the case that underspecification may be necessary in cases of alternations (Inkelas 1995). 5. See also Gussenhoven (this volume) and Kager (this volume) about vowel length in Dutch. 6. There are four exceptions: the loan words mazzel ‘good luck’, puzzel ‘puzzle’, razzia ‘raid’, and the name for railway kiosks wizzl, a form intentionally coined as a marked form with a high attention value.
21
The Dental Preterites in the History of English
Aditi Lahiri
21.1
Introduction
Any shift of lexical items from one morphological class to another can be couched in terms of analogical change. Kiparsky (1965, 1988, 2000) has argued that analogical change has the result of grammar simplification. This chapter addresses a case of a morphological split in the English dental preterite that, if an instance of analogy, appears to have the opposite e¤ect of complicating the grammar; we argue that, to the contrary, the change can be construed as phonological simplification when the grammar as a whole is taken into consideration. The morphological split in question involves the past tense in Middle English (discussed, e.g., by Marckwardt 1935, Minkova and Stockwell 1998). While involving a single regular inflectional paradigm in Old English, past tense formation in Middle English requires weak verbs to be split into two groups, one regular and the other irregular. The division is clearly observable in Modern English.1 In regular verbs (1a), there is no change in the root vowel. The past tense su‰x is d, which assimilates in voicing to the root consonant (as in reaped [ript]), or triggers the addition of [ ] if the root consonant ends in a coronal stop (as in treated [trit d]). In the irregular group (1b), the root vowel is shortened and the ending is a voiceless t (e.g., kept [kpt]), unless the root ends in a coronal stop, in which case there is no overt ending (e.g., fed [fd]). e
e
past moaned [d] healed [d] filled [d] reaped [t] begged [d] treated [ d] needed [ d] tended [ d] e e e
(1) Weak verbs in Modern English a. Regular: pres root vowel ¯ moan V ¯ heal V ˘ fill V ¯ reap V ˘ beg V ¯ treat V ¯ need V ˘ tend V
root vowel ¯ V ¯ V ˘ V ¯ V ˘ V ¯ V ¯ V ˘ V
508
Aditi Lahiri
b. Irregular:
pres mean feel keep meet feed hide send dwell
root vowel ¯ V ¯ V ¯ V ¯ V ¯ V ¯ V ˘ V ˘ V
past meant [t] felt [t] kep [t] met [Ø] fed [Ø] hid [Ø] sent [t] dwelt [t]
root vowel ˘ V ˘ V ˘ V ˘ V ˘ V ˘ V ˘ V ˘ V
Distinguishing between these two past tense verb formation processes requires positing two di¤erent past tense su‰xes in present-day English: /t/, accompanied by vowel shortening of the root, and /d/, with predictable assimilation in voice or insertion of schwa. The most convincing synchronic analysis of these patterns comes from the closed syllable account in Myers (1987b), who proposes that a bimoraic constraint is in force when the past tense su‰x /t/ is added. Assuming lexical levels (Kiparsky 1982a, 1982b, 1985), the analysis (which is based on Myers 1987) would be as follows. (2) Closed-Syllable Shortening (CSS) in Modern English Level I: Present tense and past tense /t/ (irregular verbs) Extrametricality CSS Coronal cluster simplification
keepþt keephti kephti
feedþt feedhti fedhti
meetþt meethti methti
—— [kpt] ‘kept’
fed [fd] ‘fed’
met [mt] ‘met’
Level II: Past tense /d/ (regular verbs) begþd —— —— [bgd] ‘begged’
moanþd —— —— [mond] ‘moaned’
needþd need[ d] —— [nid d] ‘needed’ e
reapþd —— reap[t] [ript] ‘reaped’
e
[ ]-insertion Voicing assim. e
It is only in the past tense that the two verb classes di¤er. There are no alternations when, for example, the 3p.sg su‰x s is added. In both sets of verbs, the su‰x consonant assimilates in voice to the preceding root consonant, and the surface forms are exactly parallel: keeps versus reaps; feeds versus needs; meets versus treats, and so on.
The Dental Preterites in the History of English
509
The evolution of the Modern English su‰xes /t/ and /d/ is relatively new. The past tense su‰x of Old English was /d/, and although there were alternations in the verbal paradigm, these were due to regular, transparent phonological processes such as ce¯pan-ce¯pte ‘keep’, gre¯tan-gre¯tte ‘greet’, me¯tan-me¯t ‘meet’, where [d] assimilates in voice to the root consonant, or fe¯dan-fe¯dde ‘feed’, fe¯lan-fe¯lde ‘feel’, hælan-hælde ‘heal’, fyllan-fylde ‘fill’ with the degemination of underlying geminates in the verb stem. The dual system characterizing Modern English, patently more complex than the simpler system in Old English, arose as the result of several changes, some analogical. We will argue that this apparent complication in the grammar of English can be understood as a process of grammar simplification if one assumes that neither phonological nor morphological analogy is to be interpreted only on the basis of surface forms but makes crucial reference to the entire grammatical system, including the status of lexical representations. The claim is that the split in the behavior of verbs like heal@healed and feel@felt is due to the fact that the Germanic dental preterite, precursor to the modern-day past tense su‰x d, used to be analyzed on par with derivational endings and stem extensions (theme vowels) rather than as an inflectional su‰x. As a result it was subject to a di¤erent set of constraints from those a¤ecting person/number/mood inflections. Eventually, the preterite su‰x split into two di¤erent su‰xes: a voiceless stop, added at level I of the morphology, which had a more restricted use, and a voiced stop, added at level II of the morphology, which became the regular past tense. This split is of particular interest in the context of the usual assumption in the grammaticalization literature in historical linguistics that a‰xes develop from clitics to unrestricted a‰xes to highly restricted a‰xes; the expectation in terms of level ordering would be that a level II a‰x would turn into a level I a‰x but not the reverse. To support the proposed split in the English dental preterite, we first explore the synchronic status of the dental preterite in Old English (section 21.2) and then examine the subsequent changes from Old to Middle English (section 21.3). 21.2
Weak Verbs in Old English
In Old English, phonological alternations in the weak verbal paradigms depended on the weight of the roots. Light roots consisted of a short vowel followed by a single consonantal coda, while all the others were heavy. (As will be discussed below, phonological alternations, such as consonant gemination, can obscure the di¤erences between light and heavy roots; see also Kiparsky and O’Neill 1976.) Most weak verbs were derived from nouns or adjectives by adding the derivational su‰x /j/ to the root. Here we focus primarily on the class I weak verbs, whose root vowels are umlauted by the /j/ su‰x. Their principal parts are listed in (3).
510
Aditi Lahiri
(3) Old English class I weak verbs derivational suffix /j/; inf /an/; 1p.pres.sg /e/; 2p.pres.sg /ist/; 3p.pres.sg /i/; 3p.past.ind /de/ light roots rootþ/j/ infinitive 1p.sg.pres.ind 2p.sg.pres.ind 3p.sg.pres.ind 3p.past.ind gloss
/framþj/ fremman fremme fremest freme fremede ‘perform’
/dinþj/ dinnan dinne dinest dine dinede ‘resound’
heavy roots /co¯pþj/ ce¯pan ce¯pe ce¯p(e)st ce¯p(e) ce¯pte ‘keep’
/fo¯dþj/ fe¯dan fe¯dan fe¯d(e)st fe¯d(e) fe¯dde ‘feed’
/fo¯lþj/ fe¯lan fe¯le fe¯l(e)st fe¯l(e) fe¯lde ‘feel’
/ha¯lþj/ hæ¯lan hæ¯le hæ¯l(e)st hæ¯l(e) hæ¯lde ‘heal’
/fullþj/ fyllan fylle fyll(e)st fyll(e) fylde ‘fill’
As seen in the table, there are several critical di¤erences in the behavior of light and heavy roots. a. In the infinitive and first person singular forms, final consonants of light roots undergo gemination, while final consonants of heavy roots do not. b. In the past tense forms, a mid vowel (phonetically schwa) appears before the preterite /d/ in verbs with a light root but not in verbs with a heavy root. c. In past tense forms with heavy roots, the preterite /d/ assimilates in voicing to the preceding, root-final consonant. This is the only context in Old English in which the coronal past tense marker is voiceless. d. The su‰x /j/ merges with the high vowels of -ist (2p.sg) and -i (3p.sg), blocking gemination. The resulting i is then lowered to e (phonetically schwa) as mentioned above in (b). The critical phonological processes that govern these alternations are gemination, high vowel deletion, glide deletion, degemination, voicing assimilation, and lowering. Crucial to an understanding of how these processes apply is the assumption that words are assigned an asymmetric resolved moraic trochee whose head must contain at least two morae (see Dresher and Lahiri 1991; Lahiri and Dresher 1999; Lahiri et al. 1999). We illustrate this trochee in the following examples, where the head of the foot is marked in square brackets. (4) Old English resolved moraic trochee (x ) (x .) ([m m]) ([mm] m) ho fu wor da ‘dwelling nom.pl’ ‘word gen.pl’
(x .) ([m mm] m) fæ rel de ‘journey dat.sg’
(x .) ([m m] m) æ e le ‘noble nom.sg’
We now proceed to talk in turn about each of the phonological alternations mentioned above. The first is syllabification, which is responsible for vocalizing the glide /j/ between consonants and deleting it before (or merging it with) /i/.
The Dental Preterites in the History of English
511
(5) SYLLABIFICATION A parsing procedure (rather than a phonological process) that applies continuously at each level. Requires all segments to be syllabified and remedied according to language-specific constraints. Old English glides are vocalized when consonants follow, and a /ji/ sequence is modified to /i/. Unsyllabifiable final glides can remain until the word level, when they must also be vocalized. Onset maximization based on the sonority hierarchy is preferred. Syllabification bleeds gemination, in which su‰xal /j/ assimilates to the preceding root-final consonant, producing a geminate. (6) GEMINATION (GEM) C i þ/j/ ! C i C i (C i 0 [r]); blocked if the result is a trimoraic head. Gemination fails in two circumstances: when the root-final consonant is /r/ or when a trimoraic head would result. In these situations the /j/ is retained. Examples illustrating gemination are given in (7) for both verbs and nouns. The nouns belong to the ja-class and thus /j/ comes between the stem and the case and number su‰x. (7) Constraining gemination in Old English a. Gemination permitted2 ‘perform 1p.sg’ ‘desert dat.sg’ fremme /frem-j-e/ we¯stenne /we¯sten-j-e/ (x ) (x .) (x .) (x ) (x .) ([m m]) ([mm] m) ([mm] m) m ([mm]) ([mm] m) fre mje > frem me we¯ ste nje > we¯ sten ne b. Gemination blocked to prevent trimoraic foot head ‘moan 1p.sg’ ‘noble dat.sg’ mæ¯ne /mæ¯n-j-e/ æele /æel-j-e/ (x .) (x .) ([mm] m) *([mmm] m) ([m m] m) *([m mm] m) mæ¯ nje > *mæ¯n ne æ e lje > *æ el le Gemination freely applies to /frem-j-e/ and /we¯sten-j-e/ since the resulting long consonant does not make the head of the foot trimoraic. In /we¯sten-j-e/ ! we¯stenne, the foot structure has changed, but the head of the initial foot remains bimoraic. On the other hand, after gemination /mæ¯n-j-e/ would become mæ¯n.ne, with a trimoraic head *mæ¯n, and hence gemination is blocked. The same holds true for æele. Glides to which gemination does not apply are subject to further alternations, depending on what su‰xes follow; when final, /j/ is treated as an appendix and left unsyllabified until the word level, where it gets vocalized. Examples will be provided below.
512
Aditi Lahiri
In addition to syllabification and gemination there are three vowel/glide-related phenomena: deletion, lowering, and umlaut. (8) Vowel/Glide-related phenomena a. HIGH VOWEL DELETION (HVD) High vowels are deleted in the weak branch of a foot. b. LOWERING (LOW) Unstressed high vowels are lowered. c. UMLAUT (UML) Vowels are fronted before /i/ and /j/ (later shown to be level I). In addition, degemination applies as a filter when geminates happen to occur before a consonant or after a long vowel. (9) DEGEMINATION (DEGEM) Geminates in a closed syllable when preceded by a long vowel or followed by another consonant are degeminated. Word-final geminates are also degeminated.3 These alternations are illustrated by the examples in (10), in which syllabification (along with foot structure), high vowel deletion, lowering, and degemination are shown applying to several verbs. Umlaut is assumed to have already applied. (10) a. mæ¯nde ‘moan 3p.past.ind’ (x .) /mæ¯n-j-d-e/ > mæ¯ ni de > mæ¯n de syll hvd b. fremede ‘perform 3p.past.ind’ (x .) /frem-j-d-e/ > fre mi de > fre me de syll low c. ce¯pe ‘keep 3p.sg.pres’ (x) (x) /ce¯p-j-i/ > ce¯ pi > ce¯ pe syll low d. fylde ‘fill 3p.past.ind’ (x .) /fyll-j-de/ > fyl li de > fyll de > fyl de syll hvd degem e. mæ¯ne ‘moan 1p.sg.pres’ (x .) /mæ¯n-j-e/ > mæ¯ ni e > mæ¯n e syll hvd
The Dental Preterites in the History of English
513
Given the existence of degemination, one might argue that gemination should apply everywhere without constraint and that the degemination sorts things out later. That is, mæ¯ne ‘mean’ (1p.sg.pres) would go through the following steps: /mæ¯nje/ > mæ¯nne (gem) > mæ¯ne (degem). However, evidence from ja-nouns proves that this cannot be the case. For ja-nouns the nominative singular su‰x is null, with surface forms like wı¯te ‘punishment’ coming from /wı¯t-j-Ø/, or cynn ‘race’ from /cyn-j-Ø/. The final e in wı¯te comes from the vocalized /j/, which is retained since it is word final and remains as an appendix till the very end. Allowing gemination to apply freely and absorb the /j/ would incorrectly predict degemination to delete all trace of /j/, yielding derivations like /wı¯t-j-Ø/ > wı¯tt (gem) > *wı¯t (degem). The correct analysis is /wı¯t-j-Ø/ > wı¯t( j) (gem blocked, /j/ retained at word edge) > wı¯ti (syll) > wı¯te (low). The deletion of /j/ in the context of a high front vowel (glide deletion), which accounts for derivations like /frem-j-i/ 3p.sg.ind > fremi (*/ji/ > i) > freme (low), bleeds gemination: /frem-j-i/ > *fremmi > *fremme. 21.3
The Dental Preterite
We now demonstrate how the various phonological phenomena discussed thus far a¤ect the realization of the dental preterite. Until now we have operated as though all su‰xes in Old English—derivational /j/, past tense /d/, person/number su‰x— are added to the root simultaneously. It will be assumed in this section, however, that derivational and inflectional su‰xes are added at di¤erent, ordered levels (levels I and II) and that the dental preterite is added at level I, along with derivational suffixes, rather than at level II, with the inflectional su‰xes. A similar position is taken, for Old High German and Old Norse, in Lahiri (2000b); Kiparsky and O’Neill (1976) also treat /d/ di¤erently from person/number su‰xes, though this does not play a crucial role in their derivations. The evidence for the separation into di¤erent su‰xational levels comes from the fact that various rules like umlaut operate on an independent domain that includes the stem extensions /j/, /i/, and /d/ but not the true inflectional su‰xes. Dresher (1993, 331–333) makes this argument convincingly for the Old English dialect Mercian on the basis of rules like umlaut and breaking. Compare the 3p.sg.pres forms of the strong verb haldan ‘hold’ and the derived weak verb onhældan ‘lean down’ below, based on a derivation from Dresher 1993 (p. 331). The weak verb undergoes umlaut, triggered by the derivational su‰x -j at level I; the rule does not apply at level II, where the /i/ is added as part of the agreement su‰x in /hæld-i/. (11) Level I uml Level II syll þ low
Weak verb /on-hald-j/ onhældj onhældj-i onhælde
Strong verb /hald/ — hald-i halde
514
Aditi Lahiri
The following derivations illustrate the dental preterite being added to the root, along with /j/, at level I, independent of the person/number or case/number su‰xes, which are added at level II. The phonological rules are as discussed above, with the addition of a rule of voicing assimilation, which devoices the dental preterite following a voicelesss root-final consonant at level II. Parentheses indicate unsyllabified material, which is resolved (when necessary) at the word level. (12) Weak verbs: 1p.sg.pres (root-j-e) and 3p.past (root-j-d-e) Level I input syll uml
present framþj fram( j) frem( j)
narþj nar( j) ner( j)
fullþj full( j) fyll( j)
framþjþd narþjþd co¯pþjþd fullþjþd fra.mid na.rid co¯.pid ful.lid fre.mid ne.rid ce¯.pid fyl.lid
1p.sg
Level II input syll gem hvd degem voi.ass low
co¯pþj co¯p( j) ce¯p( j)
past
3p.past
fremje —— frem.me —— —— —— ——
nerjþe ne.rje —— —— —— —— ——
ce¯pjþe ce¯.pi.e —— ce¯pe —— —— ——
fylljþe fyl.li.e —— fylle —— —— ——
fremidþe —— —— —— —— —— fremede
neridþe —— —— —— —— —— nerede
ce¯pidþe —— —— ce¯pde —— ce¯pte ——
fyllidþe —— —— fyllde fylde —— ——
fremme
nerie
ce¯pe
fylle
fremede
nerede
ce¯pte
fylde
(13) Weak verbs: 3p.sg.ind (root-j-i) and past participle (root-j-d) Level I input syll uml
present framþj fram( j) frem( j)
fullþj full( j) fyll( j)
framþjþd framid fremid
3p.sg
Level II input syll gem hvd degem voi.ass low
co¯pþj co¯p( j) ce¯p( j)
past
fremjþi fre.mi —— —— —— —— freme
ce¯pjþi ce¯.pi —— —— —— —— ce¯pe
co¯pþjþd co¯pid ce¯pid
fullþjþd fullid fyllid
participle fylljþi fyl.li —— —— —— —— fylle
fremid —— —— —— —— —— fremed
ce¯pid —— —— —— —— —— ce¯ped
fyllid —— —— —— —— —— fylled
The nominal forms in (14) provide additional support for the formulation of the different processes.
The Dental Preterites in the History of English
515
(14) ja-nouns: nom.sg (root-j-Ø) and nom.pl (root-j-u) Level I input syll uml
nouns cunþj cun( j) cyn( j)
wı¯tþj wı¯t( j) ——
Level II input syll gem hvd
we¯stenþj we¯sten( j) ——
aelþj ael( j) æel( j)
nom.sg
nom.pl.neut
cynjþØ —— cynn ——
wı¯tjþØ —— —— ——
we¯stenjþØ —— we¯stenn ——
æeljþØ —— —— ——
cynjþu —— cynnu cynn
wı¯tjþu wı¯tiu —— wı¯tu
we¯stenjþu æeljþu —— æeliu we¯stennu —— we¯stenn æelu
cyn —— ——
—— wı¯ti wı¯te
we¯sten —— ——
—— æeli æele
cyn —— ——
—— —— ——
we¯.sten —— ——
cyn
wı¯te
we¯sten
æele
cyn
wı¯tu
we¯sten
Word level degem syll low
—— —— —— æelu
Note that in both levels I and II, the fate of final, unsyllabifiable material is left up to the phonology of the subsequent level. For example, final unsyllabified glides not absorbed by gemination in level I remain unsyllabified (as in /fyll( j)/ in (12) and /wı¯tj/ in (13)), pending potential later su‰xation at level II. Glides that are still final, and unsyllabified, in level II are vocalized (and lowered) at the word level, when no further su‰xation is possible (e.g., wı¯te, in (13)). Degemination operates in a similar manner. The numerous phonological alternations we have discussed result in many instances of ambiguity or opacity. For example, the nom.sg wı¯tu (< /wı¯t-j-u/) is opaque with respect to high vowel deletion, which deletes the vocalized glide /i/ but not the plural su‰x /u/. Gemination produces a situation of near ambiguity between the present tense forms and the infinitive forms of light roots, on the one hand, and heavy roots with underlying long consonants, on the other. Except in in 2/3p.sg (cf. cysse versus teme), these two root types behave indistinguishably; cf. cyssan ‘to kiss’, stillan ‘to still’, py¤an ‘to pu¤ ’, fyllan ‘to fill’, which have underlyingly geminate consonants, versus temman ‘to tame’, clynnan ‘to sound’, dynnan ‘to make a noise’, whose geminates are derived. It is often only the comparision between present and past tense forms of a verb that reveals whether the root is underlyingly heavy or not. For instance, the language learner can figure out that fylde must have an underlyingly heavy root, since otherwise the glide /j/ would surface as a low vowel, yielding *fylede like temede. Note particularly the di¤erence between the past participial forms fremed, ce¯ped, fylled (13) and the 1p.sg.past.ind forms fremede, ce¯pte, fylde (12). Because of the
516
Aditi Lahiri
di¤erences in foot structure, hvd applies in fylde (< /fylide/) and ce¯pte (< /ce¯pide/) but not in fylled (< /fyllid/) or ce¯ped (< /ce¯pid/); only in the former set of forms is the high vowel in the weak branch of a foot. As a result, fylled maintains its underlying geminate and ce¯ped has no opportunity to undergo voicing assimilation. These di¤erences between the past indicative forms and the past participle play an important role in the Middle English developments. 21.4
Old to Middle English
In this section we address the development in Middle English in which the dental preterite eventually splits into two su‰xes, -t and - d, the first having a closer a‰nity to the root and the second behaving like a regular inflectional ending, as in the other Germanic languages. We argue that these changes are motivated by the addition of several phonological rules in Middle English. The following observations characterize the changes in the behavior of heavy roots. e
a. Vowel quantity alternations appear in Late Middle English, with some past tense forms (cepte, felde, etc.) exhibiting short vowels. This process of vowel shortening counteracts a regular process of vowel lengthening before voiced homorganic clusters, particularly -ld, -nd, and -mb, which was introduced in Late Old English, as in we¯lden ‘to wield’ and fe¯ld ‘field’. Long consonants as in fillen were real geminates. b. The preterite is manifested as [t] after sonorants, starting in Late Middle English, in just those verbs with the new short vowel in the past, such as felte. (After medial [ ], as in he¯led, the preterite consonant remains voiced; that is, *[ t] was never attested.) e
e
Table (15) illustrates the changed situation for verbs with heavy roots in Early and Late Middle English.4 (15) Class I alternations (original heavy roots) infinitive
past
OE
ME
OE and Early ME
Late ME
ce¯pan fe¯dan fe¯lan me¯tan fyllan de¯man hæ¯lan
ce¯pe(n) fe¯de(n) fe¯le(n) me¯te(n) fille(n) de¯me(n) hæ¯le(n)
ce¯pte fe¯dde fe¯lde me¯tte filde de¯mde hæ¯lde/he¯lde
cepte fedde felde (later felte5) mette filde dempte (later deemed) he¯led
gloss
‘keep’ ‘feed’ ‘feel’ ‘meet’ ‘fill’ ‘deem’ ‘heal’
The Dental Preterites in the History of English
517
A similar argument can be made on the basis of the behavior of Middle English verbs with light roots. Examples are shown in (16) (forms marked with z are hypothesized). (16) ME alternations (original light roots) OE infinitive fremman werian styrian dynnan trymman
Early ME past fremede werede styrede dynede trymede
infinitive fremmen/fremen werien stirien din(n)en ‡ trymmen
Late ME past fremede werede stirede dinede/dı¯ned ‡ trymede
infinitive fre¯me we¯re stire dinne trimme
‡
past fre¯med we¯red stired dinned trymmed
gloss ‘perform’ ‘defend’ ‘stir’ ‘resound’ ‘strengthen, trim’
The new process of open syllable lengthening is evident in these data, applying in the past tense forms of verbs with light roots (see also Lahiri and Dresher 1999; Lahiri and Fikkert 1999; Fikkert et al. 2006), for example, OE fremede > ME fre¯med. The lengthy literature on the split in behavior in the dental preterite, illustrated above, contains several possible explanations. For example, it has been suggested that the use of [t] in the preterite is an extension of the OE 3p.sg (thus sende > sent) (Morsbach 1896) or was borrowed from verbs like ME cepte (Moore and Marckwardt 1951). What remains unexplained, however, is why the [t] was extended only to the class of heavy roots, even when the roots ended in a sonorant (cf. ME felte > felte) class. The loss of the final schwa in the original trisyllabic past tenses like fremede > fre¯med is also unclear, as is the absence of intermediate forms like *fre¯mede. Brunner (1960), taking a pessimistic attitude toward these explanations, simply says that the reasons for these changes are not quite clear. It is claimed here that the split in the dental preterite follows from the phonological changes that occurred in Middle English. Below is a list of the principal components that led to the split of preterite /d/ into two separate preterite morphemes, /t/ and /d/, the first remaining a level I su‰x, and the second becoming a level II su‰x, resulting in present-day alternations like feel – felt versus deem – deemed. a. The innovation of open syllable lengthening produces a long initial stem vowel in original light syllables (e.g., we¯red ). b. Trisyllabic shortening (Lahiri and Fikkert 1999), which shortens antepenultimate stressed long vowels, ensures the absence of a medial stage like *fre¯mede; additionally, there was a later tendency for trisyllabic words with three light syllables to become disyllabic: fremede > fre¯med, stirede > stired.
518
Aditi Lahiri
c. Because the OE rule of high vowel deletion had never applied to light roots, such roots exhibited a tendency to end in -ed rather than -de, thus stired < OE stirede rather than *stirde, but felde < OE fe:lde /fe:lide/. d. Original heavy roots ending in sonorants went in two directions; some began to exhibit vowel alternations in the present and past. Verbs like fe¯lan @ fe¯lde > felde (‘feel’ @ ‘felt’) exhibited vowel shortening in the past, but not with hæ¯lan @ hæ¯lde > he¯led (‘heal’ @ ‘healed’). When the long vowel was maintained in the past, the ending was -ed. e. The vowel shortening in the past tense is often attributed to a special case of Closed-Syllable Shortening (see, e.g., Morsbach 1896; Moore and Marckwardt 1951; Jespersen 1961; and many others), although vowels before other homorganic consonant clusters like ld, nd, mb, and so on, were usually lengthened, as in we¯lden ‘to wield’ and fe¯ldes ‘field-gen’. f. Heavy roots ending in voiceless stops also underwent vowel shortening, and maintained the surface su‰x -te: ce¯pte > cepte, but ce¯pan. g. The voiceless su‰x is later used also for the past of sonorant roots but only when the root vowel is shortened: fe¯lde > felde > felte > felt, but hæ¯lde > he¯led > healed. h. At no point was there a possibility for the su‰x to be *-et, regardless of the weight of the root. One important factor is the opacity the interacting processes introduced into the system. Open syllable lengthening applied to light roots, obscuring the original weight distinction; also as a result of open syllable lengthening, gemination, which in Old English had served to distinguish light and heavy roots, no longer applied transparently, and therefore even more evidence for the distinction between light and heavy roots was lost. This had consequences for the analysis of the preterite. While in Old English the surface distribution of the three allomorphs of /d/ ([t], [d], [ d]) was entirely predictable from root weight and verb class, in Middle English the conditions in which they appeared became more arbitrary. Dialectal di¤erences and changes in verb class contributed further to this arbitrariness, setting the stage for a reanalysis of the preterite allomorphy. Middle English did still exhibit three consistent di¤erences between the light and heavy roots. In the original light roots (cf. (16) and (17c)), regardless of whether the root vowel became long in an open syllable in the past tense ( fre¯med ) or whether the geminate consonant was retained (dinned ), there was no syncope of the medial unstressed vowel e (from original OE /j/ > [i] before /d/) even in Late Middle English. Thus, regardless of original weight distinctions becoming opaque and regardless of dialectal di¤erences, original Old English short roots always had a full syllabic -ed in the past tense, a situation that continued for quite a while in Middle English, as e
The Dental Preterites in the History of English
519
in stired (see Moore and Marckwardt 1951; Minkova 1991; Lass 1992), even when the root vowel was lengthened. By contrast, verbs formed from heavy roots exhibited past tense forms both without syncope (he¯led ), as well as with syncope when ending in -de ( fedde) or in -te (cepte). There was also a di¤erence in the final vowel, which was lost in verbs with light roots (see (16)) that were originally trisyllabic in the preterite (stirede ! stired ) but, even in Late Middle English, was preserved in verbs with heavy roots (e.g., cepte; see (15)). A third crucial di¤erence between the original light and heavy roots in Middle English involves the preterite /d/. While heavy roots can show up with a final /t/, the original light roots never do, a situation that has been maintained in Modern English, where one finds past tenses of original light root verbs like stirred and tamed, but none with a final [t]. Our claim is that the opacity introduced by the phonological changes detailed above caused the dental preterite to be ambiguous in its pattern of alternations. The language learner was faced with competing hypotheses, and for the reasons detailed below, this led to a split such that for some verbs like feel, the preterite was still used as an indicator of verb class, functioning therefore more like a derivational level I su‰x, while for others it was treated as a level II inflectional ending. The following scenario sketches the history of the reanalysis. The first step occurred when the verb-deriving su‰x /j/ became opaque and was no longer productive. In OE, which had alternations like fremman/fremede versus fyllan/fylde, along with the short roots in /r/, which did not geminate as in werie(n)/werede, the su‰x was entirely transparent. However, when open syllable lengthening became active in Middle English, obscuring the di¤erence between light and heavy roots, and the final schwa in trisyllabic words was lost, the language learner was confronted with alternations like we¯re(n)/we¯red versus de¯me(n)/de¯mde and dinne(n)/dı¯ned/dinede versus fylle(n)/fylde. Further, words like OE temman, which had been reanalyzed as another class of weak verbs without gemination, now inflected as te¯me(n)/te¯med. The point is that gemination was no longer synchronically predictable. The main reason for the morphological split was the conflicting alternations confronting the language learner as a result of the phonological behavior of the verbs with original light roots. Some had become long by open syllable lengthening (e.g., we¯re(n)/we¯red ), while others geminated in the present and alternated in length in the past (e.g., dinne(n)/dı¯ned/dinede). The interaction between open syllable lengthening and trisyllabic shortening meant that past tenses like *we¯rede and *dı¯nede would not occur. Either the vowel was lengthened and the final vowel was deleted or the trisyllabic form remained with an initial short vowel. Gradually the trisyllabic forms became less preferred.
520
Aditi Lahiri
(17) Alternations confronting the language learner in Late ME root
past
infinitive
long v long v long v short v gem gem gem
ho¯ped felde cepte stired dinede/dı¯ned fylde wedded
ho¯pe fe¯le ce¯pe stire dinne fylle wedde
Speakers analyzing present-past alternations like dinne @ dinede/dı¯ned were faced with a choice of postulating a geminated root where the consonant in the past was also long (like trimme @ trimmed ), or a root that undergoes open syllable lengthening in both the past and the infinitive (ho¯pe @ ho¯ped ). Given these alternations, no consistent phonological analysis was possible assuming a single preterite su‰x. However, a systematic pattern of alternations was still possible by assuming two distinct preterite su‰xes: -ed and -d, which surfaces as -de or -te with inflection. (18) Reorganized pattern of alternations based on the preterite su‰x rootþsuffix alternation
root
past
infinitive
-d(e)/-t(e) (level I)
˘ C-de V ˘ C-te V ˘ C-de V
long v long v gem
felde cepte fylde
fe¯le ce¯pe fylle
-ed (level II)
¯ C-ed V ˘ CC-ed V ˘ C-ed V ¯ VC-ed
long v gem short v gem
ho¯ped wedded stired dinede/dı¯ned
ho¯pe wedde stire dinne
suffix
The bifurcation of the dental preterite meant that the /d/ su‰x on level I was now added directly to the root (there being no longer an intervening derivational /j/). Level I preterites were subject to root-level constraints like *VVCC, a ban on VVCC roots already applicable in Old English (and a factor in blocking gemination). Now extended to the root plus preterite /d/, *VVCC led to vowel shortening in the past tense.6 In contrast, the inflectional su‰x /- d/ was added in level II and was not subject to any such constraint. The situation after the morphological split was as follows. e
The Dental Preterites in the History of English
521
(19) Preterite in Late Middle English present
*VVCC Level II degem v-assim Mod.E
——
past /d/ fe¯lþd ce¯pþd feld cepd
fylldþd ——
fyllþe —— —— fylle
1p.sg / d/ de¯mþ d —— —— de¯med
1p.sg /e/ feldþe —— —— felde
cepdþe —— cepte cepte
fylldþe fylde —— fylde
‘fill’
‘deemed’
‘felt’
‘kept’
‘filled’
present de¯m fe¯l —— ——
ce¯p ——
fyll ——
1p.sg /e/ de¯mþe —— —— de¯me
fe¯lþe —— —— fe¯le
ce¯pþe —— —— ce¯pe
‘deem’
‘feel’
‘keep’
e e
Level I
past
It is level I /d/ that has come down to Modern English as the ‘‘special past’’ marker /t/ as found in felt. Thus, at the next stage of the language, after there had already been a split between the two past tense markings, the level I su‰x became underlying /t/. Development of voiceless /t/ depended on the loss of unstressed schwas in the 1p.sg, resulting in forms like cept, lovd. (Schwas were retained only after coronal stops, as in hunted.) The loss of unstressed schwas also led to assimilation in level II, producing forms like ho¯pt. Moore and Marckwardt (1951) observe that the /t/ was generalized based on words like cepte; the question, however, is why. The explanation o¤ered here is that it had become necessary to unambiguously mark the di¤erence between the two past tenses, whose distribution was no longer transparent. By the time of the loss of the schwa, the original light verb roots had either been restructured as having an underlyingly long vowel (e.g., ho¯p-) or an underlying geminate (dinn-), which later degeminated. The past su‰x assimilated to the root with the loss of schwa and forms like ho¯pt surfaced from ho¯ped, clashing with forms like cept. That is, the surface forms did not reflect the di¤erence between ho¯p(e) @ ho¯ped > ho¯pt and ce¯p @ cept. One way of distinguishing the su‰xes was to keep the underlying forms apart (i.e., not just allomorphs)—one su‰x as /d/ and the other as /t/, introducing a distinction between -ed and -t with root-final sonorants, as in felt versus healed. Another important reason for /t/ to be the special plural morpheme was that the only source for preterite /t/ was assimilation to root-final voiceless consonants in heavy roots that could later be shortened, as in cept. Thus, /t/ was always associated with vowel shortening and could be easily interpreted as indicating the special status of the level I su‰x. Our analysis of the splitting of the dental preterite into two su‰xes and the eventual root-internal shortening in the context of the level I /d/ has support from several sources. First, vowels that were lengthened in Late Old English before voiced homorganic clusters (see above) did not shorten with level II inflections: ME fe¯ld, fe¯ldes
522
Aditi Lahiri
‘field’; gru¯nd, gru¯ndes ‘ground’; or in disyllabic adjectives like wı¯lde ‘wild’.7 Neither did shortening apply in verb roots that ended in these clusters: ho¯ld, f ¯ınd, and so on. Original Old English long vowels in this context retained their length as in OE fe¯onda ‘fiend’, but again not in certain weak verbs such as OE mæ¯nde ‘meant’. Thus, the past tenses with shortened vowels were created by level I /d/, which patterns with other level I su‰xes that cause shortening such as wide @ width, or other level I constraints like trisyllabic shortening as in child @ children, wild @ wilderness (see Lahiri and Fikkert 1999). Second, there was no shortening in the past participle, which was simply the earlier /d/ later restructured to / d/. The surface forms never changed: fe¯led, dined, de¯med, ce¯ped (see (13)). According to our analysis, in Old English, the past participle forms would be identical to the past, except that high vowel deletion would not apply since the high vowel was in a closed syllable rather than in a weak branch of a foot: /de¯m-j-d/ > de¯mid (no hvd) > de¯med (lowering) (cf. past de¯mde). This is also why voicing assimilation did not apply (ce¯ped; cf. ce¯pte). In Middle English, when the past participle generally was restructured to / d/, it caused neither shortening nor voicing assimilation. Only later when the past and past participle merged did the root vowel shorten. A third argument supporting /d/ as a separate su‰x from the person/number endings comes from other inflected forms. Neither the consonantal second or third person singular su‰xes st and had this e¤ect, as seen below. e
e
(20) Lack of vowel shortening with other inflectional endings8 OE 3p.sg ce¯p(e) > ME ce¯p, ce¯p o cf. Mod. E ‘keeps’ 2p.sg ce¯p(e)st > ME ce¯pst, ce¯p st OE 3p.sg.fe¯l(e) > ME fe¯l, fe¯l o cf. Mod. E ‘feels’ 2p.sg.fe¯l(e)st > ME fe¯lst, fe¯l st e e
e e
Note that in some dialects (e.g., the Midland dialects and, especially, the Northern dialects), the second and third person singular endings were preceded by schwa ( ) (Wright and Wright 1928, sec. 150), which could be taken as the reason that these endings behaved di¤erently from the preterite /d/. However, not all dialects show schwa. In the Southern dialects, the ending was usually []. Even in the Midland dialects, presence or absence of schwa was conditioned. After heavy roots, the absence of schwa was more frequent in the Midland dialects than after light roots. Thus, in both the Southern and the Midland dialects, fe¯l would have been a more frequent form than fe¯l . In any event, there was no vowel shortening in the third or second person singular even in dialects where the ending clearly lacked [ ]. That is, in the Southern dialects, alternations like fe¯le(n) infinitive, fe¯l 3p.sg, felde past (> felte) occurred regularly. Thus it can be said with certainty that a segmentally similar consonantal inflectional ending did not have the same e¤ect as the preterite. e
e
e
The Dental Preterites in the History of English
523
The splitting up of the preterite had the e¤ect that many verbs changed classes in the course of the history of English, usually from being originally level I to becoming level II. Curme (1935) gives a detailed list of the various past and participle forms of weak verbs at di¤erent stages. But neither the shift nor the initial division was random. What is interesting is that the general tendency was for long roots that took the level I su‰x to take the level II su‰x or in other words to become ‘‘regular.’’ According to our analysis, the split was generally based on the original weight of the root and syncope of the vowel. The crucial contrasts were de¯mde versus we¯red. The original light verbs with gemination ( fremmen) or a long initial vowel (we¯ren) always took the past -ed, and when the split took place, these verbs were all treated as level II and therefore ought to have been treated at a later period as ‘‘regular.’’ This was indeed the case. Finally, since all verbs that now take the level I su‰x /t/ originally came from the class I weak verbs, they were always umlauted. Consequently, synchronically English has only front vowels in roots that take the /t/ su‰x. 21.5
Conclusions
In sum, there is a host of evidence suggesting that the Old English dental preterite split into two di¤erent inflectional su‰xes. Our claim is that this split could not have occurred had it not been the case that the dental preterite marker was a su‰x on par with derivational su‰xes and stem extensions, marking a particular class of verbs. The chronology of this evolution of the weak past is summarized below. [{root þ j þ d} þ infl sufx] [{root þ d} þ infl sufx] or [{root} þ { d þ infl sufx}] [{root} þ t] or [{root} þ d] e
(21) Evolution of the weak past Old English Early Middle English (i) (ii) Late Middle English (i) (ii)
Grammaticalization of ‘do’ from the Proto-Germanic to English went through a stage of compounding to cliticization, after which the clitic was reinterpreted as a su‰x. (On these earlier stages of development from ‘do’, see Lahiri 2000b and references therein.) In English, this /d/ has again been reinterpreted and split up into two morphemes. The two ways of making a past tense can be traced to this split. 21.6
Implications
Under the classical view of language change, there are two possibilities: sound change and analogy. In the strictest sense, the reanalysis of /d/ from a class marker
524
Aditi Lahiri
to two inflectional su‰xes is neither a sound change nor really an analogical change. If the latter, it would be a severe complication in the grammar, counter to Kiparsky’s views of analogy (e.g., Kiparsky 1982a). In fact, however, the split in the dental preterite is best interpreted as grammar simplification. English, along with other Germanic languages, maintained a clear distinction between heavy and light roots that was relevant to the realization of the dental preterite. Phonological changes destroyed this clear distinction, a situation that was repaired by splitting the preterite in two. The two morphemes were chosen from existing morphophonemic variants, each being assigned to a separate level. The analysis presented here also has consequences for accounts that depend on paradigm levelling, again a topic extensively addressed by Kiparsky (e.g., 1971b, 2000a). When independent phonological rules led to irregular paradigms in the Old English preterite, rather than leveling the paradigms, the original /d/ su‰x in question was split into two in an attempt to maintain old quantity distinctions. It was reanalyzed into a level I /d/ (later /t/) and a level II / d/ (later /d/). Particularly significant is that a level II su‰x was able to arise from a level I su‰x in this process, the opposite of the more familiar situation in which a level II su‰x becomes a level I su‰x over time. This phonological analysis of Old English is considerably influenced by Paul Kiparsky’s earlier work (particularly Kiparsky and O’Neill 1976). Although details di¤er and some controversial issues remain, the fundamental assumptions concerning Old English remain the same: the di¤erence between the ja- and i-nouns, the relationship between gemination, high vowel deletion, and metrical structure (‘‘strong marking’’ in Kiparsky and O’Neill 1976). Any historical phonological research, including that on Old English and especially that relating to metrical structure, owes an extraordinary debt to Kiparsky’s work. This research is no exception. e
Notes The author is much indebted to the invaluable comments from the editors, Kristin Hanson and Sharon Inkelas, and is also grateful to Elan Dresher, Jennifer Fitzpatrick, the late Mirco Ghini, Astrid Kraehenmann, Frans Plank, and Henning Reetz for their detailed comments and support. The research was funded in part by funds from the SFB 471, the Max-PlanckForschungspreis, and the Leibniz Preis. An earlier version of this paper was presented at the ICEHL 1998 in Manchester. 1. For reasons of space, we ignore alternations like tell – told. They do not contradict the central argument but need further explanation regarding the vowel alternations. 2. The underlying forms include the umlauted vowels here only for convenience. 3. Sometimes the original double consonants are still reflected in the orthography. 4. The Middle English data are put together based on the Oxford English Dictionary, Brunner (1960), Curme (1935), Moore and Marckwardt (1951), Sievers and Cook (1903), and Wright
The Dental Preterites in the History of English
525
and Wright (1908, 1928). Moore and Markwardt as well as Curme discuss in detail some of the early and late forms. 5. Moore and Marckwardt 1951, 162–163. 6. This constraint also a¤ected other level I su‰xes like /y/: heal – health. 7. There were dialectal di¤erences, but overall, before /ld/ the lengthening was preserved everywhere. English always had a tendency to lengthen vowels before /nC/ clusters; earlier it was more of a compensatory nature: cf. OE f ¯ıf, Gothic fimf. 8. In Northeast Midland, the third person ending was -es.
22
Analogical Morphophonology
Andrew Garrett and Juliette Blevins
It is the forge`d feature finds me; it is the rehearsal Of own, of abru´pt se´lf there so thrusts on, so throngs the ear. —Hopkins, ‘‘Henry Purcell’’ 22.1
Introduction
Since the neogrammarians first distinguished analogical change in morphological systems from what Ostho¤ and Brugmann (1878, xiv) called ‘‘mechanical’’ sound change, the study of word-based sound patterns has raised significant questions about these two types of change. Are sound changes indeed ‘‘mechanical,’’ or, as V. Kiparsky (1963, 7), Kiparsky (1965, 1967a, 1970, 1988a, 1995), and others have suggested, are they constrained by the linguistic systems in which they are embedded? Is diachrony solely responsible for the common similarity of word-based sound patterns and phonetic processes, or do phonetic principles actively constrain the form of words? A related question is how to define the diachronic relationship between what are sometimes called ‘‘phonological’’ (automatic exceptionless) patterns and ‘‘morphophonological’’ patterns (which may have morphological or lexical restrictions inter alia). The standard answer to this question is given by Kiparsky (1993a, 309): it is ‘‘a characteristic trajectory of phonological rules’’ for them to be ‘‘confined to the lexical phonology’’ over time. Dressler (1985, 149) writes in a similar vein that morphophonological processes evolve from phonological ones ‘‘by acquiring morphological and reducing phonological domains,’’ and Spencer (1991, 126) adds that they ‘‘reflect a stage in historical development of phonological rules which are becoming morphologized or lexicalized but which still retain a certain degree of generality.’’ In short, on this view, morphologically restricted sound patterns reflect formerly exceptionless patterns.1 Mechanisms of morphophonologization include morphological domain restriction, rule inversion (Vennemann 1972; McCarthy 1991; Blevins 1997), and telescoping (Wang 1968; Bach and Harms 1972; Anderson 1981; Blevins and Garrett 1993).
528
Andrew Garrett and Juliette Blevins
A typical example is the evolution of the English pattern of /g/ deletion in the adjectives long and strong (vs. longer and stronger).2 This can be analyzed as a process of coda /g/ deletion after /Ð/ (Borowsky 1986) if the morphology is organized so that resyllabification saves /g/ in some contexts (e.g., comparatives) but not others (e.g., agent nouns like singer). Since few forms alternate, the linguistic reality of this case can be challenged, but the point is precisely that the alternation is restricted in scope. Interestingly, there is evidence that the alternation was once general: coda /g/ was not realized after /Ð/, but onset /g/ was. In his Principles of the English Language (1765), the orthoepist James Elphinston wrote that /g/ is pronounced after /Ð/ when required ‘‘to articulate either a vowel or a liquid; which it does not only if the vowel or liquid follow in the same word, but even, upon solemn occasions, if either feebly commence the word following in immediate connexion and dependance’’ (Mu¨ller 1914, 215–216; cf. Horn 1901; Dobson 1968; Rohlfing 1984). That is, in conservative (‘‘solemn’’) speech, /g/ was retained after /Ð/ if followed in the same phrase (or word) by an unstressed (‘‘feeble’’) syllable beginning with a vowel or liquid; this can be treated as postlexical resyllabification. In colloquial speech styles, presumably, word-final /g/ was always deleted after /Ð/. Though Elphinston had lost wordinternal coda /g/ in cases like length and kingdom, he specifically cites all the examples in (1) as retaining their highlighted /g/. (1) a. angle, angling; anger, angry, finger, hunger, linger; longer, younger b. hanger, singer, slinger; hanging, singing, prolonging, bringing, belonging, longing, singing, springing, twanging; hanged, longed c. sing aloud, prolong it, strong and mighty, spring eternal, long repose, young Leander In most modern English dialects, /g/ is still pronounced as in (1a) but not (1b) or (1c).3 Over the last two centuries Kiparsky’s ‘‘characteristic trajectory’’ has been realized. Coda /g/ deletion after /Ð/ was a postlexical pattern in the conservative speech variety Elphinston described; it is now restricted to the lexical phonology. This diachronic scenario is common, but we will argue here that there is another, previously undescribed historical source of morphophonological patterns. The pathway we identify di¤ers significantly from morphological domain restriction, telescoping, and rule inversion. These processes yield patterns whose substance actually reflects one or more earlier sound changes (telescoped, inverted, or restricted in domain). They may result in phonetically unmotivated patterns, but phonetically based sound changes underlie them. The examples we will discuss show a di¤erent diachronic profile. In these cases, phonetically unnatural patterns arise not through a series of ordinary sound changes but by analogy, that is, morphological or paradigmatic generalization. Analogical extension of phonetically unnatural patterns is
Analogical Morphophonology
529
interesting because it shows that such patterns can be as productive and morphologically salient as any sound patterns in language. Extensions of this type, which has some points of contact with ‘‘phonological analogy’’ (Kiparsky 1982a, 58), result in what we will call ‘‘analogical morphophonology.’’ In sections 22.2–4 we discuss three cases, in ancient Greek, East Cushitic, and South-Central Dravidian, from di¤erent language families. The relevant historical developments are partly documented in the first case but must be reconstructed in the second and third cases. In each case we identify the crucial pattern and explain why we consider it unnatural, and we then analyze its evolution.4 22.2
Greek
Nasal-induced coronal spirantization corresponds to no observed sound change and has no articulatory or perceptual basis as far as we know.5 Nevertheless, in ancient Greek, a process by which coronal stops (/t/, /t h /, /d/) surface as s before /m/ is regular and thoroughly embedded in the verbal morphology.6 This is shown in (2) with two /m/-initial perfect middle verbal su‰xes, 1sg. /-mai/ and masculine nominative singular participle /-menos/. (2) Root a. poiestelderb. grap h plekc. peit h pseud-
‘make’ ‘send’ ‘flay’ ‘write’ ‘weave’ ‘persuade’ ‘deceive’
Perfect middle 1SG. pe-poı´e-mai e´-stal-mai de´-dar-mai ge´-gram-mai pe´-pleÐ-mai pe´-peis-mai e´-pseus-mai
Participle pe-poie-me´nos e-stal-me´nos de-dar-me´nos ge-gram-me´nos pe-pleÐ-me´nos pe-peis-me´nos e-pseus-me´nos
Before /m/, vowels and sonorant consonants surface intact as in (2a), noncoronal stops assimilate in nasality as in (2b), and coronal stops surface as s as in (2c). As shown in (3), the same pattern is found in nouns derived from verbal roots. (3)
Root a. peit h pret h pseudb. dathedknet h c. ot h od-
‘persuade’ ‘swell’’ ‘deceive’ ‘divide’ ‘sit’ ‘scratch’ ‘thrust’ ‘smell’
Derived noun peıˆs-ma preˆs-ma pseuˆs-ma das-mo´s hes-mo´s knes-mo´s os-me´ os-me´
‘persuasion’ ‘swelling’ ‘untruth’ ‘division of spoil’ ‘(a) swarm (of bees)’ ‘itching’ ‘(a) thrust’ ‘(a) smell’
530
Andrew Garrett and Juliette Blevins
The change in (2c) does not occur morpheme internally, as shown by words like atmo´s ‘steam’ and stat h mo´s ‘doorpost’.7 The origin of this pattern has long been known (Brugmann 1878, 81 note 1; Schmidt 1885). Its starting point was the finite paradigm of the perfect middle. The perfect as a whole is one of three basic aspectual types in the language (the others are the present and aorist); its regular formation is shown in (4) for the verb /stel-/ ‘send’, whose perfect stems are active /e-stal-k-/ and middle /e-stal-/. (4) a. Perfect active SING. 1. e´-stal-k-a 2. e´-stal-k-as 3. e´-stal-k-e b. Perfect middle SING. 1. e´-stal-mai 2. e´-stal-sai 3. e´-stal-tai
DUAL e-sta´l-k-aton e-sta´l-k-aton DUAL e´-stal-t h on e´-stal-t h on
PLUR. e-sta´l-k-amen e-sta´l-k-ate e-sta´l-k-asi PLUR. e-sta´l-met h a e´-stal-t h e
Note that the perfect active endings are vowel initial while the perfect middle endings are consonant initial. (A periphrasis replaces the third person plural form.) Verbs ending in vowels and liquids undergo no relevant changes in the perfect middle. As seen for /grap h -/ ‘write’ (perfect stem /ge´-grap h -/) in (2b) and (5), verbs in noncoronal stops do undergo assimilatory processes in the perfect middle. (5) a. Perfect active SING. 1. ge´-grap h -a 2. ge´-grap h -as 3. ge-grap h -e b. Perfect middle SING. 1. ge´-gram-mai 2. ge´-grap-sai 3. ge´-grap-tai
DUAL ge-gra´p h -aton ge-gra´p h -aton DUAL ge´-grap h -t h on ge´-grap h -t h on
PLUR. ge-gra´p h -amen ge-gra´p h -ate ge-gra´p h -asi PLUR. ge-gra´m-met h a ge´-grap h -t h e
In the perfect middle paradigm of verbs in noncoronal stops, the surface nasality and laryngeal features of the stem-final consonant are always predictable from the following consonant. In the first person singular and plural, the final stop assimilates in nasality to m; it is realized as p before /s/ or /t/, and as p h before the aspirate /t h /. These assimilations are fully regular in Greek. Another regular sound pattern is a spirantization process by which /t/ ! s before a coronal obstruent, as in /ı´d-te/ ! ´ıste ‘you (pl.) know’ and /anut-to´-s/ ! anusto´s
Analogical Morphophonology
531
‘practicable’. Showing this pattern in (6) are the reconstructed perfect middle forms of two verbs ending in coronal stops, /pseud-/ ‘deceive’ (perfect stem /e-pseud-/) and /peit h -/ ‘persuade’ (/pe-peı´t h -/).8 (6) a. Perfect middle: /pseud-/ ‘deceive’ SING. DUAL 1. *e´-pseud-mai 2. *e´-pseus-sai *e´-pseus-t h on 3. *e´-pseus-tai *e´-pseus-t h on h b. Perfect middle: /peit -/ ‘persuade’ SING. DUAL 1. *pe´-peit h -mai 2. *pe´-peis-sai *pe´-peis-t h on 3. *pe´-peis-tai *pe´-peis-t h on
PLUR. *e-pseu´d-met h a *e´-pseus-t h e
PLUR. *pe-peı´t h -met h a *pe´-peis-t h e
Note that the linguistic stage in (6) is reconstructed. Because the second and third person endings of the perfect middle all begin with coronal obstruents, spirantization alters an underlying stem-final coronal stop everywhere except in the first person singular and plural forms. Based on paradigms like those in (6), a simple analogical change occurred. The stem-final /s/ that appeared throughout the perfect middle of verbs in coronal stops was extended to the first person forms. This was an instance of paradigm leveling similar to many well-known changes, and its result was the new pattern in (7). (7) a. Perfect middle: /pseud-/ ‘deceive’ SING. DUAL 1. e´-pseus-mai 2. e´-pseus-sai e´-pseus-t h on 3. e´-pseus-tai e´-pseus-t h on b. Perfect middle: /peit h -/ ‘persuade’ SING. DUAL 1. pe´-peis-mai 2. pe´-peis-sai pe´-peis-t h on 3. pe´-peis-tai pe´-peis-t h on
PLUR. e-pseu´s-met h a e´-pseus-t h e
PLUR. pe-peı´s-met h a pe´-peis-t h e
Note that this is the actually attested paradigm, in which—due to an analogical extension from the second and third person forms—an underlying coronal stop now surfaces as s before m. In the earliest documented stages of Greek, only the analogical extension in (7) had taken place; the change of coronal stops to s before m in verbs (and forms derived from verbs) had not yet become general. Early sources instead show stop-m clusters, as in the archaic perfect middle participles in (8).
532
Andrew Garrett and Juliette Blevins
(8) a. ke-korut h -me´nos /korut h -/ b. ke-kad-me´na /kad-/ c. (pro-)pe-p h rad-me´na /p h rad-/ d. pe-puka´d-menon /pukad-/
‘armed’ (frequent in Homer) ‘arm, equip’ ‘surpassed’ (Pindar, Olympian 1.27) ‘surpass’ ‘told (beforehand)’ (Hesiod, Works & Days 655) ‘show’ ‘covered’ (Sappho 166) ‘cover’
Only in the classical (postarchaic) period did the pattern in (2c), with surface sm, become regular in the language. No doubt the extension was favored by the paradigmatic connection between (first person and participial) forms of the perfect middle; extension to /m/-initial su‰xes generally will then have followed the extension to participial forms. In any case, the point is that a case of ordinary paradigm leveling led to a fortuitous pattern that, though phonetically unmotivated, was nonetheless generalized to become quite regular. Analogy has created a phonetically unnatural morphophonological process.9 22.3
East Cushitic
We have argued elsewhere that metathesis (linear order inversion) of nasals and oral stops is unnatural phonetically; a survey of proposed cases shows that it is undocumented as a sound change in the languages of the world.10 This is because metathesis arises when a segment is perceived as being in a position other than its historical one. Such misperceptions are often faciliated by the temporally extended cues of segments like rhotics and pharyngeals. Yet in stop-nasal clusters, since nasality cannot leak across an adjacent oral stop without rendering it nasal, a TN cluster cannot be perceived as NT (nor can NT > TN). Synchronically productive nasal-obstruent metatheses are nevertheless documented in several East Cushitic languages. According to Bender (1976b, 3), the Cushitic branch of Afroasiatic contains four subbranches: North Cushitic (the Beja language), Central Cushitic (Agaw), South Cushitic, and East Cushitic. East Cushitic itself contains the three branches and various languages shown in (9), following Heine (1978) and Hayward (1984, 9). Nasal-obstruent metathesis alternations have been identified in the seven languages highlighted in boldface in (9): four Highland languages and three Lowland languages.11 Here it is impossible to present a full treatment of the evolution of each set of alternations. Instead, we will describe one system (that of Bayso) and propose a diachronic account of its origin. We will then sketch our reasons for believing that our account will also be applicable in the other East Cushitic cases.
Analogical Morphophonology
533
(9) East Cushitic Highland East Cushitic: Burji, Darasa, Hadiyya, Kambata, Sidamo Werizoid (Black 1976) Lowland East Cushitic Qafar, Saho Konso, Oromo Omo-Tana (a.k.a. Macro-Somali) Proto-Baz Western Omo-Tana: Arbore, Dasenech, Elmolo Northern Omo-Tana: Bayso Eastern Omo-Tana: Boni, Rendille, Somali The Bayso language has been described by Hayward (1978), and some aspects of its history have been explored by Heine (1978); its consonant inventory is shown in (10). (10) p b f m, m
t, t d s, s (¼ [ts]), z n, n, l, l, r, r
c (¼ [tS ]), c j (¼ [d‰]) sˇ (¼ [S ]) y, y
k, k g
w, w
h
The relevant alternations arise only in the verbal morphology, where personal endings begin with vowels, /t/, or /n/. This is shown in (11) for a representative verb ending in a labial. (11) Bayso simple perfect: /dub-/ ‘bake’ 1sg. /dub-e/ ! dube 1pl. 2sg. /dub-te/ ! dubte 2pl. 3sg.f. /dub-te/ ! dubte 3pl. 3sg.m. /dub-e/ ! dube
/dub-ne/ ! dubne /dub-ten/ ! dubten /dub-en/ ! duben
Labial roots like /dub-/ exhibit no relevant alternations with /n/-initial and /t/initial su‰xes, but such alternations do occur with coronal roots. These alternations are of several types. First, whenever a /t/-initial ending is added to a root in /sˇ/, the two segments surface as c ([tS ]). This ‘‘coalescence’’ is a sibilant metathesis (Blevins and Garrett 2004): /sˇt/ ! [tS ]. With roots ending in other coronal obstruents, the result is total progressive assimilation. These alternations are illustrated in (12) for several roots in coronal obstruents; the endings cited are the masculine and feminine third person singular endings /-e/ and /-te/. (12)
Simple perfect 3SG.M. a. /osˇ-e/ ! osˇe b. /doot-e/ ! doote /sˇigid-e/ ! sˇigide
Simple perfect 3SG.F. /osˇ-te/ ! oce (¼ [ot§e]) /doot-te/ ! dootte /sˇigid-te/ ! sˇigidde
‘dig’ ‘want’ ‘anoint’
534
Andrew Garrett and Juliette Blevins
/wut-e/ ! wute /fooc-e/ ! fooce /ajees-e/ ! ajeese
/wut -te/ ! wut t e /fooc -te/ ! fooc ce /ajees-te/ ! ajeesse
‘sow’ ‘chew’ ‘speak’
As shown in (13), roots of the same type exhibit obstruent-nasal metathesis when followed by /n/-initial endings. This metathesis is regular for underlying sequences of coronal obstruents plus /n/. It is this synchronic metathesis alternation that is the object of our analysis here. (13) Simple perfect 3SG.M. /doot-e/ ! doote /wod-e/ ! wode /wut-e/ ! wute /fooc-e/ ! fooce /osˇ-e/ ! osˇe /gilis-e/ ! gilise /kees-e/ ! keese
Simple perfect 1PL. /doot-ne/ ! doonte /wod-ne/ ! wonde /wut-ne/ ! wunte /fooc-ne/ ! foonce /osˇ-ne/ ! once /gilis-ne/ ! gilinse /kees-ne/ ! keense
‘want’ ‘drive’ ‘sow’ ‘chew’ ‘dig’ ‘swim’ ‘raise’
The assimilatory pattern in (12), with underlying sequences of coronal obstruent (other than /sˇ/) plus /t/, is progressive: /t/ regularly assimilates to the obstruent preceding it. A di¤erent pattern of assimilation is found in two other contexts in Bayso. The first context arises when any of the three consonant-final ‘‘auxiliary’’ su‰xes /-gir-/, /-ar-/, /-r-/ is followed by a consonant-initial verb ending. The result in such cases is an underlying /rt/ or /rn/ cluster. However, unlike /rt/ and /rn/ clusters in /r/-final verbs, which ordinarily surface without change (Hayward 1978, 554, note 25), these /rt/ and /rn/ clusters in ‘‘auxiliary’’ contexts surface with total regressive assimilation. For instance, the imperfect past is formed using the auxiliary /-ar-/, and representative forms are cited in (14). (14) Bayso imperfect past: /dub-/ ‘bake’ 1pl. /dub-in-ar-ne/ ! dubinanne 2pl. /dub-in-ar-ten/ ! dubinatten 3pl. /dub-in-ar-en/ ! dubinaren The second context where verbal endings trigger regressive assimilation is in the formation of ‘‘extended-stem verbs with a radical extension in -at, -sat, -aat, or -oot’’ (Hayward 1978, 558). The final /t/ of these extensions regularly assimilates to a following verbal ending in /n/, as illustrated in (15). (15) Bayso simple perfect: /kor-at-/ ‘climb’ 1pl. /kor-at-ne/ ! koranne´ 2pl. /kor-at-ten/ ! koratte´n 3pl. /kor-at-en/ ! korate´n
Analogical Morphophonology
535
It is crucial that the final /t/ of such extended-stem verbs (which are numerous and productively derived) di¤ers in its behavior from ordinary verb-final /t/ as in (12) and (13). Metathesis a¤ects /t/ in ordinary verbs, but /tn/ clusters undergo regressive assimilation in extended-stem verbs. Based on these facts, we propose the following account. We contend that two factors led to the creation of metathesis in the formation of unextended verbs with /n/initial su‰xes. Within the Bayso verbal system, both factors in e¤ect made it seem superficially plausible that forms with /n/-initial su‰xes should be derived by metathesis. The creation of synchronic metathesis was thus a straightforward generalization based on an apparent pattern elsewhere in the language. The first relevant factor was the regressive assimilation pattern seen in (14) and (15), and in particular the /tn/ ! nn assimilation found in first person plural forms of extended-stem verbs. What these patterns suggest is that the direction of assimilation in CC clusters is regressive, not progressive. The second relevant factor was the /sˇt/ ! [tS ] metathesis (with ‘‘coalescence’’) seen in (12a). At least for certain verb types, what this suggests is that metathesis is involved in the formation of paradigms such as the simple perfect. Now consider the task confronted (by a language learner or a potential language changer) in analyzing the forms in (12b). In these forms, diachronically, a su‰xinitial /t/ has assimilated to a stem-final coronal obstruent. But geminates derived by total assimilation are in principle ambiguous. They could reflect progressive or regressive assimilation, and the choice between these two analyses can only be based on other evidence in the language. Our proposal is that the geminates in (12b), though derived historically as shown by progressive assimilation, were reinterpreted synchronically as the result of regressive assimilation. In (16) we sketch this reinterpretation based on the examples in (12b). The reinterpretation generalized the metathesis pattern that independently arose in forms like the one in (12a), and it was further supported by the presence of regressive assimilation in the contexts in (14) and (15). (16)
Surface form a. oce (¼ [ot§e]) b. dootte sˇigidde wut t e fooc ce ajeesse
Originally /osˇ-te/ /doot-te/ /sˇigid-te/ /wut -te/ /fooc -te/ /ajees-te/
Reinterpreted as /otsˇe/ /dootte/ /sˇigitde/ /wutt e/ /footc e/ /ajeetse/
‘dig’ ‘want’ ‘anoint’ ‘sow’ ‘chew’ ‘speak’
The only actual change suggested so far is a restructuring of synchronic intermediate representations: a ‘‘covert reanalysis’’ (Kiparsky 1998) with no visible e¤ect. The
536
Andrew Garrett and Juliette Blevins
old structure /ajees-te/ and the new structure /ajeetse/ (synchronically the result of a metathesis from /ajees-/ þ /-te/) both yielded the same surface form ajeesse ‘speak’ (simple perfect 3sg.f.). The crucial change occurred when metathesis was extended throughout the verbal paradigm of verbs ending in coronal obstruents. Once verbs of this phonotactic class were seen as inverting the position of their final consonant and a su‰x-initial consonant, inversion was generalized for all su‰x-initial consonants. Metathesis was thus extended to the context where it is now visible, with /n/initial su‰xes as in (13). Just as /ajees-/ þ /-te/ now yielded a representation /ajeetse/ (surface ajeesse), so now /ajees-/ þ /-ne/ yielded /ajeense/. The latter underwent no assimilation, though, and surfaced intact as ajeense. We have explained the creation of nasal-obstruent metathesis alternations as a generalization that was motivated (in part) by the opacity of derived geminates.12 The fact that Bayso metathesis occurs only with coronal obstruents is a direct consequence, on our account, for the fact that only coronal obstruents triggered total progressive assimilation. Only in such cases was it plausible to reinterpret the historical reflex of a *Ct cluster as synchronically /tC/, that is, metathesized, and it is this pattern that was extended to original *Cn clusters. We cannot present equally detailed accounts of the nasal-obstruent metathesis alternations in other East Cushitic languages, but we do wish to note a striking fact that seems to us to support the general model we have applied to Bayso. In the three other East Cushitic cases (Qafar, Oromo, and a group of four Highland East Cushitic languages), the set of obstruents involved in *Cn > nC metathesis is always a subset of the set of obstruents that undergo total progressive assimilation in *Ct clusters (i.e., /Ct/ ! [CC]). The relevant facts are summarized in (17), where the second and third columns show the classes of obstruents that undergo metathesis and assimilation. (17)
Languages a. Qafar, northern dialect (Bliese 1981; Parker and Hayward 1985) b. Oromo, Wellegga dialect (Gragg 1976; Lloret 1995) c. Bayso (see above) d. Highland East Cushitic languages other than Burji (Moreno 1940; Hudson 1976)
/Cn/ ! [nC] /n/ ! [— ]
/Ct/ ! [CC] coronal obstruents (i.e., /t s/)
/t’n/ ! [n¢]
alveolar and palatal stops (i.e., /t d t’ ¢ c j c’/)
all coronals
all coronals other than /sˇ/ (which metathesizes) all nongeminate obstruents
all nongeminate obstruents
Analogical Morphophonology
537
On our account, the implicational relationship between elements in the second and third columns in (17) has a general explanation. The reinterpretation giving rise to metathesis alternations was based on a misanalysis of geminates that had resulted historically from progressive assimilation. Therefore, in each language showing metathesis, consonants undergo metathesis only if they also geminate. This is an accident if the cause of metathesis in each case was a *Cn > nC sound change.13 Though we have presented a detailed analysis for only one case in (17), we conclude that East Cushitic nasal-obstruent metathesis alternations do not pose a problem for the view (discussed in Blevins and Garrett 2004) that such alternations do not arise directly via sound change. In other words, nasal-obstruent metatheses are phonetically unnatural, and where they do exist they have evolved by processes of analogical change: the generalization to nasal-obstruent contexts of morphological alternations with independent sources.14 22.4
South-Central Dravidian
Though stop-stop metathesis does exist, we have argued elsewhere that metathesis of velar-labial (KP) stop sequences in particular is unnatural phonetically and undocumented as a sound change. By contrast, PK clusters do naturally and in several welldocumented cases undergo metathesis.15 In this section we discuss the only alleged KP > PK sound change known to us, and we argue that in fact it originated via analogical generalization of a fortuitous morphological pattern. (18) South Dravidian: Tamil, Malaya¯lam, Kannada, and others ˙ ˙ South-Central Dravidian Telugu Gondi ˙ Konda ˙ Kondh languages: Ku¯i, Ku¯vi, Pengo, Manda ˙ Central Dravidian: Kolami, Naiki, and others North Dravidian: Brahui and others The seven languages belonging to the South-Central subgroup of Dravidian languages are shown in (18), together with the other major Dravidian subgroups and a few other relevant languages of the family.16 The four Kondh languages are the locus of an apparent metathesis. All of them have synchronic alternations whereby underlying /kp/ and /gb/ sequences surface as pk and bg, respectively. Since there are no underlying or derived morpheme-internal /kp/ and /gb/ sequences and since the only su‰xes beginning with labial stops are verbal su‰xes, the relevant alternations are restricted to the verbal morphology and in particular the formation of verb stems.17 In addition to those shown below, relevant su‰xes include the Ku¯vi
538
Andrew Garrett and Juliette Blevins
desiderative su‰x /-p-/ (Israel 1979, 172) and a participial su‰x /-pi-/ (with other allomorphs, including /-bi-/) used in various periphrastic and grammaticalized periphrastic constructions. Most of these su‰xes show considerable allomorphy, but metathesis a¤ects only allomorphs beginning with /p/ or /b/. A context where metathesis is easily illustrated is the ‘‘plural action’’ formation, which is found in all Kondh languages and expresses plurality of actions or participants. This is marked by a su‰x whose allomorphs include /-ka/, /-pa/, and /-ba/, shown for Pengo by the data in (19a–c), respectively.18 (19)
Verb root a. katkerhiprazb. kapgrutpahonc. tubkavenhi-
Plural action stem kat-kaker-kahip-karas-kakap-pagrut-papa-pahon-patub-baka-baven-bahi-ba-
Verb gloss ‘cut’ ‘sing’ ‘sweep’ ‘cut’ ‘bite’ ‘fell’ ‘break’ ‘run’ ‘blow’ ‘burn’ ‘hear’ ‘give’
Examples of the /-pa/ and /-ba/ allomorphs with metathesis are shown in (20). Metathesis yielding pk is found only with verb roots ending in /k/, as in (20a), but as seen in (20b,c), respectively, roots in /k/ and /g/ can also select the /-ba/ allomorph (yielding surface bg). The examples in (19) and (20) all involve the Pengo plural action form, but metathesis is found in all Kondh languages. (20)
Verb root a. rikkukek b. kak›akhokc. pagpagtog-
Plural action stem ripkakupkaepkakabga- (/kak-ba-/) ›abga- (/›ak-ba-/) hobga- (/hok-ba-/) pabgapabgatobga-
Verb gloss ‘break’ ‘call’ ‘seek’ ‘vomit’ ‘sacrifice’ ‘wash (clothes)’ ‘kill’ ‘be split’ ‘trample’
The roots of the Kondh metathesis lie in Proto-Dravidian. In Dravidian languages, there are two widespread morphological mechanisms for deriving causative
Analogical Morphophonology
539
verbs or transitive counterparts of intransitive verbs. Devoicing a stem-final consonant is one mechanism, and the other involves a su‰x containing a *-p-.19 The two types of causative-transitive formation are illustrated in (21) with a few examples from the Kondh language Ku¯i. (21)
‘be small’ ‘be complete’ ‘be joined to’ ‘arrive’ ‘descend’ ‘return (intr.)’
Derived verb kok‘reduce’ muk‘complete’ a‘join in’ e-p‘cause to arrive’ d‰a-p‘bring down’ v›e-p‘return (trans.)’ )
)
Base verb a. kogmugab. ed‰av›e-
Especially relevant here is the causative formation in *-p-. Evidence from several Dravidian languages shows that there was originally also a formation in which the *-p- su‰x replaced the root-final consonant rather than simply being added to the verb root. We cite data from three languages in support of this claim.20 Telugu has several types of causative formation synchronically. One of these is shown in (22) with data from Krishnamurti (1961). (22) Base verb a˜gu uugu medugu cucu nil(u)cu mayu
‘to ‘to ‘to ‘to ‘to ‘to
stop, stay’ end’ be pounded’ see’ stand, cease’ be hidden’
Derived verb a˜-pu uu-pu medu-pu cu-pu nil(u)-pu ma-pu
‘to ‘to ‘to ‘to ‘to ‘to
hold back, restrain’ cause to end’ pound’ show’ set up, stop’ hide, screen, destroy’
According to Subrahmanyam (1971, 25), ‘‘-pu is substituted for the final syllable of some verbs ending in gu, cu, and yu,’’ though a simpler analysis synchronically might be that p replaces the final /g/, /c/, or /y/ of a causative base. In either case the point is that there is a substitution.21 (The final consonant of a Kondh verb stem corresponds via vowel reduction to a Telugu final CV sequence.) A second language showing this kind of replacement is Kolami (Emeneau 1961). The relevant causative surfaces as -(i)p- as in (23a). Shown in (23b) are cases where the causative su‰x replaces the final consonant of the base verb. (23)
Base verb a. serarudb. negayneray-
‘go’ ‘become dry’ ‘sit’ ‘fly’ ‘be spread’
Derived verb ser-p‘let (cattle) get lost’ ar-p‘dry (trans.)’ ud-ip‘make to sit’ nega-p‘make to fly’ nera-p‘spread (trans.)’
540
Andrew Garrett and Juliette Blevins
melgpergdig-
‘shake (intr.)’ ‘grow (intr.)’ ‘descend’
mel-pper-pdi-p-
‘shake (trans.)’ ‘rear’ ‘make descend’
As in Telugu, /y/ and /g/ are among the replaced consonants in these cases. A similar pattern is seen in Kolami’s close relative Naiki, whose g/p pairs are typically the exact cognates of their Kolami counterparts, to judge from the data cited by Emeneau (1961) and Subrahmanyam (1971, 42). Within South-Central Dravidian, the same pattern is seen not only in Telugu but also in Gondi, a language generally considered to be very closely related to the Kondh languages. The examples in (24) are from the Koya dialect of Gondi. (24) Base verb aБto stop (intr.)’ kaБto be boiled’ igg‘to descend’
Derived verb a-p‘to stop (trans.)’ ka-p‘to make boil’ i-pp‘to cause to get down’
Note that the first pair in (24) and the first pair in (22) are cognate, and the last pair in (24) and the last pair in (23b) are cognate; the second pair in (24) also has Kolami cognates. (We do not know the cause of the gemination recorded in the last Gondi example by Subrahmanyam 1968.) We conclude, for verbs ending in *g, that the Kondh languages inherited three mechanisms of causative-transitive formation: *p su‰xation; devoicing (*g ! *k), which operated very generally; and the more restricted *g ! *p replacement. Our analysis is based on this conclusion. A *g ! *p rule for causative formation would have been opaque and extremely vulnerable to a reinterpretation of some sort. A very natural reinterpretation of surface causatives with *p would have been that surface *p reflected underlying /kp/ via a synchronic /kp/ ! *p reduction. The /p/ would then be analyzed as the ordinary causative su‰x, with voicing assimilation and loss of the resulting /k/ (i.e., /g/ ! /k/ ! q before /p/). Hypothetical (‘‘Proto-Kondh’’) causatives like those in (25) would have been reanalyzed as shown. What is crucial here is that forms that had originated via the *g ! *p replacement process were construed as evidence for a synchronic /kp/ ! *p reduction. (25) Base verb *ag‘to stop (intr.)’ *pag‘to be split’
Derived verb *ap- (trans.) *pap- (trans.)
Reanalyzed as /ag-p-/ ! /akp-/ ! /ap-/ /pag-p-/ ! /pakp-/ ! /pap-/
The newly created process could then be extended to other contexts where a *pinitial su‰x was added to a velar-final root. In particular, we contend that this /kp/ ! *p process was employed in the formation of plural action verbs. The lexical
Analogical Morphophonology
541
details cannot be known (since they have been obliterated by subsequent changes), but two hypothetical examples are shown in (26). In each case, the result of the creation of a /kp/ ! *p reduction process was a new plural action form in *p, not *kp (as would previously have been the case). (26) Base verb *kuk‘call’ *pak‘split (trans.)’
Plural action form /kuk-p-/ ! *kup/pak-p-/ ! *pap-
Note that *pak- is the other productive causative-transitive of *pag- ‘to be split’ (i.e., derived by devoicing). The developments posited so far include only a relatively straightforward extension based on a demonstrably inherited Dravidian alternation. There is direct evidence also for the next step we posit: the hypercharacterization (double marking) of plural action forms like *kup- and *pap- in (26). It is crucial here that the plural action formation has two basic su‰x variants in the Kondh languages, /-p-/ and /-k-/. We propose that new plural action forms from verbs in *k were created as shown in (27). (27) Base verb *kuk‘call’ *pak‘split (trans.)’
Plural action form *kup- ! kup-k*pap- ! pap-k-
Here the basic idea is that *kup-, and so on, were insu‰ciently clear as plural action forms, and so the productive (default) plural action su‰x *-k- was added. This hypercharacterization process is generally comparable to the English creation of plurals like children and kine (in which a plural su‰x with an n was added to preexisting plural forms), and there is evidence for the same process elsewhere in the system of Kondh plural action forms. For a set of six Ku¯i verbs whose apparent plural action su‰x is /-pk-/, Emeneau (1975/1994, 228) suggests ‘‘that they should be compared with the remainder of the verbs of that [‘3rd’] conjugation . . . , which have -v- as the plural action allomorph and that -pk- should be regarded as representing a su‰xal doubling (-v-k- > -pk-).’’ The schematic representation in (28) is modeled after the proposed change in (27). (28) Ku¯i base a‘be, become’ si‘give’
(Pengo) (¼ a-) (¼ hi-)
Plural action form *a-v- (¼ Pengo a-b-) ! apk*si-v- (¼ Pengo hi-b-) ! sipk-
For the Ku¯i forms in (28) the only result of hypercharacterization was that a small set of verbs came to have /-pk-/ as their plural action su‰x, but for Proto-Kondh forms like those in (27) the result was far more profound. Here, because the base
542
Andrew Garrett and Juliette Blevins
forms (e.g., *pak- ‘split’) also accidentally had a *k, the -pk- sequence that arose from double marking looked exactly like the result of a metathesis. Precisely this appearance was taken as linguistic reality, in our view. Plural action forms like those in (27) were reinterpreted as the result of a synchronic /kp/ ! *pk metathesis. Obviously the motivation was that /kp/ is what these sequences ought to be underlyingly, since they were formed from verbs in *k and a su‰x *-p-. Here it is important to distinguish the basis for the reinterpretation we propose from its result. The result was that plural action forms like *papk- were interpreted as underlyingly /pak-p-/, with synchronic metathesis; but diachronically the *k originated as a (hypercharacterizing) form of the plural action su‰x. Two additional trivial changes occurred en route to the attested state of a¤airs. First, the new /kp/ ! *pk metathesis pattern was extended to all su‰xes beginning with *p, not just the plural action su‰x. Second, because of the productive causative relationship between verbs in *g (e.g., *pag- ‘be split’) and *k (e.g., *pak- ‘split’), the metathesis was also generalized to voiced velar stops. As a result, verbs like Ku¯i /pag-/ ‘be split’ have plural action forms like pabg-. These two changes completed the creation of a synchronic KP ! PK metathesis. Though unnatural phonetically, this process is pervasive in the morphology and apparently exceptionless in the language. 22.5
Conclusion
We have discussed three cases in which regular morphophonological patterns arose via analogical extension of fortuitous morphological patterns. The extensions are summarized in (29). (29) a. Greek
b. Bayso
c. Kondh
Fortuitous pattern Prevalence of T ! s in the perfect middle, leading to paradigm leveling and first person T ! s Independent assimilation patterns, causing regressive coronal stop þ /t/ assimilation to look like metathesis Hypercharacterization (with k) of forms derived by /k/ ! p substitution; then reinterpreted as metathesis
Extended to Other su‰xes with /m/
Other endings (i.e., with /n/)
Other su‰xes with /p/, /b/
The resulting sound patterns are schematically characterized in (30).
Analogical Morphophonology
(30) a. Greek: b. Bayso: c. Kondh:
T!s/ TN ! NT KP ! PK
543
m
Each alternation in (30) is restricted morphologically, operating only in verbs or verbal derivatives, but each is regular within its morphological domain. For the Kondh languages in particular, moreover, since underlying KP sequences do not arise elsewhere in the grammar, the alternation in (30c) is exceptionless. These patterns are of interest because (as far as we know) they could not occur (and are not attested) as sound changes. Two general conclusions follow. First, the new class of sound patterns identified here fills out the diachronic typology of morphophonological alternations. It is well known that ‘‘crazy rules’’ can evolve over time due to the cumulative e¤ect of changes that are each phonetically motivated but have a collectively arbitrary e¤ect. It is also well known that phonetically natural morphophonological patterns can arise by the mechanism of ‘‘morphophonemic analogy’’ (Moulton 1960, 1967). Yet it is usually assumed that morphophonological patterns as in (30) correspond to earlier purely phonological patterns. We have shown here that phonetically unnatural patterns can also arise by analogical processes. Since they are phonetically unnatural, they do not have purely phonological origins but reflect instead the generalization of fortuitous morphological patterns. A second conclusion is that even the most regular morphophonological patterns may lack phonetic origins. Phonological theories must therefore be broad enough to encompass phonetically unnatural alternations as well as natural ones. At the same time, theories must be properly constrained so that the two types are not confused, since confusion would falsely predict that alternations like those in (30) could arise through ordinary sound change.22 In sum, diachronic typology supports the traditional dichotomy between phonetically motivated sound change and word-based analogical change. As a concrete case where diachrony informs synchrony, analogical morphophonology reopens Kiparsky’s (1968b, 174) ‘‘window on the form of linguistic competence.’’ Once we recognize unnatural sound patterns with nonphonetic grammatical origins, constrained models of phonetically motivated sound change come into view. Notes For comments and helpful discussion we would like to thank Bruce Hayes, Dick Hayward, Larry Hyman, Sharon Inkelas, Joe Malone, Donca Steriade, and the late Murray Emeneau. The following abbreviations apply throughout this essay: sing ¼ singular; plur ¼ plural; 1sg ¼ first person singular; 2sg ¼ second person singular; 2sg.f ¼ second person singular, feminine; 3sg.m ¼ third person singular, masculine; 1pl ¼ first person plural; 2pl ¼ second
544
Andrew Garrett and Juliette Blevins
person plural; 3pl ¼ third person plural; intr ¼ intransitive; trans ¼ transitive; tr ¼ transitive. 1. For further discussion see also Wurzel (1980) and Anderson (1992, 339–346). 2. The phonetic origins of this deletion and related English sound patterns deserve further study. 3. The pattern in (1) is retained in some (e.g., West and Northwest Midlands) English dialects (Wells 1982). 4. The second and third cases, which constitute the bulk of this chapter, also contribute to our ongoing research on the diachronic origins of synchronic metathesis patterns (Blevins and Garrett 1998, 2004). 5. A phonetic explanation appealing to ‘‘rhinoglottophilia’’ (Matiso¤ 1975; Ohala 1975; Blevins and Garrett 1993) is implausible given the restriction to /m/ and coronal stops and because this change (unlike ordinary sound changes) did not operate morpheme internally. 6. Steriade (1982, 257) states the process as follows: [þcor, son] ! [þcont] / [þlab, þnas]. Basic references on Greek historical phonology and morphology include Schwyzer (1953), Lejeune (1972), Meillet and Vendryes (1979), and Rix (1992). 7. Note that no su‰xes beginning with /m/ are productively added to consonant-final noun stems. For discussion and data collections see Stratton (1899), Chantraine (1933), and Buck and Petersen (1945). Morpheme-internal clusters were apparently sometimes reinterpreted as heteromorphemic, leading to sporadic replacements like bat h mo´s > basmo´s ‘step’ and rhut h mo´s ‘rhythm’ > rhusmo´s. The coronal spirantization before /m/ discussed here should be distinguished from a more specific process, restricted to the Attic dialect (Threatte 1980), whereby original *dm > sm without exception, even morpheme internally (e.g., in the proper name Ka´dmos > Ka´smos). A likely explanation is that coda d became a fricative [D], and then [D] > [z] (spelled s) (cf. Schwyzer 1953, 208). 8. A chronologically later change not shown here is the degemination of ss, for example, in the 2sg. forms. 9. A synchronic stem-based analysis is also possible. Bases ending in coronal stops could be analyzed as having /s/-stems in the perfect middle, with the /s/-stem also phonologically conditioned by /m/-initial su‰xes. But the alternation remains phonetically unnatural whether it is characterized as phonologically conditioned stem selection or a segmental process T ! s / m. For relevant general discussion, see, for example, Kiparsky 1996a, 1996b. 10. See Blevins and Garrett (2004). Outside the domain of sound change, it is well known that sporadic metatheses are often unexpected phonetically, as in the typical folk-etymological case underlying the Old Icelandic variants rosmhvalr and romsval ‘walrus’ < *morsa < Sami morsˇsˇa (V. Kiparsky 1952, 30–44). 11. The four Highland languages with nasal-obstruent metathesis apparently form a subgroup within Highland East Cushitic. Note that since not all East Cushitic languages are well documented, similar metathesis alternations may well exist in other languages not highlighted in (9). 12. Compare Sasse (1976, 219), who also invokes a reinterpretation of opaque assimilated clusters to explain how /n/-initial verbal endings were replaced by /t/-initial endings in Dasenech.
Analogical Morphophonology
545
13. Note that Qafar retroflex / / and the Oromo implosive /¢/ are cognate, reflecting ProtoEastern-Cushitic *d’ (Sasse 1979; Ehret 1991); we do not know what bearing this may have on the distribution of metathesis in the two languages. 14. Joe Malone calls to our attention a similar sequence of changes in Classical Mandaic. (See also Malone 1971, 1985.) In that language, a synchronic alternation involves total regressive assimilation in /nC/ clusters; for example, the causative form of /npq/ ‘go out’ is /(h)anpeq/ ‘bring out’, which surfaces as [appe´q]. Variation appears in su‰xed third person masculine singular forms (e.g., [app qı´i] @ [hanp qı´i]). In such cases, it is unclear whether the regressive assimilation rule itself is subject to variability or whether some subsequent change of geminates into nasal-stop clusters is in progress (nC > CC > nC). Internal evidence supports the second account. Like most other Aramaic languages, Classical Mandaic has a minor rule (applying only to certain roots) by which /Cl/ ! [CC]. From a root /slq/ ‘go up’ ([s leq] ‘he went up’), which undergoes this minor rule, the attested causatives [asse´q] and [hanse´q] point to a sequence of developments sl > ss > ns. In this case ss > ns is not by morphological analogy; the direction of change has been determined by the dominant source of derived geminates from nasal-stop clusters. The sequence of changes sl > ss > ns is similar to the cases we discuss here. e
e
e
15. See Blevins and Garrett (2004) for discussion. The existence of PK > KP changes but not KP > PK changes is related to general properties of unordered labial þ dorsal gestures. 16. For general information about Dravidian (especially phonology), we have relied heavily on Emeneau (1970), Zvelebil (1970), Subrahmanyam (1971, 1983), Burrow and Emeneau (1984), and Steever (1998b). For information about South-Central Dravidian languages, we have used these sources: Burrow and Bhattacharya (1960), Subrahmanyam (1968), and Steever (1998a) on Gondi; Krishnamurti (1969) and Krishnamurti and Benham (1998) on Konda; ˙˙ ˙ Winfield (1928, 1929) and Burrow and Bhattacharya (1961) on Ku¯i; Burrow and Bhattacharya (1963) and Israel (1979) on Ku¯vi; Burrow (1976) on Manda; and Burrow and Bhattacharya ˙ (1970) on Pengo. We have profited from analyses of the˙ Kondh plural action formation by Subrahmanyam (1965), Emeneau (1975/1994, 223–262), and Steever (1993). 17. Once derived, infinitives may also be potential nouns, and some may be used in other functions. The point is that the su‰x undergoing metathesis is one that derives a verb stem. 18. The distribution of the three su‰x allomorphs is only partly predictable based on the shape of the verb root. For details see Burrow and Bhattacharya (1970, 82–85), who use the term ‘‘intensive-frequentative’’ for this formation. 19. See in general Subrahmanyam (1971, 1–101). What surfaces as a voicing alternation in the Kondh languages (and some other Dravidian languages) is elsewhere a consonant length alternation (with voiceless consonants corresponding to geminates); the correct reconstruction for Proto-Dravidian is debated. For convenience we will treat this here as a voicing alternation. 20. Since these languages belong to the South-Central and Central branches of Dravidian, it is possible that the process was restricted to those branches (and was not Proto-Dravidian). This is immaterial here, since the Kondh languages are South-Central Dravidian languages. 21. Historically it seems that distinct su‰xes are reflected here, but synchronically, though the labial su‰x has a clear function, there is no segmentable su‰x /-gu/. 22. Compare Hyman (2001). For the view that phonetically grounded constraints interact to determine phonological structure see Archangeli and Pulleyblank (1994), Flemming (1995), Hume (1997, 2001), Hayes (1999a), Steriade (1999b, 2001), Kirchner (2000), and Hayes, Kirchner, and Steriade (2001).
IV
Syntax and Semantics
23
Second-Position Clitics in Tagalog
Stephen R. Anderson
23.1
Introduction
In the study of clitics there are, as originally stressed by Arnold Zwicky 1977, two quite distinct notions that need to be disentangled, as in (1). (1) a. Simple clitics: prosodically weak elements that form part of a (phonological) word with other material from which they are syntactically distinct. b. Special clitics: elements that fill the grammatical function of some specifiable structural element but appear in an unusual linear or structural position, often one not available to corresponding nonclitic elements. One understanding of the notion of a clitic is represented by traditional views such as that of Wackernagel 1892, for whom a clitic was definitionally a prosodically weak element. Special behavior, including placement in a unique position, might well flow from this fact, but the basic property that makes something a clitic on this picture is the fact that for prosodic reasons it forms part of a word (or other prosodic unit) with other material from which it is syntactically distinct. An item with the property of prosodic dependency, and no other distinctive behavior, is the sort of thing Zwicky proposed to call a ‘‘simple’’ clitic. We can reconstruct this notion in contemporary terms in the following way: lexical items normally are assigned a full prosodic analysis with their segmental content organized into syllables that form parts of feet that in turn are grouped as a prosodic word. Exceptionally, some elements (including the simple clitics, as well as some of those items that display other unusual properties associated with other senses of the notion clitic) may be prosodically ‘‘deficient,’’ in that they constitute stray syllables not assigned to feet or feet that have not been assigned the status of words. A variety of phonological principles require that if such material is to be pronounced, it must be incorporated into a full prosodic structure. The rule or rules of Stray Adjunction, which perform this function, then describe the phonological dimension of cliticization: the incorporation of a prosodically isolated element into an adjacent preexisting word.
550
Stephen R. Anderson
A second, logically quite distinct dimension of cliticization is exemplified by the large number of studies that have dealt with elements such as the weak or conjunct pronominal forms of the Romance languages. These elements, along with others for which Zwicky proposed the term ‘‘special clitic,’’ are unusual in that, while they seem to fill the grammatical function of some specifiable structural element, they appear in an unusual linear or structural position, often one not available to corresponding nonclitic elements. The distinction between simple and special clitic elements suggests the existence of two mutually exclusive dimensions of clitic behavior—the phonological and the morphosyntactic dimensions, roughly. Of course, many special clitics are also prosodically dependent and thus display the phonological properties of simple clitics as well, establishing the notion that these two dimensions of cliticization are orthogonal rather than mutually exclusive. Not all special clitics are prosodically weak, though, just as not all prosodically dependent elements are syntactically unusual. The properties of being a simple clitic and those of being a special clitic are thus quite distinct per se, even if they often coincide in the same elements. In developing a general theory of clitic behavior, then, it is important to attend to both of these aspects of what we (presystematically) lump together as ‘‘clitics’’ and to recognize their potential interaction as well as their potential independence. The system of ‘‘second position’’ clitics in Tagalog provides a particularly interesting and intricate example of this point, and the goal of this chapter is to provide an explanatory account of the behavior of these elements. Tagalog (Schachter and Otanes 1972; Schachter 1973) displays clitic elements with two distinct sorts of function, all following the first element of the sentence. Note that within the pronominal clitics, there are two subsets, depending on the grammatical role of the element they represent. Since the analysis of grammatical relations in Tagalog clause structure is not at issue here and nothing in particular follows from the terminology, we simply follow Schachter in referring to these as the ‘‘Topic’’ (t) set and the ‘‘Complement’’ (c) set. (2) a. Pronominal clitics Topic 1sg 2sg 3sg dual 1pl (excl) 1pl (incl) 2pl 3pl 2sg t þ 1sg c
Complement
ako ko ka mo siya niya kata nita kami namin tayo natin kayo ninyo sila nila kaþko ! kita
Second-Position Clitics in Tagalog
b. ‘‘Particles’’ ba kasi kaya daw din ho lamang man muna
(interrogative) ‘because’ (speculation) (reported speech) ‘too’ (politeness) ‘only’ ‘even’ ‘for a while’
551
na naman nga pa pala po sana tuloy yata
‘already’ ‘instead’ ‘really’ ‘still’ (surprise) (politeness) (optative) ‘as a result’ (uncertainty)
All of these elements constitute special clitics in the sense introduced above, because they are all subject to a particular regularity in their placement: all come in second position, following the first element of the sentence. When more than one clitic is present, other regularities govern their sequence relative to one another. These will form the subject of discussion below; but before we can address those issues we must first develop a theory of second-position phenomena within which we can understand the overall placement of these classes of special clitics. 23.2
Background: The Theory of Second Position
We adopt the proposal of Anderson (1992, 2000, and elsewhere) that grammatical clitics, such as pronouns, tense/aspect markers, and so on are phrasal a‰xes, the overt realization of functional content associated (as features) with a phrasal node such as S (or IP), NP, and so on. This move is motivated in part by the extensive similarities between a‰xal material in words and the way clitics appear within phrases and sentences. These similarities are explored in Anderson 1992 and include the points in (3), among others. (3) a. They have a similar rigidity of ordering. b. Both can be nonconcatenative. c. Both make a distinction between ‘‘inflectional’’ (syntactically relevant) and ‘‘derivational’’ (semantically/pragmatically relevant). d. The same set of parameters describes in both cases the possibilities for placement (initial, final; postinitial; pre- or posthead).1 On this approach, where grammatical clitics serve as analogs at the phrasal level to the inflectional morphology of words, the question arises of how the correct surface order of elements is to be derived. In Anderson 1992, it was assumed that the clitics, like word-level a‰xes, were introduced one at a time by rules; and on that view, an obvious possibility was to relate the linear order of a‰xes and of clitics to the order of application of the rules that introduced them. In fact, however, there turn out to
552
Stephen R. Anderson
be problems with this, and a series of papers culminating in Anderson 2000 develops an alternative based on principles drawn from Optimality Theory (OT; cf. Prince and Smolensky 1993). In OT, descriptive order reflects the ranking of element-specific constraints, rather than a sequence in which the elements are introduced one at a time. Relative ordering results from the fact that where a number of elements should all preferably be located in the same position, the demands of some outrank those of others. To treat clitics within the terms of OT, we need to assume that clitic material is introduced within some domain and then located within that domain so as to conform as well as possible to a system of constraints. In the analysis of second-position phenomena, these constraints are of three general types. First, let us assume there is a family of constraints on morphosyntactic structure (or perhaps PF) EdgeMost(e,E,D), each of which says that the element e should be as close to the edge E (left or right) of the domain D as possible. Depending on the value of the edge parameter E, I will refer to these as LeftMost and RightMost. A given clitic/a‰x is characterized as a prefix or as a su‰x, depending on whether LeftMost(cli ,D) dominates RightMost(cli ,D) or vice versa. The descriptive order of two clitics cli and clj , both prefixes or both su‰xes, is determined by the dominance relation that obtains between their corresponding EdgeMost constraints. Infixes located with respect to the beginning of the word and second-position clitics can be described by saying that the element in question should, on the one hand, be as close as possible to the left edge of its domain (a word or a syntactic constituent) but, on the other hand, should not be absolutely initial. This e¤ect recalls a descriptive generalization known as the Tobler-Mussafia law, valid for some of the Old Romance languages according to which pronominal clitics cannot appear in absolute sentence-initial position. To describe this, let us assume a second constraint family Noninitial(e,D) each of whose representatives says the element e should not be initial within domain D. We characterize a clitic (or a‰x) cli as ‘‘second position’’ on this view by saying that Noninitial(cli ,D) dominates (otherwise high-ranking) LeftMost(cli ,D). The clitic/a‰x will then go as far to the left as possible without actually becoming initial: it will appear in second position. In some instances, the Tobler-Mussafia e¤ects of Noninitial(cli ) can be derived without the need to treat them as grammatical, since the independently motivated phonology will have this as a consequence. This is the case in Warlpiri, for example. In this language, auxiliary clitics are located in either first or second position within the clause. The principle that determines the choice is that when the base of the auxiliary is monosyllabic (or j), the auxiliary follows the first word of the sentence; while if the base of the auxiliary is bisyllabic, the auxiliary can appear either initially or in second position.
Second-Position Clitics in Tagalog
553
The ‘‘minimal word’’ in Warlpiri, as in most languages, consists of a bimoraic foot, which because there are no heavy syllables in the language must also be bisyllabic. Monosyllabic elements thus cannot be independently footed and must be treated as prosodically deficient. Assume that Stray Adjunction in Warlpiri always operates leftward; then in order to be incorporated into prosodic structure, a monosyllabic auxiliary base will have to be preceded by some other material. When we say that auxiliary bases are subject to LeftMost(cli ,S), the farthest to the left that they can go is after the first word, if they are to be incorporated into prosodic structure. That accounts for their appearance in second position. What shows us that it is really the phonology that accounts for the noninitial position of prosodically weak bases is the following additional fact: according to Simpson 1991, even ‘‘small’’ auxiliary bases can appear in sentence-initial position when the sentence in question is closely preceded in connected speech by other material. (4) Warlpiri ‘‘. . . in connected speech, monosyllabic aux bases are found sentence initially, because the last element of the previous sentence provides a phonological host for the clitics.’’ (Simpson 1991, 69) In that case, the auxiliary attaches phonologically to the final word of the preceding sentence. Clearly the relevant requirement on the base is not that it be noninitial within its clause but that it be located so as to be incorporated into preceding material within a phonological phrase (or some other relevant domain). Such an account may be available for some languages, but not for all. For instance, some second-position clitics in some languages are not prosodically deficient, in which case there is no phonological reason they could not be initial: it is just a fact that they must satisfy Noninitial(cli ,D). In some languages, some of the clitics appear in second position, but other (prosodically weak) clitics are allowed to appear initially (e.g., Bulgarian ne, sˇte as opposed to others; cf. Legendre 2000a). In such a language, Stray Adjunction must in principle be able to work in either direction and thus cannot replace the constraint Noninitial(cli ,D) entirely. In languages for which evidence of the sort found in Warlpiri is not available, then, we will express the notion that an element e appears in second position within a domain D by saying that LeftMost(e,D) is highly ranked within the overall hierarchy of constraints governing PF in that language but below Noninitial(e,D). We have assumed that the domain D to which these two constraints are relevant is the same, but some second-position phenomena are somewhat more complex. For instance, Richardson 1997 argues that in Czech, second-position clitics are subject to LeftMost(e,IP) but are required to be Noninitial(e,CP). Since the left edges of IP and CP generally coincide in main clauses, this di¤erence only shows up when other material, such as certain topicalized elements, is part of CP but external to IP. Such
554
Stephen R. Anderson
examples are part of the reason we analyze second position as resulting from the interaction of two constraints, rather than from a single constraint alone. The next issue that arises in connection with second-position clitics is that of just how much material appears to their left. In some systems, this is exactly one syntactic constituent, while in others it is a single word. Assume a third type of constraint, Integrity(Word), requiring that material not be introduced inside an existing word when that material does not represent part of the content of that word. In practically all languages, this constraint is undominated. A possible exception to this ranking is Pashto, about which there is some controversy (see Tegey 1977; van der Leeuw 1995, 1997 and references cited there; as well as Roberts 1996 for a di¤erent view). As a result, where Integrity(Word) prevails, there will always be at least one word between a second-position clitic and the left edge of the phrase. The only way there could be less would be by violating either Integrity(Word) or Noninitial(cli ,D); but by assumption both of these dominate LeftMost(cli ,D). On the other hand, given high-ranking LeftMost(cli ,D), the only way there could be more than one word before the clitic would be if some other, even higher-ranking constraint required it. We then describe cases where ‘‘second position’’ means ‘‘after the first phrase’’ by generalizing the notion of Integrity to a constraint family, the most inclusive instantiation of which is (unqualified) Integrity(XP), requiring that phrases not properly contain elements that are not members of that phrase. In some languages, some phrasal types may be more ‘‘permeable’’ than others, requiring a more specific parameterization of the class of phrases whose interruption counts as a violation of Integrity(XP). Now suppose that Integrity(XP), like Integrity(Word), is undominated in some language. In that case, the earliest that second position can come is after the first phrasal daughter of the containing domain, and so that is exactly where second-position clitics will lodge. That shows us how to describe each type of second position: in languages where this means ‘‘after the first phrase,’’ Integrity(XP) is undominated. In languages where it means ‘‘after the first word,’’ on the other hand, Integrity(Word) is undominated, but Integrity(XP) is dominated by the LeftMost(cli ,D) constraints for the various clitics, meaning that their desire to get to the left can violate phrasal (but not word) integrity where necessary. Integrity(Word) and Integrity(XP) are obviously variations on a single theme: constraints to the e¤ect that material cannot be properly contained within a domain unless it represents a member or element of that domain. Integrity(C) is a family of constraints where C ranges over prosodic and syntactic category types. Given that, as noted above, the class of constitutent types displaying such e¤ects di¤ers to some extent from language to language, we need to assume that C in Integrity(C) is parametrically variable.2
Second-Position Clitics in Tagalog
555
The theory of clitics (especially second-position clitics) outlined above is significantly di¤erent from the more standard assumption that these are syntactically autonomous elements that are placed by rules of the syntax. These two points of view have been contrasted in previous work (such as Anderson 2000), and the plausibility of the theory defended here has been further supported by its application to a variety of linguistic systems. In particular, work by Legendre in a series of papers (1998a, 1998b, 2000a, 2000b) has pursued essentially the same approach, yielding important insights into the clitic systems of several Balkan languages. Legendre’s work is explicitly related to that discussed here and shares with it the central assumption that in order to achieve the potential benefits of an OT-like description, it is essential to treat the clitics in question as phrasal a‰xes rather than as syntactic heads. 23.3
Clitics in Tagalog
On the basis of this overall view of second position, we can now proceed to the specific properties of the Tagalog clitics mentioned earlier. The overall framework for the description of clitics proposed above attributes their complex surface properties to the interaction of (individually quite simple) principles from a number of areas of grammar—at least phonology, morphology, and syntax. The analysis of the Tagalog clitics illustrates virtually the full range of considerations that can interact in determining the placement of clitics. We note at the outset that the property of appearing in second position characterizes only some clitics in Tagalog. It is not the case, that is, that all phonologically weak elements appear in this position (as Wackernagel assumed in his 1892 analysis of the early Indo-European languages), since Tagalog also has a set of sentence-final particles a, e, ha, o (Schachter and Otanes 1972, 461–463). When more than one of the second position clitics is present in a sentence, their relative ordering is governed by a set of principles that have been the subject of some discussion. The basic ordering regularities are the following. (5) a. Monosyllabic members of the set of pronominal clitics (i.e., the monosyllables in (2a)) always precede particles (the elements listed in (2b)). b. Particles have an internal ordering among themselves. c. Particles always precede disyllabic pronominal clitics. Within each of the three sets of clitics ordered by these principles, there is a fixed relative ordering. In the case of the two subsets of pronominals, this ordering is presumably stipulated by the ranking of corresponding LeftMost constraints, since it does not appear to follow from any more general principles. In the case of the particles, there is also a determinate ordering among the elements. These fall into several
556
Stephen R. Anderson
subclasses, whose relative order we presume can be made to follow from a combination of semantic scope and language-particular stipulation, much like the relative order of derivational elements in morphology, though we do not attempt to develop the relevant principles here. There is some agreement in the descriptive literature, such as Bloomfield 1917 and Schachter and Otanes 1972, that these clitic elements generally follow the first stressbearing word of the sentence, though there are as usual some complications in this picture. The examples that are most directly consistent with this simple picture involve clitics that appear to interrupt otherwise unitary constituents so as to come after a single sentence-initial word. (6) Ganu ¼ka ¼na ¼ba kakinis? how you already int clever ‘How clever are you now?’ (Bloomfield 1917, 143) On the other hand, there are various constructions consisting of more than one phonological word that do not get broken up (Schachter and Otanes 1972, 187¤ ): proper names, numerical expressions, times of day, ages, amounts of money, and so on. (7) Bukas ng gabi nang alas otso ¼siya aalis tomorrow night at eight o’clock he leaves ‘It’s tomorrow night at eight that he’s leaving.’ (Schachter and Otanes 1972, 188) These facts are described by Schachter and Otanes 1972 as based on treating certain construction types, the ones they call ‘‘obligatory non-pre-enclitics,’’ as uninterruptable. In the present theory the expression of this uninterruptability is an appropriately parameterized high-ranking Integrity constraint. I have nothing particularly illuminating to say here about why some constructions have this property while others do not. Where a sentence-initial constituent can be interrupted, Integrity(Word) in combination with Noninitial(e,D) nonetheless prevents the clitics from appearing any farther to the left than after the initial word, as in cases like (6). One further refinement relates to the domain in which the clitics are placed and involves the fact that second position sometimes has to be within a domain that excludes certain preposed, focused constituents. This point will be further discussed below; for now let us simply assume that the domain within which some element appears in second position is that within which it is introduced, which follows (in some way) from its grammatical properties. These complications do not a¤ect the major points to be made below, however: what matters is that second position has the same basic sense for all of the elements in (2).
Second-Position Clitics in Tagalog
557
Tagalog clitics pose problems of varying severity for an account based on purely syntactic mechanisms. First, there is the usual di‰culty of placing clitics ‘‘after the first word,’’ a notion that does not have a natural reconstruction in terms of those parts of grammar that normally deal in terms of phrases as basic elements. This is especially true if the relevant notion of ‘‘word’’ is a phonological one, which would result in di‰culties for a principle of ‘‘Phonology-free Syntax’’ (Pullum and Zwicky 1988). A more serious problem is that of deriving the regularities of order among clitics. If these are as stated in (5) above, it is not clear how they are to be described syntactically at all. This matter is the subject of Schachter 1973: the problem is that the order appears to depend on the phonological property of how many syllables a given clitic has, and syntactic rules, even of the sort one accepted back in 1973, ought not to have access to the specifics of the phonological realization of elements they a¤ect. But how can this be avoided? Note in particular the portmanteau element kita: this clitic represents the combination of a 2sg.Topic and a 1sg.Complement and replaces the expected sequence kaþko. Each of the latter elements individually is monosyllabic and thus should precede any particles (by (5a)), but when they come together, the combination is replaced by a disyllabic portmanteau that must follow particles (by (5c)). Even a set of rules that places the pronouns (as syntactic elements) one by one and that treats the length regularities as accidental will fail in this case. Schachter 1973 evaluated several brute force solutions to these di‰culties utilizing reordering transformations and showed that they either fail altogether or at least fail to capture the generalization about length. His own proposal, following Perlmutter’s 1971 analysis of French and Spanish clitics, is a surface constraint: a template explicitly stipulating that monosyllabic pronouns precede particles, which in turn precede longer pronouns. (8) Pro1s < Particles < Pro2s This, he suggests, could serve either to filter the output of a general scrambling rule or to linearize elements at surface structure. Such a description would at least allow us to state the regularity about syllable structure, but it does not explain either of the relevant facts: (i) monosyllabic pronouns come first; and (ii) otherwise, pronouns come after particles. Subsequent work (e.g., Kroeger 1993) similarly leaves this pattern as a matter of stipulation. In Anderson 1992, I noted that an explanation of the fact that at least most of the pronouns follow particles might be derived from the analysis of clitics as sentential morphology. Within words, Derivational Morphology comes ‘‘inside of ’’ Inflectional Morphology. If we ignore the monosyllabic pronouns for the moment and think of clitics as su‰xes to the first word or phrase, then we could say the ordering of
558
Stephen R. Anderson
particles before pronouns reflects the same regularity: particles have various semantic and pragmatic content and are thus ‘‘derivational,’’ as opposed to pronouns, which are (on this analysis) agreement markers and thus ‘‘inflectional.’’ It would be natural for pronominal elements to be su‰xed outside of particles, then, and thus to follow them. What accounts for the extension of the ‘‘derivation inside of inflection’’ theorem from morphology to the description of clitics? In the word-internal case, this follows (as argued in Anderson 1992) from the place of lexical insertion in a grammar, given the nature of (productive) inflection. In the case of clitics, the corresponding result probably also follows from the architecture of grammar. Adopting the overall terms of Chomsky 1995, chapter 4, we might consider the introduction of semantically contentful particle clitics to be a generalized instance of the operation Merge. In the development of complex syntactic structures, the usual step consists of merging a word (taken from the structure’s numeration) with (some part of ) an existing syntactic structure. In some instances, though, I suggest that the way Merge works is to introduce some a‰x-like material (rather than an autonomous word) into the structure. If we suppose, as seems natural, that all instances of this generalized Merge operation must be completed prior to the point at which clausal agreement-like material has to be realized phonologically, the particle clitics will all be in place at the point where the structure exists to support the introduction of grammatical (pronominal) clitics. As a result, grammatical or inflectional clitics (such as the Tagalog pronominals) will be introduced ‘‘outside of ’’ semantically and pragmatically contentful (or derivational) elements such as the Tagalog particles. Obviously, this involves a certain amount of serialism in the overall derivation of a sentence—a notion that is not fundamentally at odds with the invocation of principles from OT to describe what happens at a given stage of the derivation, though it is at odds with the common (but incorrect) assumption that constraint systems must always involve a single monolithic collection of constraints applied in parallel. A similar point is developed in a much broader context in Kiparsky (in press), where it is suggested that much of the nonmonolithic architecture of Lexical Phonology is in fact perfectly consistent with the basic insights of a constraint-based approach such as OT. But what about the syllable length regularity? If the explanation just o¤ered is correct, it is clearly an embarrassment to have some inflectional clitics coming inside of the derivational ones. In Anderson 1992, I proposed that the monosyllabic clitics are introduced by ‘‘head-inflection’’ rules, as opposed to all the others, which are ‘‘word inflection.’’ This analysis, however, still misses the fact that the di¤erence depends on prosodic size. In a step backward from Schachter’s (1973) analysis, it does not even state this fact, let alone explain it.
Second-Position Clitics in Tagalog
559
Is it possible to eliminate this anomaly? Suppose that the relation between particles and disyllabic pronouns does indeed follow from an appropriate generalization of the ‘‘derivation precedes inflection’’ theorem. That is, all of the derivational clitics are already in place, attached to their host, at the point where the pronominal ones are introduced. The di¤erence between principles (5a) and (5c) must then follow from some other aspect of the grammar of clitics. Where might it be appropriate to locate a di¤erence between the behavior of disyllabic and monosyllabic clitics? Since the difference is itself phonological, the obvious place to look is in the phonology. Let us suppose that monosyllabic and disyllabic clitics di¤er in the way they undergo Stray Adjunction and that this di¤erence interacts with the system of constraints determining their behavior as second-position elements to yield the observed placements. As in most languages, the minimal foot in Tagalog is bimoraic. As a result, disyllabic pronominal clitics (and particles) are prosodically complete enough to be structured as feet. Monosyllabic pronominals, on the other hand, are monomoraic and thus cannot constitute feet on their own: they are simply stray syllables. Both are simple clitics, however, in the sense introduced in section 23.1, since they are not phonological words; and so both types must be incorporated into an adjacent word by Stray Adjunction. (9) a. Disyllabic clitic
b. Monosyllabic clitic
How are these orphan elements to be provided with homes? That is, how does Stray Adjunction operate? Suppose we say, in the spirit of the widely accepted Prosodic Hierarchy of Nespor and Vogel 1986 and other work, that a stray foot can (and must) be incorporated into the prosodic word to its left, while a stray syllable is adjoined to a foot on its left. The pronouns are true second-position clitics—that is, they are subject to high-ranking LeftMost and Noninitial constraints while not being allowed to violate a variety of Integrity constraints. In particular, lexical items are integral in this sense, as are various uninterruptable phrasal types that we noted above.
560
Stephen R. Anderson
Stray Adjunction thus presents two cases. (10) a. Stray foot adjunction
b. Stray syllable adjunction
The di¤erence between incorporation (in the foot case, (10a)) and adjunction (in the case of stray syllables, (10b)) in these rules follows from the well-known fact that the internal constituency of feet is rather narrowly constrained, while phonological words are somewhat more open ended and display an organization that is flatter and internally more homogeneous. To see how this provides an account of our problem, consider a sentence such as the following, with a monosyllabic clitic, a particle, and a disyllabic clitic, in that order. (11) nakikita ¼ka ¼na ¼niya. sees you (sg.) already he ‘He sees you (sg.) now.’ In the construction of this phrase, we begin with a sentence-initial host consisting of a single word, nakikita. Within the syntactic system, this comes to be followed by a particle ¼na as a consequence of the application of derivational cliticization. The structure in (12a) thus becomes that of (12b) at the output of the syntax. (12) a.
Second-Position Clitics in Tagalog
561
b.
We then add the pronominal clitics. When the pronoun is disyllabic, it must be parsed as a foot and thus has to follow the entire lexical word plus particle sequence. The latter is (as a whole) a phonological word (perhaps derived by previous instances of Stray Adjunction incorporating particle clitics into the word defined by the host). The clitic must follow this entire sequence because (by rule (10a) above) the only place where a foot can undergo Stray Adjunction is at the right edge of a phonological word. Where the pronominal clitic is monosyllabic, on the other hand, it is only a syllable, not a foot. Potentially it can (and thus, by LeftMost, must) be placed farther to the left within the word. These clitics must be placed at the right edge of a foot so that they can undergo Stray Adjunction by rule (10b). They must also not violate higher-ranking Integrity constraints, including the integrity of lexical words, but they must otherwise be as close as possible to the left edge without being initial. As a result of the interaction of these factors, they will lodge at the right edge of the leftmost foot that is not properly included within a lexical word (or other integral domain). This means, in particular, that they appear to the left of the particles. These two types are shown below in (13).
562
Stephen R. Anderson
(13)
We have assumed here that the added monosyllable ¼ka is simply adjoined to the ¼na]. Depending on innermost foot within the already complex foot [ f [ f ki ta] the details of stress contours in such a configuration, we might well replace this analysis with one where the two adjoined syllables are restructured to establish a new foot of their own, yielding the following structure. (14) [PrWd [ f na ki] [ f ki ta] [ f ka na] [ f ni ya]] We do not attempt to develop the mechanical details of this refinement of our proposal here, but they do not appear to present conceptual problems. The phonology of Stray Adjunction, then, taken together with the constraints that define second position, can actually explain why monosyllabic clitics behave differently than disyllabic ones. It is important to note that this result depends crucially on the constraint-based approach adopted here. That is, on the present picture, instead of specifying a unique position in hierarchical (or linear) structure in which clitics are generated or to which they move, the present analysis says that clitics are attracted as far to the left of the relevant domain as possible, subject to the higherranking requirement that they not be absolutely initial. The definition of ‘‘second position’’ that results is a consequence of the interaction of a number of other e¤ects (especially the details of the language’s phonology and of the relevant Integrity constraints). Since these factors play out di¤erently in the cases of monosyllabic and of disyllabic clitics, these elements appear in di¤erent positions, in ways that could not be described adequately by syntactic mechanisms per se. To this point, we have assumed that the Tagalog clitics—particles and pronouns alike—are to be characterized collectively as second-position elements in the sense that for each clitic clj there is a constraint Noninitial(clj ,D) outranking an otherwise
Second-Position Clitics in Tagalog
563
highly ranked constraint LeftMost(clj ,D). We noted in an earlier section, however, that in some languages (e.g., Warlpiri), the Noninitial e¤ect results not from an independent constraint but rather from phonological requirements alone. Could this be the case for the Tagalog clitics? A di¤erence between the e¤ects of Noninitial(clj ) and the phonological behavior at issue might be explored if we could find a circumstance where the domain in which some clitic is subject to Noninitial(clj ,D) is closely associated phonologically with preceding material. In that case, if the situation is in fact similar to that in Warlpiri, we might expect that the clitic would appear initially within the domain D; while if Noninitial(clj ,D) in fact outranks LeftMost(clj ,D), the clitic should continue to be post initial within D. It is always di‰cult to be absolutely certain that the phonological adjacency necessary to establish this kind of situation is in fact met, but there is one construction in Tagalog that appears to provide us with evidence of the type we seek. A range of inversion constructions (described by Schachter and Otanes 1972, 485¤ ) involves the movement of a topical constituent to an initial position, presumably SpecCP . When this happens, there may be a pause between the topic phrase and the rest of the sentence (which would, we assume, block stray incorporation of an IP-initial clitic); but there need not be, especially in the case where the topic is marked by the element ay (presumably in C). It would appear that this preceding phrase would provide an appropriate prosodic host for an IP-initial clitic, if the latter element were not subject to Noninitial(clj ,IP), parallel to the facilitating role of a preceding connected sentence in Warlpiri. In such sentences in Tagalog, however, a pronominal clitic continues to appear in second position within IP, as in example (15) below. (15) Ang sulat ay tinanggap ¼ko kahapon. the letter ay received 1sg yesterday ‘I received the letter yesterday.’ (Schachter and Otanes 1972, 486) This is at least weak evidence that the position of pronominal clitics is in fact governed by high-ranking grammatical constraints of the type Noninitial(clj ,IP). Furthermore, there is no evidence similar to that found in Warlpiri that would argue that this phenomenon is purely phonological in origin. When we consider corresponding data involving particle clitics, however, a rather surprising fact emerges. Consider the following pair of sentences. (16) a. [CP [IP
sasayaw dance ‘Will they dance b. [CP bukas tomorrow ‘Will they dance
¼ba ¼sila ng pandanggo bukas ng gabi?]] Q they fandango tomorrow night a fandango tomorrow night?’ ¼ba ng gabi ’y [IP sasayaw ¼sila ng pandanggo?]] Q night ay dance they fandango a fandango tomorrow night?’
564
Stephen R. Anderson
In sentence (16a) we see the particle ba preceding the pronoun sila ‘they’ as expected, both coming in second position. In sentence (16b), however, where the phrase bukas ng gabi ‘tomorrow night’ appears in fronted topic position, the two clitics are separated. Sila appears as expected in second position within the nucleus of the clause, but ba appears within the topic phrase. We cannot account for the facts in (16) by saying that particles, unlike pronouns, are not subject to Noninitial(cl,IP) and are constrained to be noninitial only by phonological factors. If that were the case, we would expect the particle to appear at the left edge of IP, as in the ungrammatical (17). bukas ng gabi ’y [IP ¼ba sasayaw tomorrow night ay Q dance ‘Will they dance a fandango tomorrow night?’
(17) *[CP
¼sila they
ng pandanggo?]] fandango
The fact that the particle appears embedded within the phrase bukas ng gabi (following its first word) shows that it must in fact be located with respect to the left edge of a larger, including constituent (probably CP). In nontopicalized main clause constructions, the left edges of IP and CP will coincide, and so the particles and the pronouns will form a single sequence located in the same ‘‘second position.’’ Where there is a preposed topic phrase, however, the di¤erence between the two domains is revealed and the clitics become separated. One group of particles behaves like ba in (16) above, coming inside the topic phrase. Other particles, however, behave like the pronouns in always appearing within IP, such as lamang in the sentences in (18). sasayaw ¼lamang ¼sila ng pandanggo bukas ng gabi]] dance just they fandango tomorrow night ‘They’ll just dance a fandango tomorrow night.’ b. [CP bukas ng gabi ay [IP sasayaw ¼lamang ¼sila tomorrow night ay dance just they ng pandanggo]] fandango ‘They’ll just dance a fandango tomorrow night.’ c. *[CP bukas ¼lamang ng gabi ay [IP sasayaw ¼sila tomorrow just night ay dance they ng pandanggo]] fandango ‘They’ll just dance a fandango tomorrow night.’
(18) a.
[CP [IP
It seems, then, that the particles fall into at least two subclasses, one of which is subject to high-ranking Noninitial(cl,CP) while the other is subject to Noninitial(cl,IP). This di¤erence is a property of the particles as grammatical elements and does not follow simply as a consequence of their phonological form.
Second-Position Clitics in Tagalog
565
According to Schachter and Otanes 1972, 429¤, the particles that behave like ba in (16) include both monosyllabic elements (ba, man) and bisyllabic ones (kasi, kaya). Similarly, the set that is excluded from sentence-initial topic phrases includes not only bisyllabic particles like lamang, muna but also monosyllables (na, pa). A third subset, including both monosyllabic and bisyllabic items, can appear in either position, depending (apparently) on considerations of semantic scope.3 If this speculation is indeed correct, it points the way to the essence of the di¤erences we are examining here. Assuming that particles are introduced within a domain that constitutes their scope, as we might wish if the operation Merge is to be semantically compositional, we could say that the relevant LeftMost and Noninitial constraints are relativized to the same domain. Since the domain of introduction of pronominal elements is presumably uniformly IP, their placement in second position within IP is consistent with this account. It seems, then, that we have no evidence in favor of the notion that the noninitial position of any of the Tagalog clitics follows from phonological considerations alone. Instead, each clitic’s location is governed by the interaction of appropriate LeftMost and Noninitial constraints, relative to the grammatical domain (either IP or CP, depending on the clitic) within which it is introduced. The suggestion of Legendre 1998a that Tagalog clitics are placed in second position within a phonological constituent (specifically, a phonological phrase) rather than a grammatical one is thus not supported. Tagalog thus presents a particularly interesting and intricate illustration of the interaction of a variety of factors in determining the exact location of second-position clitics. (19) a. A constraint system, including i. LeftMost alignment constraints; ii. Noninitial alignment constraints within an appropriate domain (IP or CP); iii. Hierarchical ranking of alignment constraints for specific clitics; iv. Integrity constraints, which characterize uninterruptible subparts of a domain and thus define what counts as occupying ‘‘first’’ position; b. An architecture of grammar in which ‘‘derivational’’ clitics result from an operation of Merge, while ‘‘inflectional’’ ones realize the functional categorial morphology of phrases and clauses; and c. Phonological e¤ects, in particular the details of Stray Adjunction phenomena. This complex picture has a natural place within a theory that draws on the basic ideas of OT, such as that pursued in the research program summarized in the introductory sections of this chapter.
566
Stephen R. Anderson
Notes The research on which this chapter is based was supported in part by NSF award BCS9876456 to Yale University. I am grateful to a number of audiences to whom this material has been presented in recent years. The first occasion on which I discussed the treatment of Tagalog here was in a series of talks at Stanford in January, 1998, where I benefited from the comments of Paul Kiparsky among others. Too many people have contributed to my understanding of the problems below to list individually—or for there to be any real excuse for the remaining errors and misconceptions, which are entirely my own. The following abbreviations apply throughout this chapter: 3sg ¼ third person singular; 2sg ¼ second person singular; 1sg ¼ first person singular; 3pl ¼ third person plural; 2pl ¼ second person plural; 1pl ¼ first person plural; aux ¼ auxiliary verb; t ¼ topic; c ¼ complement; incl ¼ inclusive; excl ¼ exclusive. 1. There is of course a large theoretical literature on the way clitics, especially ‘‘secondposition’’ clitics should be characterized with respect to their placement. These include Kaisse 1985, Klavans 1985, Inkelas 1989, Halpern 1992, and Anderson 1993, to cite only a few who have contributed more or less directly to the view presented here. See these works for more general discussion and exemplificiation of the full range of special clitic constructions in natural languages. 2. Notice that we have explicitly provided for the possibility that the domains in question may be defined either syntactically or prosodically. The same is true, in principle, for the domains within which constraints of the EdgeMost and Noninitial families are applicable. There may well be generalizations that restrict some of these constraints to one sort of domain or the other, as argued (mutatis mutandis) in Zec and Inkelas 1990, but we neglect that refinement here. 3. See Condoravdi and Kiparsky 1998 for a discussion of the interaction of scope with alignment constraints in the placement of a‰xal material in the Tagalog word.
24
‘‘Elsewhere’’ in Gender Resolution
Stephen Wechsler
24.1
Introduction
Coordinate structures such as the subject of the French sentence (3) pose an interesting challenge for the grammar of agreement. (1) Le garc¸on est compe´tent (m.sg) / *compe´tente (f.sg). ‘The boy is competent.’ (2) La fille est *compe´tent (m.sg) / compe´tente (f.sg). ‘The girl is competent.’ (3) [Le garc¸on et la fille] sont compe´tents (m.pl) / *compe´tentes (f.pl). ‘The boy and the girl are competent.’ The predicate adjective shows gender agreement with its subject, masculine in (1) and feminine in (2). When masculine and feminine are conjoined as in (3), the predicate appears in masculine plural in French. More generally, languages follow two strategies to handle coordination of unlike conjuncts, either ‘‘resolution’’ or ‘‘partial agreement’’ (this terminology follows Corbett 1991). A resolution rule derives the agreement features of a coordinate NP on the basis of the features of all the individual conjuncts. Example (3) illustrates the purported French resolution rule dictating that mixes of genders are resolved to the masculine. In ‘‘partial agreement,’’ on the other hand, agreement consults one conjunct and ignores the other(s), as in this example from Ndebele (Bantu) taken from Moosally (1998, 88). (4) In-khezo lemi-ganu i-qamukila. 12-spoon conj.4-plate 4-broke ‘The spoons and plates broke.’ In Ndebele partial agreement, the predicate agrees with the closest conjunct, ignoring any others.
568
Stephen Wechsler
This chapter largely sidesteps partial agreement (on which see, inter alios, Corbett 1991; Johannessen 1996, 1998; and Moosally 1998), focusing instead on gender resolution. While there has been relatively little formal generative work on this problem, some recent generative treatments of resolution (see section 24.3 below) have proposed formal mechanisms for computing the agreement features of the coordinate NP on the basis of the features of the conjuncts (Dalrymple and Kaplan 2000; Vincent and Bo¨rjars 2000). To date, these feature computation mechanisms have not succeeded in capturing the attested range of systems, both undergenerating (SerboCroatian and Slovene are problems; see sections 24.3 and 24.6) and perhaps overgenerating as well (some unattested systems we predict to be impossible could be generated by Dalrymple and Kaplan 2000; see section 24.6). Recognition of a default or ‘‘elsewhere’’ (Anderson 1969; Kiparsky 1973b) gender, together with a theory relating grammatical to semantic gender categories, will be shown below to solve these problems. Even more important than the empirical problems is the standard question of explanation: Why do the feature computation rules and the featural representations of the genders take the specific forms that have been posited? A feature resolution mechanism specialized for coordinate NPs fails to relate the details of the resolution system in a given language to other aspects of the grammar of that language. In contrast, the present study explains the gender resolution facts in terms of the semantics of grammatical gender, understood within a theory of markedness. As Corbett (1999, 402) observed, ‘‘genders always have a semantic core (there are no purely formal systems).’’ Though grammaticalized, genders can be shown to have optional semantic content that surfaces under predictable circumstances, with consequences throughout the grammar of agreement. Among those consequences are the patterns of agreement with coordinate NPs. 24.2
Gender Resolution Patterns
Continuing with the French example above, consider the e¤ect on predicate agreement when we conjoin two feminines (5), two masculines (6), and a mixture of masculine and feminine (7)¼(3). The system is summarized in (8). (5) [La fille et sa soeur] sont *compe´tents (m.pl) / compe´tentes (f.pl). ‘The girl and her sister are competent.’ (6) [Le garc¸on et son fre`re] sont compe´tents (m.pl) / *compe´tentes (f.pl). ‘The boy and his brother are competent.’ (7) [Le garc¸on et sa soeur] sont compe´tents (m.pl) / *compe´tentes (f.pl). ‘The boy and his sister are competent.’
‘‘Elsewhere’’ in Gender Resolution
569
(8) French rule (M, F) a. f.pl: f þ f b. m.pl: elsewhere The letters m, f on the first line of (8) indicate that French has a two-gender system, with the genders masculine (m) and feminine (f). As shown, a group of feminine conjuncts triggers feminine plural agreement, and masculine plural is the default used in all other cases. The three-gender system of Serbo-Croatian (masculine, feminine, neuter) has a similar pattern. A group of feminines yields feminine plural agreement, as shown in (9). All other combinations result in masculine plural agreement. Perhaps most surprising, a coordinate structure with all neuter gender conjuncts triggers, not neuter plural but masculine plural agreement, as shown in (11). (9) Sve snage i sva paznja bice posvecene toj borbi. all powers(f.pl) and all attention(f.sg) will be dedicated-f.pl to this struggle (10) Ogledalo i cetka za kosu su bili na stolu. mirror.nt.sg and brush-f.sg for hair aux were-m.pl on table ‘The mirror and the hair brush were on the table.’ (11) Ogledalo i nalivpero su bili / *bila na stolu. mirror.nt.sg and fountain pen-nt.sg aux were-m.pl / *nt.pl on table ‘The mirror and the fountain pen were on the table.’ The Serbo-Croatian rule is actually identical to the French rule above but has a different e¤ect because the language has three rather than two genders. (12) Serbo-Croatian rule (M, F, NT) a. f.pl: f þ f b. m.pl: elsewhere Turning to Icelandic, when gender is homogeneous across all conjuncts then that gender is inherited by the coordinate structure. Any heterogeneity of gender is resolved with the neuter plural form, as in (13) and (14). The Icelandic pattern is given in (15). (13) Drengurinn og telpan eru reytt. the.boy and the.girl are tired.nt.pl (14) E´g sa´ a´ og lamb, boeDi svort. I saw ewe(f) and lamb(nt), both.nt.pl black.nt.pl (Corbett 1991, 283)
570
Stephen Wechsler
(15) Icelandic rule (M, F, NT) a. f.pl: f þ f b. m.pl: m þ m c. nt.pl: elsewhere (including nt þ nt and all mixtures) 24.3
Syntactic Feature Computation
Dalrymple and Kaplan (2000) propose a theory of feature resolution based on the operation of set union, implemented within the framework of Lexical-Functional Grammar. Gender values are represented with sets of abstract features. For example, in French the masculine value is represented with the singleton set containing the element m, while feminine is represented with the null set. (16) French masc: fem:
{m} { }
To calculate the gender value for a coordinate structure, simply union the sets associated with the conjuncts. The French resolution pattern at left is correctly captured by the set designations in (16) above. masc & masc ¼ masc masc & fem ¼ masc fem & fem ¼ fem
{m} W {m} ¼ {m} {m} W { } ¼ {m} { }W{ } ¼ { }
Turning now to the three-gender system of Icelandic, Dalrymple and Kaplan designate the following sets for the gender values. (17) Icelandic masc: {m} fem: {f} neut: {m, f} The Icelandic coordination pattern is also captured on the Dalrymple and Kaplan system, as shown by the set union operations at right. masc & masc ¼ masc fem & fem ¼ fem neut & neut ¼ neut masc & fem ¼ neut masc & neut ¼ neut fem & neut ¼ neut
{m} W {m} ¼ {m} {f} W {f} ¼ {f} {m, f} W {m, f} ¼ {m, f} {m} W {f} ¼ {m, f} {m} W {m, f} ¼ {m, f} {f} W {m, f} ¼ {m, f}
Dalrymple and Kaplan (2000) note that Slovene resolution poses a problem for this approach. Exactly as in Serbo-Croatian as described above, coordinated neuter NPs
‘‘Elsewhere’’ in Gender Resolution
571
yield masculine plural agreement: neut & neut ¼ masc. Since set union is idempotent (for any set X, X W X ¼ X), any feature set assigned to neut will incorrectly yield neut and not masc agreement.1 Vincent and Bo¨rjars (2000) propose a modification of Dalrymple and Kaplan 2000, using set intersection instead of union. This too fails for Slovene and Serbo-Croatian, since intersection, like union, is idempotent (for any set X, X X X ¼ X). A modified, semantically motivated set intersection approach is proposed in section 24.6 below. 24.4
Syntactic versus Semantic Resolution
The grammatical feature computation mechanism described above ignores an important distinction drawn by Corbett (1991) between ‘‘syntactic’’ and ‘‘semantic’’ gender resolution: ‘‘Gender resolution by the semantic principle involves reference to the meaning of the conjoined elements even if this implies disregard for their gender. The syntactic principle operates according to the gender of the conjoined items irrespective of their meaning’’ (p. 269).2 Among Corbett’s examples of semantic gender resolution are those from Bantu languages. Bantu coordination resolution is based on meaning and not morphological Noun Class (NC)—even though agreement with simple, noncoordinate NPs is based on NC. Consider Luganda, as described by Givo´n (1970, 252). Conjuncts of the same NC in Luganda can be conjoined, yielding plural agreement of that class (Givo´n suggests that all Bantu languages allow this). Now consider agreement when the conjuncts di¤er in NC. In Luganda, human-denoting nouns typically fall into class 1/2 (i.e., class 1 in singular, class 2 in plural), but there are many exceptions.3 NC 1/2 is the resolution class for humans, while NC 7/8 is used for nonhumans. This applies irrespective of the NCs of the conjuncts (Corbett 1991). (18) Ek-kazi, aka-ana ne olu-sajja ba-alabwa. 5-fat.woman 12-small.child and 11-tall.man 2-were.seen ‘The fat woman, small child, and tall man were seen.’ (19) En-te, omu-su, eki-be ne ely-ato bi-alabwa. 9-cow 3-wildcat 7-jackal and 5-canoe 8-were.seen ‘The cow, the wildcat, the jackal, and the canoe were seen.’ Of course, with a normal noncoordinate subject the predicate must agree in NC: ekkazi ‘fat woman’ normally triggers class 5 agreement, and so on. But the semantic principle operates in coordination resolution. Turning to French, it is unclear a priori whether the masculine plural agreement in example (3)/(7) results from a syntactic rule (mixes of the two grammatical genders trigger masculine) or instead a semantic rule (NPs denoting mixed-sex groups
572
Stephen Wechsler
trigger masculine). Corbett (1991, 279) notes that inanimates follow the same rule as animates.4 (20) Le livre et le cahier sont neufs / *neuves. the book(m) and the notebook(m) are new.m.pl / *new.f.pl (21) La mise`re et la ruine sont de´sastreuses / *de´sastreux. the misery(f) and the ruin(f) are disastrous.f.pl / *disastrous.m.pl (22) Ce savoir et cette adresse sont merveilleux / *merveilleuses. this knowledge(m) and this skill(f) are marvellous.m.pl / *marvellous.f.pl From this Corbett reasonably concludes that French employs syntactic and not semantic resolution. However, there is some evidence to support a third possibility: that inanimate NPs are subject to syntactic resolution, while animate NPs are subject to semantic resolution. In particular, it appears to be true crosslinguistically that when grammatical and natural gender diverge, resolution depends on natural, not grammatical, gender—even though simple, noncoordinate agreement follows grammatical gender. For example, consider the French nouns sentinelle ‘sentry’ and personne ‘person’. Both nouns are grammatically feminine regardless of whether they refer to men or women (indeed, a sentinelle is typically, though not necessarily, assumed to be male). This is illustrated in (23). (In all examples assume the referent of sentinelle or personne is male.) (23) a. La sentinelle / personne a` la barbe a e´te´ prise (f.sg) / *pris (m) en otage. ‘The sentry / bearded person was taken hostage.’ b. Les sentinelles ont e´te´ prises (f.pl) / *pris (m) en otage. ‘The sentries were taken hostage.’ Now consider coordinate NPs. Recall first of all that masculine plural is the resolution value for mixed gender subjects, as in (24a). But crucially this rule is based on semantic gender, namely, sex, and not grammatical gender, as shown in (24b,c).5 (24) a. Suzanne et Pierre ont e´te´ pris (m) / *prises (f.pl) en otage. ‘Suzanne and Pierre were taken hostage.’ b. La sentinelle et sa femme ont e´te´ pris (m) / *prises (f.pl) en otage. ‘The sentry and his wife were taken hostage.’ c. La sentinelle et la personne a` la barbe ont e´te´ pris (m) / *prises (f.pl) en otage. ‘The sentry and the bearded person were taken hostage.’ In (24b,c) we have two grammatically feminine NP conjuncts, yet the participle appears in masculine. Reversing conjunct order has no e¤ect on agreement. Clearly sex, not grammatical gender, is determining agreement. In (24b) the sexes are mixed,
‘‘Elsewhere’’ in Gender Resolution
573
yielding the masculine resolution form, while in (24c) both conjuncts are male denoting, again yielding masculine plural. This result for (24b) is summarized here. (25) Example (24b) Type of resolution rule Syntactic: feminine þ Semantic: male þ
feminine ) female )
Predicted form feminine plural masculine plural
Correct? * p
Similar facts obtain in Icelandic. As shown in (26), the noun ska´ld ‘poet’ is grammatically neuter, regardless of the sex of the poet. But when this noun refers to a male poet and is coordinated with another male-denoting noun, the predicate is not neuter plural—the form used for mixed gender subjects, as we saw above. Instead it appears in masculine plural (27). (26) Ska´ldi er frægt / ??frægur. the.poet(nt) is famous.nt.sg / ??m.sg ‘The poet is famous.’ (27) Ska´ldi og Jo´n eru frægir / *fræg. the.poet(nt) and Jon are famous.m.pl / *nt.pl ‘The poet and Jon are famous.’ (assume the poet is male) (28) Type of resolution rule Syntactic: neuter þ masculine Semantic: male þ male
) )
Predicted form neuter plural masculine plural
Correct? * p
In Serbo-Croatian the diminutive devojce ‘(little) girl’ is grammatically neuter while devojka ‘girl’ is grammatically feminine. In (30), summarized in (29), we see that resolution based on sex is preferred to resolution based on grammatical gender. (29) Type of resolution rule Syntactic: feminine þ Semantic: female þ
neuter female
) )
Predicted form masculine plural feminine plural
Correct? * p
(30) Ova velika devojka i ovo malo devojce su se this-f.sg big girl(f.sg) and this-nt.sg little-nt.sg girl(nt.sg) aux.pl refl lepo igrale / ?igrali. well played-f.pl / played-m.pl ‘This big girl(f) and this little girl(nt) played well.’ Recall from above that unlike conjuncts (indeed, anything but a group of feminine conjuncts) yield masculine plural. Although (30) contains a mix of genders (feminine and neuter), the preferred form for the predicate is nevertheless feminine plural. Summarizing, gender resolution for animates proceeds by consulting the meaning (‘‘natural gender,’’ i.e., sex) of the coordinate NP and not the form (‘‘grammatical gender’’) of the conjunct NPs.
574
24.5
Stephen Wechsler
Animates and Genders with Semantic Correlates
Grammatical genders often have semantic correlates: masculine gender correlates with male sex and feminine gender with female sex. What is the exact nature of this ‘‘correlation?’’ A careful answer to this question will e¤ectively dissolve the problem of gender resolution in coordinate structures. Consider how an NP receives its gender feature. Typically an NP inherits its gender from the lexical gender feature of its head noun: the NPs la table and la sentinelle a` la barbe have feminine gender because the underlined head noun is inherently specified in the lexicon as feminine. The grammatical mechanism responsible for the lexical gender feature on simple (noncoordinate) NPs like la personne avec la barbe is the Head Feature Principle or its X-bar theoretic equivalent: ‘‘head features’’ are shared between the head word and the phrase it heads. Gender is a head feature; the head noun (stem) personne bears the feature [gend fem]; hence the NP itself bears this feature. We will call this the ‘‘inherent (grammatical) gender’’ of the nominal. This source of grammatical gender is lacking in either of two situations: when the head noun is lexically genderless or where the phrase lacks a head noun altogether. As an example of the former, consider proper names (31) and certain sex-neutral nouns, such as French journaliste (32) and Serbo-Croatian sudija ‘judge’ (33). (31) a. Le (m.sg) Professeur Dupont est beau (m.sg). ‘Professor Dupont (a man) is handsome.’ b. La (f.sg) Professeur Dupont est belle (f.sg). ‘Professor Dupont (a woman) is beautiful.’ (32) a. Le (m.sg) journaliste est compe´tent (m.sg). ‘The (male) journalist is competent.’ b. La (f.sg) journaliste est compe´tente (f.sg). ‘The (female) journalist is competent.’ (33) a.
Taj stari sudija je dobro sudio. that.m old.m judge aux well judged.m ‘That old (male) judge judged well.’ b. Ta stara sudija je dobro sudila. that.f old.f judge aux well judged.f ‘That old (female) judge judged well.’ c. *To mlado sudija je dobro sudilo. that.n young.n judge aux well judged.n ‘That young judge judged well.’
Neuter is impossible in (33c), a matter we return to below.
‘‘Elsewhere’’ in Gender Resolution
575
The following principle is operating. (34) Gender agreement with an animate NP that lacks inherent gender is always interpreted semantically. Professeur Dupont, journaliste, and sudija lack inherent gender. So by (34), such nouns denote males or females depending on whether they trigger masculine or feminine agreement. Note that this generalization applies only to NPs lacking inherent grammatical gender. Semantic interpretation of gender is blocked by inherent gender, for example, in la sentinelle. Next consider plurals. (35) a. Les Dupont sont belles (f.pl) / beaux (m.pl). ‘The Duponts (all women / a male or mixed-sex group) are beautiful.’ b. Les journalistes sont compe´tentes (f.pl) / compe´tents (m.pl). ‘The (female / male or mixed-sex) journalists are competent.’ Feminine plural is used for a group of females, while masculine (plural) agreement indicates a group of males or a mixed-sex group. Pronouns (36) and sexdi¤erentiated common nouns (37) follow the same pattern. (36) Ils (m.pl) / Elles (f.pl) parlent dans la cuisine. ‘They (a male or mixed-sex group / a female group) are talking in the kitchen.’ (37) a. b. c. d.
un Americain (m.sg): male American une Americaine (f.sg): female American les Americains (m.pl): male or mixed sex Americans les Americaines (f.pl): female Americans
According to a long-noted crosslinguistic generalization (Corbett 1991, 292, citing Greenberg 1966 and Schane 1970; Farkas and Zec 1995), the gender used for mixed-sex groups in a given language is the same for coordinate structures (Pierre et Marie), non-sex-di¤erentiated plural proper nouns (les Dupont), non-sexdi¤erentiated plural common nouns (les journalistes), sex-di¤erentiated plural common nouns (les Americains), and plural pronouns (ils). The other languages discussed in this chapter show the same generalization. A Serbo-Croatian masculine plural pronoun (oni) or sex-di¤erentiated common noun (e.g., Amerikanci) refers to a male or mixed-sex group, while the feminine forms (one, Amerikanke) refer to a group of females, paralleling the coordinate structure resolution rule above. In Icelandic, neuter plural is used for mixed-sex coordinate structures, as well as pronouns ( au ‘they.nt’, a mixed sex group). The French and Serbo-Croatian facts above follow if the gender features have the semantic interpretations in (38) below. Serbo-Croatian neuter lacks a semantic
576
Stephen Wechsler
correlate, as shown in (33c). So we distinguish two types of genders, those with semantic correlates, such as masculine and feminine (hereafter ‘‘s-genders’’ for semantic-gender), and those lacking a semantic correlate, such as neuter in some languages (‘‘e-genders’’ for empty or expletive gender). (38) Semantic interpretations of French or Serbo-Croatian s-genders (applying only to NPs lacking inherent gender; see (34)) a. Feminine: ‘‘female’’ b. Masculine: ‘‘nonfemale’’ Sex is a distributive property: a group is female if and only if all of its members are female and male if and only if all of its members are male. As a rule the set of distributive properties of a group is just the intersection of the sets of distributive properties of each of the group’s members. Hence for a singular NP the semantic feature ‘‘female’’ in (38a) entails (or perhaps presupposes) that the NP refers to a female, while for a plural NP the semantic feature ‘‘female’’ means that the NP refers to a group of females. However, the negatively defined semantic feature ‘‘nonfemale’’ is not distributive (since negation itself is not distributive): a ‘‘nonfemale’’ group is a group that fails to meet the description of a ‘‘female’’ group (namely a group of females). Thus any group containing at least one male is a ‘‘nonfemale’’ group.6 Returning now to (animate) coordinate structures, we find that the problem of resolution has vanished. A coordinate NP lacks a head noun.7 Hence it lacks an inherent lexical gender feature, so by (34) an agreeing predicate assigns semantic gender. Mixed-sex coordinate NPs trigger masculine gender agreement (in, e.g., French) because masculine has the appropriate meaning. 24.6
The Origin of Inanimate Resolution Rules
Let us assume that the inanimate resolution rules are grammaticalizations of the logic of semantic combination just discussed. As just noted, the set of distributive properties of a group is just the intersection of the sets of distributive properties of each of the group’s members. The same logic applies to (privative) distributive morphological features. The set of privative features associated with a group is just the intersection of the sets associated with each member of the group. Corresponding to the semantic values ‘‘female’’ and ‘‘nonfemale’’ (see (38)), each positively specified gender is represented as a singleton set such as {f}, and each negatively specified gender such as nonfeminine (i.e., masculine) as a null set { } (cp. Vincent and Bo¨rjars 2000). E-genders like neuter can be either singleton or null sets: Serbo-Croatian neuter is {n} while Icelandic neuter is { }, as we will see below. Since they lack semantic correlates, e-genders are exempt from the logic of distributivity:
‘‘Elsewhere’’ in Gender Resolution
577
there is no sense in which a ‘‘group of neuter items’’ is a ‘‘neuter group.’’ So e-gender features will be systematically removed from the computation of gender features for group-denoting NPs. To remove any e-genders from the calculated intersection set, we intersect it with the set Gs of all s-gender features available in the language. In the following proposed universal rule for the gender value of aggregate discourse referents, the biconditional A,B means that any representation satisfying description A must also satisfy B and vice versa. (39) Rule for deriving gender of inanimate aggregate discourse referents D.R.[{[GEND g1 ], . . . [GEND gn ]}] , D.R.[GEND g1 X, . . . X gn X Gs ] where g1 . . . gn are null or singleton sets and Gs is the set of s-gender features in the grammar. According to (39) the gend(er) value for an aggregate discourse referent is just the intersection of the s-gender values of the discourse referent’s elements. In a coordinate structure those elements correspond to the conjuncts; see below. This solves the problem plaguing the set-based accounts of Dalrymple and Kaplan 2000 and Vincent and Bo¨rjars 2000 (see section 24.3) and correctly accounts for the SerboCroatian/Slovene type of language. The system is illustrated with French, SerboCroatian, and Icelandic below. The semantic correlate of each s-gender is shown in parentheses. (40) French: Gs ¼ {F} fem: {f} (< ‘‘female’’) masc: { } (< ‘‘nonfemale’’) masc & masc ¼ masc fem & fem ¼ fem masc & fem ¼ masc
{ } X { } X gs ¼ { } {f} X {f} X gs ¼ {f} { } X {f} X gs ¼ { }
(41) Serbo-Croatian: Gs ¼ {F} fem: {f} (< ‘‘female’’) masc: { } (< ‘‘nonfemale’’) neut: {n} (e-gender) masc & masc ¼ masc fem & fem ¼ fem neut & neut ¼ masc masc & fem ¼ masc masc & neut ¼ masc fem & neut ¼ masc
{ } X { } X gs ¼ { } {f} X {f} X gs ¼ {f} {n} X {n} X gs ¼ { } { } X {f} X gs ¼ { } { } X {n} X gs ¼ { } {f} X {n} X gs ¼ { }
578
Stephen Wechsler
(42) Icelandic: Gs ¼ {M, F} masc: {m} (< ‘‘male’’) fem: {f} (< ‘‘female’’) neut: { } (e-gender) masc & masc ¼ masc fem & fem ¼ fem neut & neut ¼ neut masc & fem ¼ neut masc & neut ¼ neut fem & neut ¼ neut
{m} X {m} X gs ¼ {m} {f} X {f} X gs ¼ {f} { } X { } X gs ¼ { } {m} X {f} X gs ¼ { } {m} X { } X gs ¼ { } {f} X { } X gs ¼ { }
While permitting the Serbo-Croatian/Slovene type, this theory also rules out certain unattested resolution patterns. The following pattern, like Serbo-Croatian or Slovene but with masculine and neuter genders reversed, cannot be generated (assuming neuter is an e-gender). (43) *Unattested: (M, F, NT) a. fem & fem ¼ fem b. Elsewhere (neut & neut; masc & masc; and mixes) ¼ neut For any singleton or null sets X and Y, if X0Y then X X Y ¼ { }; so the null set represents the resolution gender. Since the sets representing the genders must be distinct, there can be at most one null set in the gender paradigm. If neuter is represented as the null set, then masculine must be nonnull (e.g., {m})—but this would yield the Icelandic pattern above, not (43). This result does not follow from feature computation mechanisms that are not rooted in the semantics. Note that our rule (39) is not specific to coordinate structures but applies more generally to aggregate discourse referents. This is because the crosslinguistic generalization noted above—that pronouns, inter alia, invariably follow the same resolution pattern as coordinate structures in a given language—applies not only to animates but inanimates as well. For example, feminine plural is the resolution gender for Rumanian inanimates, both in coordinate structures (see below) and in pronouns; the latter is illustrated in this discourse. (44) Saˇ vopsim podeaua odataˇ cu plafonul. Vor fi amandouaˇ C paint.1pl floor.def[f] same.time with ceiling.def[m] will be both albe. white.f.pl ‘Let’s paint the floori when we paint the ceilingk . Theyiþk will both be white.’ The demise of coordinate structure resolution, in favor of a discourse referent rule, has a further consequence. On the common view that the relevant discourse referents are denotations of maximal extended projections of nominals (NP or DP, depending
‘‘Elsewhere’’ in Gender Resolution
579
on one’s analysis) but that subconstituents of nominals are not referential, it follows that gender resolution should not be possible when coordinating smaller subconstituents of nominals. In fact it has been observed for NP-internal concord that resolution cannot apply and only partial agreement is possible (Serbo-Croatian; Corbett 1983, 209–210). (45) najsvirepije / *najsverepiji [kazne i mucenja] cruellest.f.pl / cruellest.m.pl punishments(f.pl) and tortures(nt.pl) ‘the cruellest punishments and tortures’ Only partial agreement with the nearer conjunct is possible. 24.7
Origins of the Featural Representations of Grammatical Genders
Obviously the featural definitions of the genders—whether they are specified as singleton or null sets—crucially a¤ect the results. Where do these feature assignments come from?8 In most languages the grammatical features directly correspond to semantic features of the language. From the French semantic features ‘‘female’’ @ ‘‘nonfemale’’, we get the grammatical features {f} @ { }, and so on, mutatis mutandis, for the other languages.9 E-genders can be of either type: Serbo-Croatian neuter is a singleton set ({n}) while Icelandic neuter is null ({ }). The tendency toward parallelism between semantic and syntactic resolution patterns can be explained as a consequence of ‘‘abductive inference’’ (also called ‘‘hypothetical inference’’ or ‘‘abduction’’).10 The philosopher C. S. Peirce explained abduction as tentatively accepting the results of reversing modus ponens.11 If m were true, p, p 0 , p 00 would follow as miscellaneous consequences. But p, p 0 , p 00 are in fact true. 8 Provisionally, we may suppose that m is true. This kind of reasoning is often called adopting a hypothesis for the sake of explanation of known facts. (Peirce 1992)
Grammar acquisition is the development of explanations for the ‘‘known facts’’ (p, p 0 , p 00 ) about the input corpus. Grammatical gender features hypothesized (‘‘abduced’’) on the basis of the French animate agreement facts would be {F} and { }, correctly yielding the pattern in (8) above: if all conjuncts are feminine, use feminine plural; elsewhere use masculine plural. The vast majority of conjoined animates confirm this hypothesis, the only counterexamples being the special cases discussed above, such as sentinelle and personne, which are feminine even when the referent is male. Hence this hypothesis is grammaticalized as the featural representation of the grammatical genders applying to inanimates. As invoked here, abductive inference is not part of the grammar but rather a functional mechanism predicting the likely development path for grammar. Other factors
580
Stephen Wechsler
could contravene the e¤ects of abductive inference, leading to a heterogenous system with di¤erent patterns for animate and inanimate agreement. For example, in Rumanian, mixed sex animates trigger masculine plural, while inanimate gender mixes resolve to the feminine plural (Corbett 1991, 288–289; Farkas and Zec 1995; Moosally 1998). The pattern for animates is identical to that of French, SerboCroatian, and so on (examples are from Moosally 1998). (46) Maria si tata au fost vaˇzuti fumıˆnd marijuana. Maria and father were seen.m.pl smoking marijuana ‘Maria and father were seen smoking marijuana.’ Pronouns and sex-di¤erentiated common nouns pattern as expected, with the masculine plural ei ‘they.m.pl’ and, for instance, scriitori ‘male writers’ (rather than scriitoare ‘female writers’) used for mixed sex groups. Rumanian nouns like santinelaˇ ‘sentry’, which is grammatically feminine, behave as in the languages described above (section 24.4): sex rather than gender determines agreement in coordinate structures, as first noted by Farkas and Zec (1995). (47) Maria si santinela au fost caˇsaˇtoriti de catre preotul local. Maria and sentry.def pst.pl were married.m.pl by priest.def local ‘Maria and the sentry were married by the local priest.’ So far Rumanian is just like French. Following the analysis above, we posit that feminine/masculine agreement indicate ‘‘female’’/‘‘nonfemale’’ semantics, a meaning arising wherever grammatical gender does not preempt it. Unlike animates, which resolve to masculine plural, inanimates resolve to feminine plural (Farkas and Zec 1995, 96). (48) MASC & MASC ) MASC.PL a. Nucul si prunul sunt uscati. walnut.def[m] and plum.def[m] are dry.m.pl ‘The walnut tree and the plum tree are dry.’ All other combinations of MASC, FEM, or NEUT ) FEM.PL b. Podeaua si plafonul sunt albe. floor.def[f] and ceiling.def[m] are white.f.pl ‘The floor and the ceiling are white.’ c. Scaunul si dulapul sunt albe. chair.def[n] and cupboard.def[n] are white.f.pl If all conjuncts are masculine we get masculine plural agreement (48a). All other combinations of masculine, feminine, or neuter yield feminine plural agreement (48b,c). This shows clearly that the resolution mechanism for inanimates is independent of the determinants of animate gender agreement.
‘‘Elsewhere’’ in Gender Resolution
581
Why is Rumanian resolution asymmetrical? A clue is found in the morphological gender system of that language. The so-called neuter gender is really a class of inquorate12 nouns that are masculine in the singular but feminine in the plural (Bley-Vroman 1977; Farkas and Zec 1995).13 Thus for plurals, feminine is the gender with the broadest distribution—the unmarked or ‘‘elsewhere’’ gender—and hence is used for resolution.14 Further evidence for the separation between animate and inanimate resolution is an asymmetry in Serbo-Croatian noted by Corbett (1983, 1991). He reports many textual exceptions to the resolution rule involving overapplication of the masculine plural default to coordinate structures consisting entirely of feminine conjuncts, as in this example (from Corbett 1991, 302):15 (49) Stula i staka bili su sve sto je tadasnja medicina wooden.leg(f) and crutch(f) been.m.pl are all that aux of.that.time medicine mogla da mu pruzi. could that him.dat o¤er ‘A wooden leg and a crutch were all that medical science of that time could o¤er him.’ Interestingly, Corbett notes that ‘‘I have found no examples of masculine agreement with feminine nouns denoting persons’’ (Corbett 1991, 302). Thus we do not find violations of the semantic content of genders, while the weaker, derivative, autonomous resolution rule for inanimates is occasionally violated. Summarizing, the principles governing resolution follow directly from the logic of distributivity. Denotations like ‘‘female’’ correspond in the morphosyntactic domain to privative morphological features, while negatively defined denotations like ‘‘nonfemale’’ correspond to the lack of a privative morphological feature, represented by the null set. Semantic and grammatical feature representations, while shown to be autonomous, tend to be parallel due to the e¤ects of abductive inference. 24.8
How Inherent Gender Blocks Semantic Gender
A crucial aspect of this analysis is that gender features have semantic content only where the agreement trigger lacks inherent gender. This blocking e¤ect can be captured formally in Lexical-Functional Grammar (LFG). Unification-based formalisms such as LFG model agreement as a correlation arising because features of a single grammatical object, namely the functional structure or f-structure, are specified by two distinct elements in the sentence. This specification occurs via equations of two types: ‘‘defining equations,’’ which build the f-structure, and ‘‘constraining equations,’’ notated with ¼c , which check the f-structure for the presence of a
582
Stephen Wechsler
feature. Suppose that gender agreement morphology encodes a disjunction between two equations, a defining equation for the semantic gender (i.e., sex) of the agreement trigger, and a constraining equation for its semantically null grammatical gender, as in the following example. (50) a. compe´tente, A
b. compe´tent, A
("pred) ¼ ‘competenth("subj)i’ ("subj gend) ¼c {f}4("subj gend) ¼ ‘‘female’’ ("subj num) ¼ sg ("pred) ¼ ‘competenth("subj)i’ ("subj gend) ¼c { }4("subj gend) ¼ ‘nonfemale’’ ("subj num) ¼ sg
Values in quotation marks, here ‘‘female’’ and ‘‘nonfemale,’’ are semantic values, which add to the semantic form the presupposition of reference to a female/ nonfemale. The values {f} and { }, on the other hand, are semantically vacuous. The lexical entries for nouns like table and personne have the semantically vacuous feminine gender (represented by the set {f}), while Dupont and journaliste are unspecified for gender. (51) a. table ‘table’, personne ‘person’, sentinelle ‘sentry’, and so on: ("gend) ¼ {f} b. crayon ‘pencil’, and so on: ("gend) ¼ {m} c. Dupont (name), journaliste ‘journalist’, and so on: no gend equation In sentence (52a) the variant of compe´tente specifying a ‘‘female’’ subject would conflict with the {f} gender of personne, so the disjunct specifying {f} must be selected instead. Thus the adjective in (52a) is silent regarding the sex of the subject referent. Conversely, the variant of compe´tente selecting a {f} subject is unavailable in (52b). The constraining equation ("subj gend) ¼c {f} would check the f-structure for the feature {f} and fail to find it, since Dupont lacks inherent gender (see 51c). Only the value ‘‘female’’ is available, so Dupont must refer to a female. (52) a. [La personne a` la barbe][gend {f}] est compe´tente[subj gend ¼c {f}] . ‘The bearded person is competent.’ b. Dupont est compe´tente[subj gend ¼ ‘female’] . ‘Dupont (a female) is competent.’ In sum, the gender feature marked on agreement targets has two possible functions: it can check for the presence of a semantically vacuous grammatical agreement feature inherent in the trigger NP or it can have semantic content. Being exocentric, coordinate structures lack inherent gender, so they generally trigger semantic agreement whenever possible (e.g., when they denote animates). The special ‘‘rule for deriving gender of inanimate aggregate discourse referents’’ in (39) is a stopgap for those cases where semantic agreement fails but the rest of the gram-
‘‘Elsewhere’’ in Gender Resolution
583
mar also fails to provide an inherent grammatical gender value. Next we consider why some languages lack such a stopgap. 24.9
Syntactic and Semantic Gender Resolution Revisited
According to Corbett (1991), languages of the world can be classified according to whether they employ syntactic resolution, semantic resolution, or some combination of the two. The present claim is that all languages employ semantic resolution. When the domain of semantic classification (male/female) extends only to sexed items, the discourse referent rule (39) covers the residue outside that domain. On the present assumptions, a pure semantic resolution language is one where the domain of semantic classification exhausts the entire universe of concepts so that there is no residue for syntactic resolution to cover. Recall the Luganda examples (18) and (19) above, where classes 2 and 8 are ‘‘resolution classes’’ for humans and nonhumans, respectively. Following our earlier approach we hypothesize these semantic genders for Luganda. (53) Luganda (NC 1/2, 3/4, 5/6, 7/8, 9/10, etc.) a. NC 1/2: ‘‘human’’ b. NC 7/8: elsewhere (‘‘nonhuman’’) Unlike systems based on sex classification, the Luganda system has human (NC 1/2) and nonhuman (NC 7/8) classes that exhaustively partition the concepts of the world. This means that in Luganda there is no residue to be covered by syntactic rules.16 The Luganda type system is also found in Dzamba, Likila, Lingala, and Swahili (Bokamba 1985, cited in Corbett 1991, 275–276).17 Corbett notes that while the Bantu systems vary in the details and there are descriptive complications (partial agreement, homophony between conjunction and comitative marker), it is nevertheless valid that ‘‘In almost all the Bantu languages investigated we find evidence for semantic resolution based on the human/non-human distinction’’ (Corbett 1991, 276). 24.10
Competition between Semantic and Syntactic Resolution
In languages employing both principles, semantic gender assignment generally blocks the syntactic gender resolution, as we saw from sex-gender mismatches like sentinelle. However, this blocking e¤ect is not always complete. The syntactic resolution system can sometimes compete with the semantic interpretations of genders. The first example comes from Serbo-Croatian. We saw in (30) above that feminine plural agreement is preferred for a coordinate NP denoting a group of females, even though the conjuncts are not all morphologically feminine. However, feminine agreement appears to require at least one morphologically feminine conjunct.
584
Stephen Wechsler
(54) Ovo moje devojce i ovo tvoje luce su se this.nt.sg my.nt.sg girl.nt.sg and this.nt.sg your.nt.sg doll.nt.sg aux.pl refl lepo *igrale/igrali. well played.f.pl/played.m.pl ‘My girl(nt) and your ‘‘doll’’ (¼ girl, nt) played well.’ With no morphologically feminine conjuncts at all, feminine agreement becomes difficult or impossible. Similar facts are reported for Ndebele (Moosally 1998). The 1/2 plural animate resolution class can apply only if at least one of the conjuncts belongs to class 1/2 (data from Moosally 1998, 91–92). (55) a.
Um-angoye le-nja ba-zazwan-an-a. 1a-cat conj.5-dog 1/2pl-friends-recip-past ‘The cat and the dog were friends.’ b. *U-sane le-nkazana ba-hle. 11-baby conj.5-girl 2-pretty ‘The baby and the girl are pretty.’
Similarly, some French speakers report uncertainty or disagreement about the judgments in (24), suggesting that the morphological rule is competing with the semantic one. 24.11
Conclusion
The fact that gender need not have a direct semantic basis—la table is feminine, but does not denote a female—does not mean that semantics can be safely ignored. It was argued that gender morphology on agreement targets (modifiers, predicates, determiners, etc.) indicates either semantic class or semantically vacuous grammatical agreement. Semantically vacuous grammatical agreement arises whenever the NP trigger is marked for the agreement feature due to the NP’s internal morphosyntactic structure—specifically, when the NP contains a head noun that is lexically marked for gender. On the other hand, where the NP lacks an inherent grammatical gender feature, target gender features have semantic content—if the nominal lies within the domain of semantic classification. If not, then grammatical resolution arises. Semantic and grammatical resolution are autonomous processes, but both are governed by the same logic of distributivity, and they tend to be parallel due to abductive inference. Summarizing, we find the following hierarchy of defaults. (56) Gender assignment hierarchy a. Inherent grammatical gender (if NP has a head with a gender value). Else: b. Semantic gender (if NP’s denotation is within the domain of the semantic classification system). Else: c. Rule for deriving gender of inanimate aggregate discourse referents.
‘‘Elsewhere’’ in Gender Resolution
585
‘‘Elsewhere’’ enters the picture in two ways. First, the traditional notion of an unmarked/default/elsewhere gender must be preserved to capture the range of attested phenomena. Second, the application of the three sources of gender summarized in (56) is governed by a hierarchy of defaults. Notes For help with the data, my thanks go to Larisa Zlatic´ (Serbo-Croatian), Jo´hannes Gı´sli Jo´nsson (Icelandic), Alexandra Teodorescu and Ileana Comorovski (Rumanian), and Pascal Denis and Knud Lambrecht (French). Thanks also go to Mary Dalrymple for valuable discussion and to Kristin Hanson and an anonymous reviewer for comments on an earlier draft. The following abbreviations follow throughout the text: m.sg ¼ masculine singular; m.pl ¼ masculine plural; f.sg ¼ feminine singular; f.pl ¼ feminine plural; masc ¼ masculine; fem ¼ feminine; neut ¼ neuter; nt ¼ neuter; gend ¼ gender; g ¼ gender; f ¼ feminine; n ¼ noun; nc ¼ noun class; dat ¼ dative; aux ¼ auxiliary; conj ¼ conjunction; 1.pl ¼ first person plural; pred ¼ predicate; subj ¼ subject; num ¼ number; def ¼ definite; recip ¼ reciprocal; refl ¼ reflexive. 1. Dalrymple and Kaplan 2000 outline two possible solutions for Slovene: either agreeing elements distinguish between coordinate and noncoordinate neuter plural NPs or the conjunction word itself could contribute a crucial feature. If we are correct in concluding (section 24.8) that languages lack resolution rules specific to coordinate structures, then neither solution is available. 2. The observations in this section are inspired by those of Farkas and Zec 1995 on Rumanian. 3. Following the Bantuist tradition, a noun is lexically associated with a representation of the form x=y, where x is the noun’s singular NC and y is its plural NC. 4. Some speakers prefer to reverse the order of conjuncts in (22), so that the masculine is closer to the agreement target. But feminine agreement in (22) is impossible, so this is not partial agreement. 5. There is cross-speaker variation in some of these judgments. See Section 24.10 below. The test in (24) is based on Farkas and Zec 1995. 6. A variant of this account would replace ‘‘nonfemale’’ in (38b) with ‘‘animate.’’ A general blocking principle would ensure that the most informative alternant in a paradigm is selected over less informative ones. For females, feminine would block masculine since ‘‘female’’ is more informative than ‘‘animate’’; other animates would be masculine. Blocking principles of this kind have been proposed by the structuralists (Roman Jakobson, inter alios) in studies of pragmatics (cp. Grice’s Maxim of Quantity) and in various forms within the generative paradigm, including the faithfulness constraints of Optimality Theory. For a recent application to agreement morphology, see Blevins 2000. 7. Or perhaps the conjunction is the (functional) head (Johannessen 1996); or coordinate NPs are multiply headed by the conjunct NPs (Farkas and Ojeda 1983). In any case, the important point is that they lack the usual clear projection line from a unique N 0 head to the NP. Partial agreement may be analyzed as a special case where head features are inherited from one conjunct or where one conjunct is the specifier (Johannessen 1996).
586
Stephen Wechsler
8. This is a question that Dalrymple and Kaplan (2000) do not ask. Their set-theoretic representations for the genders are stipulated. 9. An interesting exception to this parallelism is Rumanian, discussed below. 10. Peirce was an early theorist of abduction; more recent applications to linguistics include Strigin 1999 and Hobbs et al. 1993. For a related argument see Corbett 1983, 202. 11. Modus ponens states that from p ! q and p, we deduce q. Abduction assumes p ! q and q and tentatively concludes p. 12. The term ‘‘inquorate,’’ characterizing nouns belonging to di¤erent genders in singular and plural, comes from the notion that there are too few members to make a quorum (Corbett 1991, 1999). 13. Neuter nouns have some declensions not found in masculine or feminine (Mallinson 1986) but trigger masculine when singular, feminine when plural. 14. Farkas and Zec (1995) elegantly analyze the three Rumanian genders with a single feature [F]: masculine nouns are [F], feminine nouns are [þF], and neuter nouns are unmarked for [F]. Defaults fill in unspecified [F] values as minus in the singular and plus in the plural. This unifies the treatment of coordinates and neuters. 15. These cannot be due to the failure of agreement altogether. The Serbo-Croatian ‘‘neutral gender’’—the form of a verb when there is no nominative NP subject at all—is neuter (singular) and not masculine. 16. In Luganda, as in many Bantu languages, mixing human and nonhuman conjuncts is strongly dispreferred; a comitative construction is used instead. If it is forced then the form selected is NC 2. See Moosally 1998 for description and discussion. 17. In Swahili the crucial semantic feature is animacy rather than humanness.
25
The Force of Lexical Case: German and Icelandic Compared
Dieter Wunderlich
25.1
Introduction
Modern German (G) and Modern Icelandic (I) resemble each other in many details of their Case-marking systems. This chapter focuses on some di¤erences in control and raising-to-object constructions in the two languages, using Dative-Nominative (experiencer) verbs as the test case. German favors the argument in the Nominative as the syntactic pivot, while Icelandic favors the highest argument as the syntactic pivot. It is shown that the classical notion of subject fails to deal with the data. The chapter therefore develops a lexical account, implemented in the framework of Optimality Theory (OT; Prince and Smolensky 1993). 25.2
Case, Control and Raising-to-Object in German and Icelandic
German and Icelandic have similar Case systems. All Case patterns found in German are also attested in Icelandic (though Icelandic exhibits a few additional Case patterns not considered here). The two languages also have similar classes of lexically marked verbs (often referred to as verbs with quirky Case arguments). The examples in (1) illustrate a nom-dat verb, which deviates from the canonical transitive nomacc pattern. The Dative in this instance is assumed to be lexically marked. (1) a. G Er half mir. b. I Hann hja´lpaDi me´r. 3msg.nom helped.3sg 1sg.dat ‘He helped me.’ German and Icelandic furthermore share the property that the finite verb exclusively agrees with Nominative arguments. The examples in (2) illustrate that lexically marked Case is preserved in the passive and that default agreement features (3sg, henceforth u) appear on the auxiliary if the only argument remaining in the passive is non-Nominative.
588
Dieter Wunderlich
(2) a. G Mir wurde geholfen. b. I Me´r var hja´lpaD. 1sg.dat aux.3sg help.part ‘I was helped.’ However, if the clauses in (2) are embedded into certain syntactic contexts, German and Icelandic behave di¤erently. For instance, a clause with a lexically marked argument can be embedded under a control verb in Icelandic but not in German, as shown in (3). (3) Control a. G *Ich ho¤e geholfen zu werden. 1sg hope help.part to aux b. I E´g vonast til aD verDa hja´lpaD. 1sg hope for to aux help.part ‘I hope to be helped.’ The possibility of being controlled is one of the subject diagnostics used in the literature. On the basis of a number of tests, quirky Case arguments of Icelandic have been shown to be potential subjects, while those of German are not (Andrews 1982; Zaenen, Maling, and Thra´insson 1985). In other words, the Dative argument in (2b) is a subject, while the Dative argument in (2a) is not. The pattern in (3) then follows from the assumption that only subjects can be controlled. An alternative way to capture this distribution is the assumption that visibility of lexical Case—that is, the requirement that any lexically assigned (as opposed to structurally assigned) Case must be overtly expressed in each clause—ranks relatively high in German but relatively low in Icelandic.1 (4) MAX(lexF) Lexical Case features are visible in the output. (Visibility) Another phenomenon with respect to which German and Icelandic di¤er is argument deletion in coordination. In (5), an intransitive verb (with Nominative subject) is coordinated with an inverted dat-nom verb (where the higher argument is lexically marked for Dative, and the lower one is realized by Nominative). The DP occurring in the first position of the second clause can be deleted if it is coreferential with the corresponding DP of the first clause (in boldface). As predicted by the respective ranking of Max(lexF), Icelandic is able, but German is unable, to delete the lexically marked argument. (5) Coordination a. G Er sagt von sich fleißig zu sein, 3sg.nom claims of himself diligent to be *aber ihm ka¨me die Aufgabe zu schwer vor. but 3sg.dat find.conj the task too hard pt
The Force of Lexical Case
589
b. I Hann segist vera duglegur, 3sg.nom claims-self be diligent of ungt. en honum finnst verkefniD but 3sg.dat finds the homework too hard ‘He claims to be diligent but finds the homework too hard.’ If, in contrast, the two Nominative arguments are coreferential, German allows the lower argument to move to the preverbal position, where it can be deleted if it is identical in Case, as in (6a), whereas Icelandic does not allow the nom-argument in a preverbal position (6b). (6) a. G Er sagt von sich fleißig zu sein, 3sg.nom claims of himself diligent to be aber er kommt mir doch eher faul vor. 1sg.dat but rather lazy pt but 3sg.nom find cf. *aber mir kommt er doch eher faul vor. but 1sg.dat find 3sg.nom but rather lazy pt b. I Hann segist vera duglegur, 3sg.nom claims-self be diligent *en me´r finnst hann latur. but 1sg.dat find 3sg.nom lazy cf. *en hann finnst me´r latur. but 3sg.nom find 1sg.dat lazy ‘He claims to be diligent, but I find him (rather) lazy.’ To capture this further di¤erence between German and Icelandic, it is necessary to set out some assumptions about the order in which arguments appear. Let us assume that the unmarked order of arguments reflects their hierarchical semantic ranking.2 (7) SEMHIER The linear order of arguments corresponds to their semantic ranking (with the highest argument leftmost). How do we know the semantic ranking of arguments? Several tests have been proposed in the literature that are sensitive to semantic ranking, among them the Barss and Lasnik (1986) tests, which comprise anaphoric binding, bound pronouns, weak crossover, multiple questions, and the occurrence of negative polarity items. Not all of these tests are decisive in the languages under question (see Frey 1993 for German), but there is good reason to believe that the semantic ranking of arguments can be established for all verbs in every language. In the following, I will assume that for each individual verb the semantic ranking of arguments is known and that,
590
Dieter Wunderlich
consequently, their basic linear order is determined. Every movement of arguments gives rise to a violation of SemHier. The patterns in (5) and (6) suggest the following rankings: Max(lexF) SemHier for German but SemHier Max(lexF) for Icelandic. If Max(lexF) ranks high, ignoring lexical Case leads to ungrammaticality as in (5a), but if it ranks low, ignoring lexical Case can be tolerated, as in (5b). An additional di¤erence between German and Icelandic emerges when the inverted dat-nom verbs, illustrated in the second clauses of (5) and (6), are embedded under a control verb. As seen in (8), German allows control of only the lower, Nominative, argument (8a), whereas Icelandic allows control of only the higher, Dative, argument (8b). Therefore, (8a) and (8b) have di¤erent readings. (8) Control a. G Ich ho¤e, ihm — nom zu gefallen. 1sg hope 3msg.dat — nom to like ‘I hope to please him/to be liked by him.’ b. I E´g vonast til aD — dat lı´ka ??hann/?essi bo´k. 1sg hope for to — dat like 3msg.nom/this book.nom ‘I hope to like him/this book.’ If, however, these verbs are embedded into a raising-to-object construction, German requires the lower argument to raise and receive Accusative Case from the matrix verb (though some speakers find (9a) only marginally acceptable). In contrast, Icelandic allows only for raising of the higher argument, even if it is Dative (and remains so in the passive), whereas the lower argument obligatorily stays in the Nominative; see (9b,c) and the passive in (9d). (9) Raising to object a. G ?Ich sah ihm den Film gefallen. 1sg saw 3msg.dat the.acc film like ‘I saw him like the film.’ b. I E´g tel henni lı´ka thessi bo´k. 1sg believe 3fsg.dat like this book.nom ‘I believe her to like this book.’ ´ lafur leiDinlegur. c. I E´g tel henni hafa alltaf o´tt O 1sg believe 3fsg.dat have always thought Olaf.nom boring.nom ‘I believe her always to have found Olaf boring.’ (Zaenen et al. 1985, 101) ´ lafur leiDinlegur. d. I Henni er taliD hafa alltaf o´tt O 3fsg.dat is believed.u have always thought Olaf.nom boring.nom ‘She is believed always to have found Olaf boring.’ These constructions pose several problems for a syntactic approach to control and raising:
The Force of Lexical Case
591
(i) Since in Icelandic the missing argument of a control construction can be the one marked for lexical Case, even in the presence of a Nominative argument, one has to assume that pro relates to the highest argument of the embedded verb. However, the German data show that pro may relate to a nonhighest argument of the embedded verb. (ii) The fact that an infinitive clause of Icelandic may contain a Nominative argument (although this is the argument that can agree with the finite verb) shows that assignment of Nominative is independent of the existence of agreement. (iii) The German data suggest that multiple raising is possible, because the Dative argument in (9a) cannot be considered to be part of the infinitive clause. The Exceptional-Case-Marking account (ECM), claiming that the ‘‘raised’’ argument gets structural Case from the matrix verb, does not work for the Icelandic examples in which the raised argument is lexically marked for Case. 25.3
A Lexical Analysis
In this chapter, I will pursue a lexical account of the data discussed in section 25.2, implemented in the framework of OT. My basic assumptions are the following: (i) The properties of the control and the raising-to-object constructions are fully determined by the lexical representations of the involved verbs, in combination with a set of general syntactic constraints. (ii) The di¤erences between German and Icelandic follow from di¤erent constraint rankings. For descriptive purposes, the main di¤erences between German and Icelandic in the constructions considered so far (control, coordination, and raising to object) can be captured by the following generalization (which is uncontroversial in the literature). (10) In German, the Nominative argument is subject, whereas in Icelandic, the semantically highest argument is subject. If one conceives of Universal Grammar (UG) as a set of unfixed parameters, (10) describes a di¤erent setting of parameters. However, there are certain facts that render such a parameter account problematic:3 (i) Why is the default order of arguments in German determined by semantic ranking and not by the choice of subject? (ii) On the other hand, why is agreement in Icelandic determined by Nominative and not by the choice of subject? (iii) What is the relationship between the parameter selection in (10) and other properties, such as visibility of lexical Case in German and the fixed order of arguments in Icelandic?
592
Dieter Wunderlich
(iv) How does a matrix verb determine that the subject of a dependent verb is either unexpressed or raised? These problems indicate that the notion of subject, though it may be descriptively adequate, does not have much explanatory force. The generalization in (10) does not seem to describe an UG option but may, rather, follow from other options. I will claim in the following that both German and Icelandic exhibit a number of grammatical constructions that require a choice among the arguments of a verb: (i) The agreement morphology on the finite verb calls for a correspondent among the arguments of the verb. (ii) The syntax of Icelandic requires a designated argument in the preverbal position, contrary to German, which does not have such a position; so the necessity of having a notion of ‘‘subject’’ in positional terms may di¤er from language to language. (iii) For a passive verb of Icelandic, both the agreeing argument and the designated argument have to be determined. (iv) Under the assumption that infinitive clauses, when they appear in control or raising constructions, cannot express a full proposition, one argument has to be banned from the infinitive clause. I assume that for each of these phenomena two alternative options are available: either the semantically highest argument (High) or the argument that bears the leastmarked morphological case (Nom) is selected. In all canonical verbs, the highest argument and the Nominative argument coincide. A conflict between the two options can arise only in the passive of ditransitive verbs and in the inverted dat-nom verbs, where the Nominative argument is not the highest one. The generalization made in this proposal is the claim that the notions of High and Nom are su‰cient to replace the undi¤erentiated notion of (underlying) subject. In a way, High is the subject in semantic terms, and Nom is the subject in morphological terms. In the syntax, then, either one can be selected, where this can depend on the type of construction. Each construction type (regardless of the number of languages in which it is manifest) is determined by several constraints, including a pair of constraints with regard to High and Nom. In order to belong to UG, constraints must be specified in more general terms, but some of them may be instantiated (or contextualized) in view of the particular options made otherwise in the language. Most of the restrictions I will explicate in this study have always been assumed to be relevant, although they never have been articulated as violable constraints. Let us consider one pair of constraints that regulate the way infinitive clauses are embedded (thus generalizing about control and raising constructions, among others).
The Force of Lexical Case
593
(11) a. *HIGH-INF (a subcase of SEMHIER) The semantically highest argument is banned from the infinitive clause. b. *NOM-INF The Nominative argument is banned from the infinitive clause. I will claim that both constraints belong to UG but can be ranked di¤erently with respect to each other. German follows the ranking *Nom-Inf SemHier (thereby inducing multiple raising in certain instances), while Icelandic follows the reverse ranking SemHier *Nom-Inf (allowing Nominative in infinitive clauses). These two rankings thus express one aspect of the generalization in (10) without using the notion of subject. The constraints SemHier and *Nom-Inf conspire with Max(lexF), introduced above, in several interesting ways. If SemHier is the dominating constraint (Icelandic), Max(lexF) should play a less important role than in the reverse case (German). But if *Nom-Inf is the dominating constraint (German), the relevant argument role should be identified irrespective of its placement and its semantic ranking; in this case, Max(lexF) should play a more important role than in the reverse case (Icelandic). Therefore, I suggest the following constraint rankings. (12) a. G *Nom-Inf, Max(lexF) SemHier b. I SemHier *Nom-Inf, Max(lexF) In the remainder of this chapter, I will substantiate these proposals. Section 25.4 considers the clause-internal syntax, section 25.5 the control construction, and section 25.6 the raising-to-object construction. A major claim behind these considerations is that the syntax is driven by lexical properties as well as by violable structure-building constraints. 25.4 25.4.1
Case Assignment, Designated Argument, Agreement, and Passive Canonical Case Assignment
Lexical Decompositional Grammar (LDG; Joppen and Wunderlich 1995; Wunderlich 1997; Kaufmann and Wunderlich 1998; Stiebels 1999, 2000; Wunderlich 2000; Wunderlich and Laka¨mper 2001) accounts for all Case marking in a principled way. Following ideas of Kiparsky (1992b), LDG predicts that canonical Case assignment conforms to the semantic ranking of arguments: a medial argument is realized by Dative, the lowest argument by Accusative, and the highest argument by Nominative. (Alternatively, the highest argument is realized by Ergative, and the lowest argument by Nominative.) The basic architecture for Case marking is as follows:
594
Dieter Wunderlich
(i) In the theta-structure of a verb, the l-abstractors for argument variables are ordered according to their semantic ranking, with the lowest argument to the left and the highest argument to the right. (ii) All theta-roles are associated with abstract Case features, where [þhr] means there is a higher argument, and [þlr] means there is a lower argument.4 (iii) The structural morphological Cases are defined in terms of the same features, with Dative ¼ [þhr,þlr], Accusative ¼ [þhr], Ergative ¼ [þlr], and Nominative ¼ [ ] (the default Case). A typical canonical ditransitive verb is zeigen ‘show’, represented in (13b). (13) a. als when b. lz þhr lr
Peter dem Touristen den Dom zeigte Peter the.dat tourist the.acc cathedral showed ly lx {act(x) & see(y,z)} þhr hr þlr þlr
A theta-structure such as the one in (13b), then, forms the input for Case assignment.5 The selection of morphological Case is determined by general faithfulness and markedness constraints regarding the features [þhr] and [þlr]. In an acc(usative)-language, the feature [þhr] is favored by the order Max(þhr) *[þhr], and the feature [þlr] is disfavored by the order *[þlr] Max(þlr); the reverse holds for an erg(ative)-language. Some additional constraints are necessary to derive all possible Case patterns, including those that are lexically determined (Stiebels 2000; Wunderlich 2003). (14) a. MAX(lexF) Every lexically assigned feature in the input has a correspondent in the output. b. DEFAULT Every linking domain displays the default linker (Nominative). c. UNIQUENESS Each linker applies only once in a domain. d. MAX(þhr,þlr) Every feature combination [þhr,þlr] in the input has a correspondent in the output. Max(lexF), already introduced above, is needed to account for lexical marking. Default is motivated by the assumption of economy: Every Case pattern should be realized by minimal e¤ort, so it should include the default form of an NP. Uniqueness serves to avoid ambiguity: If two positions in a Case pattern are realized identically, it is hard to distinguish the arguments, unless the sortal restrictions imposed by
The Force of Lexical Case
595
the verb or the syntactic positions of the arguments function as discriminating factors. Finally, Max(þhr,þlr), a local conjunction in the sense of Smolensky (1995), reflects the requirement that all maximally marked theta-roles should be visible. For reasons of logic, this constraint must rank above the individual Max(þhr) and Max(þlr) constraints. (15) illustrates the optimality of the nom-dat-acc Case pattern for a ditransitive verb. This holds for German as well as for Icelandic; consequently, the constraint ranking is the same in these two languages.6 (15) Ditransitive verbs in a dat/acc system Max (þlr)
*[þhr]
+ nom dat acc
*[þlr]
Max (þhr)
z
Max (þhr,þlr)
Uniq
y
Def
Max (lexF)
x
*
*
**
**
**
nom acc acc
*!
*
nom acc nom
*!
*
*
**
*
nom nom acc
*!
*
*
**
*
nom nom nom
*!*
*
**
**
Turning to passive, I assume that a verb can passivize only if the highest role is agentive and not lexically marked. Under the assumptions of LDG, passive is a lexical rule that binds the (agentive) [hr] role existentially and thus removes it from the theta-structure. More specifically, one can assume that in both German and Icelandic, this rule operates on participles, although nothing hinges on this assumption. The theta-structure of the participle is reduced, as shown in (16), so that only two arguments are left for Case assignment. (17) shows that dat-nom is the optimal Case pattern. (16) gezeigt/sy´nd ‘shown’ [þpart,þpass]: lz ly bx {act(x) & see(y,z)} þhr þhr lr þlr nom dat
596
Dieter Wunderlich
(17) Passive of ditransitive verbs in a dat/acc system
acc acc
*!
*
*
*
*
*
**
*
*[þhr]
*!
Max (þlr)
dat acc
*[þlr]
+ dat nom
Max (þhr)
z
Max (þhr,þlr)
Uniq
Def
Max (lexF)
y
*
**
nom acc
*!
*
*
*
acc nom
*!
*
*
*
*
**
*
nom nom
*!
Observe that Default is responsible for the occurrence of Nominative in the passive and that Max(þhr,þlr) ensures that the medial argument is still realized by Dative. This, again, is true of both German and Icelandic. 25.4.2
Lexical Case
All instances of lexical Case deviate from this canonical Case assignment, often for conceptual reasons (Wunderlich 1997, among others). For instance, the occurrence of Dative Case on the highest argument often correlates with experiencer subjects. LDG allows a semantically highest argument to be lexically marked by the feature [þhr], thereby becoming nonhighest in terms of Case marking, while a semantically lowest argument that is marked as [þlr] will become nonlowest in terms of Case marking. Verbs whose highest argument is lexically marked can never be passivized; this might be explained semantically but also follows from Max(lexF). By contrast, verbs whose lowest argument is lexically marked often can be passivized but preserve the lexical Case in the passive. Examples of the passive of two-place verbs whose lowest argument is marked as Dative or Genitive are given in (18) for German and in (19) for Icelandic. (In the subsequent discussion, Genitive-marked verbs are neglected because they behave somewhat di¤erently from Dative-marked verbs.7) (18) German a. Den Ma¨nnern wurde gefolgt. the man.pl.dat was.u follow.part ‘The men were followed.’
The Force of Lexical Case
597
b. Der Verstorbenen wurde gedacht. the dead.pl.gen was.u remember.part ‘The dead were remembered.’ (19) Icelandic a. Mo¨nnunum var bjargaD. man.mpl.dat was.u save.sup ‘The men were saved.’ b. Sju´klinganna var vitjaD. patient.mpl.gen was.u visit.sup ‘The patients were visited.’ (Andrews 1976, 170) (20) shows Accusative-marked intransitive verbs, and (21) shows inverted Dat-Nom verbs, which were already considered in the introduction. These verbs cannot be passivized. (20) a. G Mich du¨rstet. 1sg.acc be.thirsty ‘I am thirsty.’ b. I Ba´tana hefur brotiD i spo´n. boat.pl.acc has.u broken.u in pieces ‘The boats have broken into pieces.’ (Maling and Zaenen 1990, 145) (21) a. G Mir sind sie immer langweilig vorgekommen. 1sg.dat are they.nom always boring found b. I Me´r hefur/hafa alltaf ott eir leiDinlegir. 1.dat has.u/have.3pl always thought they.nom boring.nom ‘I have always found/thought them boring.’ The Dative-marked verbs in (18) and (19) are described by the lexical feature [þlr]; if it is combined with the default feature [þhr] for the semantically lowest argument, the Case that associates with the relevant argument is Dative, as shown in (22a). By contrast, for the verbs in (20) and (21) the lexical feature [þhr] is used. No further argument role is present in (20), so in this case the argument role is matched best with Accusative, as shown in (22b). In the inverted dat-nom verbs in (21), the lexical feature [þhr] is combined with the default feature [þlr] for the semantically highest argument, so in this case the relevant argument role is again matched with Dative, as in (22c). (22) Lexical: Default: Case:
a. NOM-DAT verbs ly lx verb(x,y) þlr þlr þhr hr dat nom
b. ACC-verbs lx verb(x) þhr lr acc
c. Inverted verbs ly lx verb(x,y) þhr þhr lr þlr nom dat
598
Dieter Wunderlich
That the annotated Case distribution is indeed the optimal one is shown in (23), with the same constraint ranking as before.8 (23) The optimal Case patterns for some lexically marked verbs9
b.
*!
*
*
*
**
*
*
acc-verbs x + acc nom
c.
*[þhr]
nom acc
Max (þlr)
+ nom dat
*[þlr]
Max (þhr)
Max (þhr,þlr)
Uniq
Def
nom-dat verbs x y
Max (lexF)
a.
*
*
*!
*
Inverted verbs x y + dat nom
* *!
dat acc
*!
acc acc nom acc
*!
* *
*
*
*
*
** *
**
*
*
All other types of lexical marking, occurring in either German or Icelandic, can be described by the same means. However, I will not go into more details here. 25.4.3
Designated Arguments
A number of syntactic di¤erences between German and Icelandic follow from di¤erences in the V and I positions. V is final in the VP of German but initial in the VP of Icelandic. Likewise, I is final in the IP of German but initial in the IP of Icelandic. German does not need to fill a SpecI position in the subordinate clause, because this position could only be vacuously distinguished from a SpecV position; therefore, I assume (following Haider 1988) that German has a matched I/VP projection. In the main clause of German, the inflected verb is realized in C, and SpecC can be filled by any XP, regardless of its status as argument or adjunct. In contrast, Icelandic exhibits the same minimal IP
The Force of Lexical Case
599
structure in both subordinate and main clauses (see also Ro¨gnvaldsson and Thra´insson 1990), and SpecI has to be filled. I call the argument that is realized in SpecI ‘‘designated.’’ In a way, the syntax of Icelandic is more ‘‘articulated’’ than that of German, because of the di¤erent positions of V and I. Subordinate clauses are illustrated in (24), main clauses in (25). (24) a. G [C 0 daß [I/VP ich Maria nie getro¤enV habeI ]] b. I [C 0 aD [IP e´g [I 0 hef [VP aldrei hitt Marı´u]]]] ‘that I have never met Mary’ (25) a. G [CP Ich [C 0 habe [I/VP Maria nie getro¤enV — ]]] b. I [IP E´g [I 0 hef [VP aldrei hitt Marı´u]]] ‘I have never met Mary.’ Both (24) and (25) display the basic word order, in accordance with SemHier, requiring that the linear order of arguments correspond to their semantic ranking. German allows scrambling within the subordinate clause, whereas Icelandic does not. Since in Icelandic SpecI is filled by the designated argument, there is no place for another argument. In the main clause, both languages allow object topicalization with a resulting CP, thereby violating SemHier. In German, it is simply the object that fills SpecC, whereas in Icelandic, a more complex structure arises through object movement, as shown in (26). (26) a. G [CP Maria [C 0 habe [I/VP ich — nie getro¤enV — ]]] b. I [CP Marı´u [C 0 hef [IP e´g [ I 0 — [VP aldrei hitt ]]]]] ‘Maria, I have never met.’ The projections with passive participles di¤er similarly. (27) a. G [C 0 daß [I/VP mir geholfen wurde]] b. G [CP Mir [C 0 wurde [I/VP geholfen — ]]] c. I [IP Me´r [I 0 var [VP hja´lpaD]]] ‘I was helped.’
(subordinate clause) (main clause)
The fact that the syntax of Icelandic requires a designated argument to fill the SpecI position, whereas German does not, correlates with another di¤erence between German and Icelandic: German has only one option for the passive of ditransitive verbs, whereas Icelandic has two. 25.4.4
DAT-NOM
Patterns in Icelandic
It has been argued in the literature that the inverted dat-nom verbs are the ‘‘unaccusative’’ counterpart of the passive of canonical ditransitives (Fanselow 2000)—the former lack an ‘‘external’’ argument, while the latter have the ‘‘external’’ argument
600
Dieter Wunderlich
existentially bound—otherwise, these two classes of verbs have the same thetastructure, hence they map to the same Case pattern. This argumentation goes through for German. In Icelandic, however, these two classes of verbs behave di¤erently. First, Icelandic allows two ways of passive formation of a ditransitive verb, as illustrated in (28) (taken from Andrews 1976/1990, 179). (28) a. Honum voru sy´ndir drengirnir. 3msg.dat aux.past.3pl shown.mpl.nom boy.pl.nom ‘He was shown the boys.’ b. Drengirnir voru sy´ndir honum. boy.pl.nom aux.past.3pl shown.mpl.nom 3msg.dat ‘The boys were shown to him.’ The sentences in (28) exhibit di¤erent arguments in SpecI: in (28a), honum is in SpecI, whereas in (28b), drengirnir is in SpecI. If these sentences are embedded under an object-raising verb such as telja ‘believe’, it is always the element in SpecI that is raised. Hence a crucial di¤erence of the two passive variants arises in the context of raising to object. The Dative subject remains unchanged, as shown in (29a), while the Nominative subject becomes Accusative, as shown in (29b). (29) a. E´g tel honum hafa veriD sy´ndir drengirnir. 1sg believe 3msg.dat have been shown.nom boy.pl.nom ‘I believe him to have been shown the boys.’ b. E´g tel drengina hafa veriD sy´nda honum. 1sg believe boy.pl.acc have been shown.acc 3msg.dat ‘I believe the boys to have been shown to him.’ (Andrews 1976, 180) It follows, then, that in the passive of a ditransitive verb either the next-to-highest argument or the Nominative argument is designated. The only other verbs in which a Nominative argument di¤ers from the highest one are the inverted dat-nom verbs; however, in these verbs only the highest argument (Dative) is designated, as shown by the ungrammaticality of (30b). E´g tel henni hafa alltaf lı´kaD drengirnir. 1sg believe 3fsg.dat have always liked boy.pl.nom ‘I believe her to have always liked the boys.’ b. *E´g tel drengina hafa alltaf lı´kaD henni. 1sg believe boy.pl.acc have always pleased 3fsg.dat ‘I believe the boys to have always pleased her.’
(30) a.
Such a di¤erence between the two types of dat-nom patterns does not occur in German because here the Nominative argument is always changed to Accusative in the context of raising to object (see section 25.6.1 below).
The Force of Lexical Case
601
Another di¤erence between the passive of ditransitives and the inverted dat-nom verbs is found with respect to agreement. Generally speaking, both German and Icelandic obey the following constraint. (31) AgrNom The finite verb agrees with the argument realized in the Nominative. If there is no Nominative argument, the verb gets the default marking u¼3sg. This is also true of the Icelandic passives of ditransitives, illustrated in (27) above. However, the inverted dat-nom verbs allow for two options: agreement with the Nominative argument or the default marking, as shown in (21b), repeated here as (32).10 (32) Me´r hefur/hafa alltaf ott eir leiDinlegir. 1sg.dat has.u/have.3pl always thought 3pl.nom boring.nom ‘I have always thought them boring.’ This di¤erent behavior of the two classes of verbs is summarized in (33). (33) Passive of ditransitive verbs Inverted dat-nom verbs
Designated argument
Agreement
dat/nom dat
nom nom/u
This di¤erence can be accounted for by the following assumptions: (i) Icelandic inverted dat-nom verbs designate the higher argument with the feature [þhr] and, simultaneously (as opposed to German), the lower argument with the feature [hr], so that the resulting Nominative argument is lexically marked. (This kind of Nominative has been called ‘‘oblique’’ in the literature.)11 (ii) The designated argument (to be realized in SpecI) is either the highest argument or the Nominative argument but never a lexically marked Nominative. The following DesHigh constraint is a subcase of SemHier: a. DESHIGH The highest argument is designated—this allows the preservation of the underlying order of arguments. b. DESNOM The Nominative argument is designated—this allows the agreement relation to hold between the element in I and the element in SpecI. c. *DESLEXNOM A lexically marked Nominative argument cannot be designated. (iii) Besides the requirement that the verb agree with the Nominative (AgrNom), there exist two prohibitions against agreement (which are both satisfied in German):
602
Dieter Wunderlich
a. *AGRNONNOM The finite verb does not agree with any argument other than the Nominative. b. *AGRLEXF The finite verb does not agree with any argument associated with a lexical feature. The tableaux in (34) show how the designated argument is selected, and the tableaux in (33) show how the agreement facts come about. (34) Optional versus obligatory designation of the highest argument a.
Passive of ditransitive verbs
DesHigh
DesNom
+ (15a): [ip dat . . . [vp . . . nom + (15b): [ip nom . . . [vp . . . dat b.
*DesLexNom
* *
Inverted dat-nom verbs +
[ip dat . . . [vp . . . nom [ip nom . . . [vp . . . dat
* *
*!
If all three Des constraints are equally ranked, two equally good options arise for the passive of ditransitives, whereas the designation of the Nominative argument leads to more violations when the Nominative is lexically marked. (35) Obligatory versus optional agreement with Nominative a.
Passive of ditransitive verbs dat Vagr:dat nom
*AgrNonNom
AgrNom
*!
*
*AgrLexF
+ dat Vagr:nom nom *!
dat Vu nom b.
Inverted dat-nom verbs dat Vagr:dat nom
*!
*
+ dat Vagr:nom nom + dat Vu nom
* *
By contrast, if all three Agr constraints are equally ranked, a lexical Nominative does not need to induce agreement; the inverted verbs thus allow for two agreement options. This analysis is based on the stipulation that the inverted dat-nom verbs have an ‘‘oblique’’ (i.e., lexically marked) Nominative, while the Nominative in the passive of
The Force of Lexical Case
603
ditransitives is only induced by Default. In addition to the two classes of dat-nom verbs just mentioned, Icelandic also exhibits a minor class of inverted dat-nom verbs (though it is not a small class due to a number of complex predicates) that alternate in the choice of the designated argument in the same way as the passive of ditransitives (Berno´dusson 1982; BarDdal 1999). Consider the examples in (36). (36) a. E´g tel henni henta drengirnir. 1sg believe 3fsg.dat please.inf boy.pl.nom ‘I believe her to be pleased by the boys.’ b. E´g tel drengina henta henni. 1sg believe boy.pl.acc please.inf 3fsg.dat ‘I believe the boys to please her.’ These verbs simply lack the lexical feature [hr] for the lower argument and therefore pattern with the passive of ditransitive verbs, shown in (34a) and (35a).12 However, they are puzzling under the hypothesis that each verb has at most one ‘‘subject’’; these verbs seem to have two subjects. Icelandic thus turns out to be symmetric not only with respect to which object is realized as subject (i.e., in SpecI) in the passive of ditransitive verbs but also with respect to which argument is realized as subject in a small class of alternating dat-nom verbs. This is due to two factors: (i) there is a syntactically designated position to be filled by an argument, and (ii) no preference is made among the constraints that determine the choice of this argument. 25.5 25.5.1
Control
Zu-Infinitives in German
German control verbs take zu-infinitives. The infinitive is a [agr] form, which does not realize any agreement; moreover, it excludes Nominative arguments in German. However, the argument that is excluded from the infinitive clause can be controlled by a matrix verb. I will consider only obligatory (subject) control, where the argument of the matrix verb is identified with the missing argument in the infinitive. A minimal pair of possible versus impossible control is shown in (37). (37) a. Ich ho¤e, — nom nicht verfolgt zu werden. b. *Ich ho¤e, — dat nicht gefolgt zu werden. ‘I hope not to be followed.’ The passive of a canonically transitive verb such as verfolgen is accepted in the control construction, whereas the passive of a Dative verb such as folgen is not. The control construction is impossible with semantically zero-place verbs (weather verbs), such as the one in (38a), and is also blocked if all argument roles of the
604
Dieter Wunderlich
dependent infinitive are realized, as in (38b) (where ihm schmeckt der Apfel ‘he enjoys the apple’ would express a full proposition), even if Nominative is avoided. (38) a. *Ich ho¤e, es zu regnen. 1sg hope it to rain ‘I hope it rains.’ b. *Ich ho¤e, ihm den Apfel zu schmecken. 1sg hope 3msg.dat the.acc apple to taste ‘I hope he tastes the apple.’ This observation can be captured by a constraint that forbids infinitive clauses to express a full proposition. (39) *INF-PROP An infinitive clause must have an open argument role. This constraint can be implemented in the representation of an obligatory control verb such as ho¤en ‘hope’ if one assumes that the dependent infinitive predicates of the same argument (x) as the control verb, as shown in (40a). Let us further assume that the control verb has an argument role that is lexically marked by [a], and the dependent infinitive has an open argument role lexically marked by [b]. The stage in which such a control verb is combined with the zu-infinitive is given in (40b). By internal functional application, x replaces u, as shown in (40c). (40) a. Control verb lP lx verb(x, P(x)) zu-inf [a] b. Plus zu-inf lx verb(x, lu dep.verb(u) (x)) [a] [b] c. Result lx verb(x, dep.verb (x)) [f] The open question is whether such a construction converges and how the resulting Case feature f is computed: is f¼a, or f¼b, or f¼a A b (by unification)? An answer is given by the first of the locality constraints in (41). (41) Locality constraints a. LOCCASE Case features are locally assigned. b. DOMAIN All arguments are realized within the licensing domain.
The Force of Lexical Case
605
LocCase implies that the controlling argument role cannot realize Case features of the controlled argument role, and vice versa, that is, f must be identical to a. Domain corresponds to the extended projection principle in syntactic accounts. In the following, I will assume that argument gaps, as found in control constructions, as well as argument raising or topicalization constitute Domain violations, whereas clause-internal scrambling does not. Therefore, Domain must be a low-ranked constraint in both German and Icelandic. A further question is which of the arguments of an infinitive clause has to be controlled. The constraints in (42) o¤er two options: either the highest argument or the Nominative argument. (42) a. *HIGH-INF The highest argument is not allowed to stay in an infinitive clause. b. *NOM-INF No Nominative argument is allowed to stay in an infinitive clause. *High-Inf can be considered a subcase of SemHier: a gap in the infinitive clause does not alter the linear order of arguments. But since the controlling argument is identical with the gap, it counts as the relevant argument; and since it is ordered before the infinitive clause, it must be the highest argument in order not to alter the linear order predicted by semantic ranking. Note that *Inf-Prop corresponds to the disjunction of the two constraints in (42). *Inf-Prop is violated only if both *High-Inf and *Nom-Inf are violated. The standard account, which assumes a syntactic pro argument in the infinitive clause (thus not violating Domain), is compatible with this view. Indeed, none of these constraints is violated if the embedded verb is a canonical one, as in (37a). However, if the inverted dat-nom verbs are involved, one of these constraints will be violated: *High-Inf in German and *Nom-Inf in Icelandic. Let us first consider passives with lexical Case (where *Nom-Inf and *High-Inf are irrelevant). If one intends to express the German utterance ich ho¤e, daß mir geholfen wird ‘I hope that I will be helped’ as a control construction (‘I hope to be helped’), three candidates come to mind. All three are ungrammatical. (43) a. *Ich ho¤e, geholfen zu werden. 1sg hope helped to aux b. *Mir ho¤t, geholfen zu werden. 1sg.dat hopes helped to aux c. *Ich ho¤e, mir geholfen zu werden. 1sg hope 1sg.dat helped to aux (43a) violates Max(lexF), since Dative is ignored, (43b) violates LocCase, since Dative is realized on the matrix argument, and (43c) violates *Inf-Prop, since the
606
Dieter Wunderlich
dependent infinitive has no open argument. From the ungrammaticality of the examples in (43) we can conclude that all three constraints are undominated in German. If there is no candidate that escapes one of the undominated constraints, it is the Nullparse (Prince and Smolensky 1993) that wins. The Nullparse simply repeats the input but does not integrate the input elements into a resulting structure that converges semantically; it only violates the constraint *q ‘‘Avoid Nullparse’’ (similar to the MParse constraint of Prince and Smolensky 1993).13 (44) shows an instance in which the Nullparse is the winning candidate. (44) Evaluation of dat-passives in a control construction LocCase
*InfProp
Ich ho¤e, geholfen zu werden (¼ (41a)) Mir ho¤t, geholfen zu werden (¼ (41b)) Ich ho¤e, mir geholfen zu werden (¼ (41c))
Max (lexF)
*q
*!
Domain *
*!
* *!
+ Nullparse
*
Nothing excludes that the control verb itself is lexically marked. In fact, (45a) shows that the controller may be lexically marked for Dative. However, LocCase and Max(lexF) exclude that such a Dative argument controls a Dative of the dependent verb, as shown by the ungrammaticality of (45b). (45) a. Ihrdat gelang es, — nom nicht verfolgt zu werden. b. *Ihrdat gelang es, — dat nicht gefolgt zu werden. ‘She managed not to be followed.’ If the dependent verb is an inverted dat-nom verb, the ranking between *High-Inf and *Nom-Inf becomes decisive. The examples in (46) show that *High-Inf can be violated, that is, a nonhighest Nominative argument can be controlled in German, whereas *Nom-Inf cannot. In the ungrammatical sentence (46c), the violation of *Nom-Inf coincides with a violation of Max(lexF). (46) a. b.
Ich ho¤e, ihrdat — nom zu gefallen. ‘I hope to be liked by her.’ Der Hund ho¤te, einmal im Leben einemdat freundlichen Herrn — nom zu geho¨ren. ‘The dog hoped to belong to a nice master once in his life.’
The Force of Lexical Case
607
c. *Ich ho¤e, — dat ernom zu gefallen. ‘I hope to like him.’ With the constraint ranking Max(lexF), *Nom-Inf *q *High-Inf, these data can be captured, as shown in (47). (47) Evaluation of inverted dat-nom verbs in a control construction Max (lexF) a.
*NomInf
+ Ich ho¤e [vp ihr — nom zu gefallen]
*
Nullparse b.
*q
Ich ho¤e [vp — dat er zu gefallen] + Nullparse
HighInf
*! *!
* *
Nullparse is excluded in (47a) because there is a candidate that does not violate one of the undominated constraints. In contrast, Nullparse wins in (47b) because the next-to-best candidate violates undominated constraints. 25.5.2
A¶-Infinitives in Icelandic
Icelandic control verbs require infinitives with the complementizer a¶. Considering the Icelandic counterparts of the candidates in (43) (‘I hope to be helped’), one can see that the first, in (48a), is acceptable, whereas the other two, in (48b,c), are ungrammatical, as in German. E´g vonast til aD verDa hja´lpaD. 1sg hope for to be helped ‘I hope to be helped.’ b. *Me´r vonast til aD verDa hja´lpaD. 1sg.dat hope for to be helped c. *E´g vonast til aD me´r verDa hja´lpaD. 1sg hope for to 1sg.dat be helped
(48) a.
I conclude from these data that in Icelandic, Max(lexF) is lower ranked than *q. The evaluation of the possible candidates is illustrated in (49).
608
Dieter Wunderlich
(49) Evaluation of dat-passives in a control construction LocCase
*InfProp
*q
Max (lexF)
Domain
*
*
+ E´g vonast til aD verDa hja´lpaD. Me´r vonast til aD verDa hja´lpaD. E´g vonast til aD me´r verDa hja´lpaD.
*!
* *!
Nullparse
*!
Andrews (1976) remarks that Icelandic speakers judge the alternative with a subjunctive clause, shown in (50), slightly better than (48a); this indicates that the speakers are aware of the Max(lexF) violation in (48a). Note that the complementizer a¶ in (50), introducing a clause with a finite verb, is di¤erent from the a¶ that combines with an infinitive clause. For obvious reasons, a finite verb clause and an infinitive clause are not in the same candidate set; hence (48a) and (50) are not direct competitors but rather paraphrases of each other. (50) E´g vonast til aD me´r verDi hja´lpaD. I hope for that me.dat be.subj helped ‘I hope for me to be helped.’ (Andrews 1976, 175) One reason Max(lexF) may be violated in Icelandic is that lexical case is often morphologically visible at some other category. Consider the following examples in which an adjective shows the marked Case of the unrealized controlled argument. This is possible because adjectival predicates do not only agree with their argument in gender and number but also show Case concord. (51a) shows an intransitive verb with lexically marked Accusative and a secondary predicate in the same Case. In (51b), this verb occurs in a control construction in which the marked Accusative is still displayed by the adjective. In (51c), we see a control verb marked for Accusative, which controls an argument marked for Dative, whereas (51d) is an example in which the reverse marking appears. (51) a. Hana rak a´ land eina. 3fsg.acc drifted ashore alone.fsg.acc ‘She drifted ashore alone.’ b. Hu´n vonast til aD reka a` land eina. 3fsg.nom hopes at to drift.inf ashore alone.fsg.acc ‘She hopes to drift ashore alone.’ (Andrews 1976, 175)
The Force of Lexical Case
609
c. Mig langar aD vera kastaD einum u´t um gluggann. 1sg.acc long to be thrown alone.msg.dat out of the window ‘I long to be thrown out of the window alone.’ (Andrews 1976, 176) d. Honum otti leitt aD verDa einn eftir. 3msg.dat thought loathsome to stay alone.msg.acc after ‘He hated to stay behind.’ (Kress 1982, 217) These examples, which instantiate the hypothetical case of (40b) with a0b, show that the di¤erent features cannot be unified, neither on the matrix verb nor on the embedded verb, as required by LocCase. It has been argued in the literature that examples like those in (51) provide evidence for the existence of pro, in this case a pro that bears lexical Case. However, there is no need for assuming pro as a syntactic element. Under the present account, it is the theta-role that bears the lexical Case feature ([þhr]); thus a secondary predicate inherits this Case feature when the infinitive clause is formed, and the theta-role is indexed with the agreement features realized on this predicate. In the next step, illustrated in (52), the infinitive clause is embedded under a control verb. The lexical Case feature of the dependent verb has to be ignored in this case, whereas the agreement features are passed to the controlling argument (according to the agreement theory in Wunderlich 1994; Pollard and Sag 1994). In other words, agreement features, which are semantically motivated, have to be unified, in contrast to Case features. (52) lx
lu fsg {drift-ashore(u) & alone(u)} (x)) þhr hope(x, {drift-ashore(x) & alone(x)})
hope(x,
¼ lx fsg
If the controlled argument bears no lexical Case, one expects that a corresponding secondary predicate is realized in the Nominative (the unmarked Case). This is indeed one option. However, as Andrews (1990) observes, it is also possible that the secondary predicate takes over the marked Case from the control verb. (53) a. Mig langar aD fara ´ı kaupstaDinn einn/einan. 1sg.acc long to go into town alone.nom/acc ‘I long to go to town alone.’ (Andrews 1990, 176) b. lP lx {long(x, P(x))} þhr This long-distance Case relationship poses a problem that cannot be solved by the assumption of pro.acc because the Case features [a] and [b] cannot be unified on the embedded verb, according to LocCase. The fact that nevertheless einan.acc is possible in (53a) can be explained as follows. According to the representation of the control verb in (53b), the infinitive clause P predicates the argument x; hence, it can
610
Dieter Wunderlich
inherit the lexical Case feature from the control verb. The infinitive itself cannot bear lexical Case, but it may pass the Case feature to a secondary predicate of the same argument (if the embedded verb is not itself marked for lexical Case). This syntactically triggered top-down strategy, leading to einan.acc, competes with the lexically triggered bottom-up strategy, which leads to einn.nom. If we now turn to the inverted dat-nom verbs in Icelandic, we see that (54a), in which the higher argument is preserved and the lower one is controlled, is strictly ungrammatical. This shows that *High-Inf ranks above *Nom-Inf. In the reverse case, the higher argument is controlled and the Nominative is preserved within the infinitive clause. Here, the judgments di¤er to some extent. Some examples, such as (54b) with a pronoun, are hardly acceptable, but a full DP, as in (54c), improves a lot on the construction. These examples violate both Max(lexF) and *Nom-Inf, which may explain why they are only marginally acceptable. The alternative shown in (54d) with a PP only violates Max(lexF) and is fully acceptable. *E´g vonast til aD henni lı´ka — nom . 1sg hope for to 3fsg.dat like — nom ‘I hope to be liked by her.’ b. ???E´g vonast til aD — dat lı´ka hu´n. 1sg hope for to — dat like 3fsg.nom ‘I hope to like her.’ c. ?E´g vonast til aD — dat lı´ka thessi bo´k. 1sg hope for to — dat like this book.nom ‘I hope to like this book.’ d. E´g vonast til aD — dat lı´ka viD hana. 1sg hope for to — dat like with 3fsg.acc ‘I hope to like her.’
(54) a.
violates *High-Inf, which is fatal violates Max(lexF) and *Nom-Inf violates Max(lexF) and *Nom-Inf violates Max(lexF)
(55b), taken from Taraldsen (1995), is another example showing that the control construction with an inverted dat-nom verb is not ungrammatical. (55) a. Henni leiddust/leiddist their. 3fsg.dat be.bored.3pl/u 3pl.nom ‘She was bored with them.’ (Taraldsen 1995, 307) b. ?ViD reyndum aD — dat leiDast hu´n ekki. we tried.1pl to — dat be.bored 3fsg.nom not ‘We tried not to be bored with her.’ (Taraldsen 1995, 322) These observations can best be captured by the constraint ranking assumed in (56); it predicts that only the highest argument can be controlled.
The Force of Lexical Case
611
(56) Evaluation of inverted dat-nom verbs in a control construction *HighInf E´g vonast . . . [vp henni lı´ka — nom ]
a.
+ Nullparse b.
Max (lexF)
*NomInf
*
*
*! *
+ E´g vonast . . . [vp — dat lı´ka thessi bo´k] Nullparse
25.6
*q
*!
Raising to Object
25.6.1
Raising in German
Only a few German verbs trigger raising to object; most of them are verbs of perception. The examples in (57) illustrate that the highest argument of the bare infinitive is raised and gets canonical Accusative by the matrix verb. If the dependent verb is transitive, a double Accusative appears, as in (57c); this indicates that only the highest argument is raised, while the lower object remains part of the VP infinitive clause. (57) a. Ich sah ihn kommen. 1sg saw 3msg.acc come ‘I saw him come.’ b. Ich sah ihn verfolgt werden. 1sg saw 3msg.acc followed up be ‘I saw him be followed up.’ c. Ich sah den Mann den Jungen schlagen. 1sg saw the.acc man [VP the.acc boy hit] ‘I saw the man hit the boy.’ In view of the data in (58), one has to assume that the expletive argument of a weather verb can be raised. If there is no argument to be raised, the construction is ungrammatical, as shown by (58b,c). (58) a.
Ich I b. *Ich I c. *Ich I
sah es regnen. saw it rain sah getanzt werden. saw danced be ho¨rte umgera¨umt werden. heard replaced be
612
Dieter Wunderlich
The following examples with lexically marked verbs are nearly acceptable for the majority of speakers. Some speakers find these examples good, while others reject them. Note that (57b–d) preserve the semantic order of arguments. (59e) with a scrambled order is judged worse than (59d), whereas (59f ) with a Nominative is ungrammatical.14 (59) a. ?Ich sah ihm geholfen werden. 1sg saw 3msg.dat helped be ‘I saw him be helped.’ b. ?Ich sah ihm den Mantel gezeigt werden. 1sg saw 3msg.dat the.acc coat shown be ‘I saw him be shown the coat.’ c. ?Ich ho¨rte ihm den Apfel schmecken. 1sg heard 3msg.dat the.acc apple enjoy ‘I heard him enjoy the apple.’ d. ?Ich sah ihm den Film gefallen. 1sg saw 3msg.dat the.acc film like ‘I saw him like the film.’ e. ??Ich sah den Film ihm gefallen. 1sg saw the.acc film 3msg.dat like Intended: ‘I saw him like the film.’ f. *Ich sah ihm der Film gefallen. 1sg saw 3msg.dat the.nom film like Intended: ‘I saw him like the film.’ Let us assume that the object-raising verb takes an infinitive VP and assigns structural Case to the raised argument. This follows from the clause union account, according to which the relevant argument role becomes part of the theta-structure of the matrix verb. Since a lexically marked argument is not assigned Accusative, as in (59a), the lexical representation of an object-raising verb should not make any reference to the arguments of the dependent predicate. Therefore, the representation of such a verb simply requires that it takes a propositional expression to be realized by an infinitive clause. (60) shows how sehen ‘see’ can be combined with an infinitive of kommen ‘come’ by functional composition. In the result, (60c), every lexical marking of lu must be preserved. (60) a. lp lx see(x,p) inf b. lp lx see(x,p) [lu come(u)] c. lu lx see(x,come(u)) Nothing in the lexical representation (60a) excludes the combination with zero-place predicates like those in (58b,c), and nothing accounts for why the examples in (59)
The Force of Lexical Case
613
are not fully acceptable. Therefore, further constraints are necessary. To the set of constraints already introduced, I add ECM, defined in (61). (61) ECM A raised argument gets structural case by the matrix verb. (62) Ranking for German Max(lexF),*Inf-Prop,*Nom-Inf ECM,*High-Inf Domain *Inf-Prop accounts for the fact that bare infinitive phrases with no argument to be raised cannot appear with a raising verb. I therefore assume that all the examples in (59a) to (59f ) are instances of raising. In (59a), a lexically marked argument is raised and does not satisfy ECM, whereas in (59b–e), a nonhighest argument is raised and satisfies ECM. Given the order of arguments, (59b–d) must be instances of multiple raising (i.e., multiple violations of Domain).15 (59f ) is an example that violates *Nom-Inf. As we will see in the next subsection, multiple raising is not possible in Icelandic, due to a di¤erent ordering of Domain. Summing up, none of the examples violates Max(lexF), but the ungrammatical instances of (56) violate *Inf-Prop. The (partially) acceptable instances of (59) either violate ECM (59a) or they violate Domain twice (59d,e). The tableau in (63) shows that an ECM violation is more tolerable than a Max(lexF) or *Inf-Prop violation. Speakers who reject the best candidate of (63) do not tolerate an ECM violation. (63) Evaluation of dat-passives in a raising-to-object construction
Nullparse
*!
Domain
Ich sah [vp ihmdat geholfen werden]
*
High-Inf
Ich sah ihnacc [vp geholfen werden]
ECM
*q
*Nom-Inf
*Inf-Prop
Max (lexF) + Ich sah ihmdat [vp geholfen werden]
* *
*! *!
(64) shows the evaluation of (59d–f ). It turns out that the candidate with multiple raising (a double violation of Domain) meets best the given constraint ranking. Speakers who reject (59d) do not tolerate multiple raising.
614
Dieter Wunderlich
(64) Inverted dat-nom verbs in a raising-to-object construction
+ ihm den Film [vp gefallen]
**
den Film [vp ihm gefallen]
*!
ihm [vp der Film gefallen] ihn [vp der Film gefallen] [vp ihm der Film gefallen] Nullparse
Domain
*High-Inf
ECM
*q
*Nom-Inf
*Inf-Prop
Max (lexF)
Ich sah . . .
*! *!
*
* *!
* * *
* *!
In the following we will see that the relative order of *Nom-Inf and *High-Inf is the opposite in Icelandic. 25.6.2
Raising in Icelandic
Icelandic exhibits more verbs than German that have the option of raising to object, among them verbs of saying or thinking. Propositions such as those in (65) can easily be embedded under raising verbs, as shown in (66). (65) a. Stra´karnir voru kitlaDir. boy.mpl.nom aux tickled.mpl.nom ‘The boys were tickled.’ b. Stra´kunum var bjargaD. boy.mpl.dat aux rescued.u ‘The boys were rescued.’ (Andrews 1990, 189) (66) a. E´g tel stra´kana (hafa veriD) kitlaDa. 1sg believe boy.mpl.acc (have been) tickled.mpl.acc ‘I believe the boys to have been tickled.’ b. E´g tel stra´kunum (hafa veriD) bjargaD. 1sg believe boy.mpl.dat (have been) rescued.u ‘I believe the boys to have been rescued.’ (Andrews 1990, 190) Of particular interest here is the fact that the raised arguments can become designated (subject) in the passive, which is illustrated in (67). These data provide evidence that the raised arguments in fact belong to the domain of the matrix verb.
The Force of Lexical Case
615
(67) a. Stra´karnir eru taldir (hafa veriD) kitlaDir. boy.mpl.nom are believed.mpl.nom (have been) tickled.mpl.nom ‘The boys are believed to have been tickled.’ b. Stra´kunum er taliD (hafa veriD) bjargaD. boy.mpl.dat is believed.u (have been) rescued.u ‘The boys are believed to have been rescued.’ (Andrews 1990, 190) The (a) sentences of (66) and (67) demonstrate that a structural Nominative becomes Accusative in the raising construction but again Nominative under passive, while a Dative argument in the (b) sentences stays inert. Thus in no case is Max(lexF) violated. The following tableau shows that the sentence in (66b) comes out best. Crucially, ECM must be dominated by Max(lexF). (68) Evaluation of dat-passives in a raising-to-object construction *InfProp
*q
Max (lexF)
+ E´g tel stra´kunum.dat [vp . . .
*
E´g tel stra´kana.acc [vp . . . E´g tel [vp stra´kunum . . . Nullparse
ECM
*! *! *!
The example in (69), involving an inverted Dat-Nom verb, illustrates another clear di¤erence from German. Here, the nonraised Nominative argument remains Nominative, which is ungrammatical in German; compare (69) with (59f ). ´ lafur leiDinlegur. (69) E´g tel henni hafa alltaf o´tt O 1sg believe 3fsg.dat have always thought Olaf.nom boring.nom ‘I believe her always to have found Olaf boring.’ (Zaenen et al. 1985, 101) The tableau in (70) provides an analysis. The optimal candidate violates both ECM and *Nom-Inf. Therefore, these constraints must be the lowest-ranked ones.
616
Dieter Wunderlich
(70) Inverted dat-nom verbs in a raising-to-object construction
*
*
**!
. . . hana.acc ´ lafur.nom] [vp . . . O
Nullparse
* *
. . . henni.dat ´ laf.acc [vp . . . ] O
. . . [vp henni.dat . . . ´ lafur.nom] O
*Nom-Inf
*!
ECM
´ laf.acc ... O [vp . . . henni.dat]
Domain
Max (lexF)
*q
*High-Inf
*Inf-Prop + . . . henni.dat ´ lafur.nom] [vp . . . O
* *!
*!
* *
*!
As pointed out above, either of the two passive variants of a ditransitive verb can undergo the raising-to-object construction, giving rise to the options in (28), repeated here as (71).16 (71) a. E´g tel honum hafa veriD sy´ndir drengirnir. 1sg believe 3msg.dat have been shown.nom boy.mpl.nom ‘I believe him to have been shown the boys.’ b. E´g tel drengina hafa veriD sy´nda honum. 1sg believe boy.mpl.acc have been shown.acc 3msg.dat ‘I believe the boys to have been shown to him.’ The postulated constraint ranking seems to predict that contrary to the facts only one option, namely (71a), can occur.
The Force of Lexical Case
617
(72) Passive of ditransitive verbs in a raising-to-object construction *Inf- *High- Max *NomProp Inf (lexF) Domain ECM Inf drengina.acc [vp . . . honum.dat]
*!
+ honum.dat [vp drengirnir.nom] honum.dat drengina.acc [vp . . . ]
* *
*
*
**!
However, if one assumes that the raising-to-object construction applies to an IP rather than a VP, the two passive structures that underlie (71a) and (71b) constitute di¤erent inputs for the evaluation of raising to object,17 so the two options in (71) cannot be compared directly. In the recent literature it has been argued that cyclic optimization is needed; for example, Kiparsky (2002b) argues for cyclic morphology (stem level vs. word level) on the basis of opacity data in Arabic dialects, and Heck and Mu¨ller (2000) argue for cyclic syntax on the basis of data on repair-driven movement. Accordingly, if the evaluation of syntactic structures in Icelandic is cyclically performed, the undesirable result of (72) is avoided. In the first cycle the embedded IP is evaluated, and in the second cycle each optimal candidate from the first cycle is embedded into a more complex construction. However, active and passive do not form separate cycles, since neither is syntactically derived from the other. The number of cycles that are necessary to yield the two options with the raising verb being passivized, shown in (73), is still two. (73) a. Drenginir eru taldir hafa veriD sy´ndir honum. boy.pl.nom are believed have been shown 3msg.dat ‘The boys are believed to have been shown to him.’ b. Honum er taliD hafa veriD sy´ndir drengirnir. 3msg.dat is believed have been shown boy.pl.nom ‘He is believed to have been shown the boys.’ In other words, only IPs, which allow for a new choice of the designated argument, should constitute a new cycle in Icelandic. 25.7
Summary
The lexical di¤erences between German and Icelandic are by and large rather marginal; both have similar types of lexical marking and follow the same constraint
618
Dieter Wunderlich
ranking for Case assignment. I have argued that some of the major di¤erences between these two languages follow from the fact that the syntactic constructions induced by control and raising verbs react di¤erently to the semantic and morphological properties of the embedded verb. In order to account for these di¤erences, the notion of ‘‘underlying subject’’ has been replaced by a pair of notions: High (the semantically highest argument) and Nom (the morphologically unmarked argument). The syntactic di¤erences between German and Icelandic reduce to V and I as final versus initial; they only explain why Icelandic needs to select a designated argument, which is irrelevant in German. The syntactic notion of ‘‘designated argument,’’ then, is the only notion of ‘‘subject’’ that is left under this account. An important piece of empirical evidence for the necessity to split the notion of subject into several subjects comes from the passive of ditransitive verbs and the small class of alternating inverted dat-nom verbs of Icelandic, where either High or Nom is the designated argument. For obvious reasons, the assumption of two underlying subjects that are alternatively mapped on the surface subject must be rejected. Moreover, the notion of an underlying subject would lead to inconsistencies, given the result that High and Nom are equally good concerning the designated argument but di¤er crucially as to which argument must not occur in the infinitive clause. Regarding the way the syntactic constructions react to the underlying properties of lexical items, I have shown that visibility of lexical Case is more important in German than in Icelandic, while faithfulness to the semantic hierarchy is more important in Icelandic than in German. This correlates with the fact that *Nom-Inf is more important in German, while Domain is more important in Icelandic. The main findings can be summarized by the following statements. (74) Partial rankings that hold for both German and Icelandic a. Case assignment Max(lexF) Default, Uniqueness Max(þhr,þlr) b. Local Case and infinitives LocCase, *Inf-prop *q (75) Partial rankings in which German and Icelandic di¤er (dependent infinitives and agreement) a. G Max(lexF), *Nom-Inf *q *High-Inf, ECM Domain AgrNom, *AgrNonNom, *AgrLexF *q b. I *High-Inf *q Max(lexF), Domain ECM, *Nom-Inf *AgrNonNom *q AgrNom, *AgrLexF Moreover, Icelandic instantiates Des constraints. The constraints regulating Case assignment and agreement largely di¤er from those that regulate control and raising. Thus, certain sets of constraints are construc-
The Force of Lexical Case
619
tion specific, while other constraints regulate various kinds of constructions. The most relevant constraints are SemHier (with *High-Inf as a subcase), which regulates the mapping between semantics and syntax, and Max(lexF), which regulates the mapping between lexical features and syntax. Given this characterization, semantics turns out to be more important in Icelandic than in German, while the morphosyntactic marking turns out to be more important in German than in Icelandic. Notes This chapter was initiated by Paul Kiparsky in 1992. I am grateful to Ingrid Kaufmann, Gereon Mu¨ller, Albert Ortmann, Barbara Stiebels, an anonymous reviewer, and the editors of this ´ rnason, Jo´hannes Jo´nsson, and Sigga volume for their comments, as well as to Kristja´n A Sigurjo´nsdo´ttir for their judgments. In particular, I would like to thank Jo´hanna BarDdal for drawing my attention to the class of alternating inverted dat-nom verbs in Icelandic. The research was supported by the German Science Foundation (DFG) in connection with the SFB 282 ‘‘Theorie des Lexikons’’. A preliminary version of the chapter appeared in the Working Papers of the SFB 282, No. 112 (2000, 81–102), and was discussed in Mu¨ller (2000, 285– 294), from which I highly profited. 1. Max(lexF) is a faithfulness constraint in the sense of Correspondence Theory (McCarthy and Prince 1995). I discuss below the question of what exactly constitutes the input; at least, lexical features are part of the input. A similar constraint is assumed by Woolford (2001). 2. In a framework in which the input is conceived of as a syntactic deep structure, SemHier can be replaced by a constraint called Parallel Movement (Mu¨ller 2001). 3. To put it di¤erently: Agreement seems to be a good subject criterion for German, but if one applies this criterion to Icelandic, this language turns out to have split subjects. All tests for semantic ranking seem to be good subject criteria for Icelandic, but if one applies them to German, this language turns out to have split subjects, too, but di¤erently from Icelandic (see also Seefranz-Montag 1983, 165–167). The problem of the notion of subject was first noted by Keenan (1976) and has been extensively discussed in the literature on Austronesian languages such as Tagalog (see Schachter 1976; Kroeger 1993; among others). Furthermore, the distinction between ‘‘syntactic ergativity’’ (where the nom-argument is the pivot in syntactic constructions, as in Dyirbal) and ‘‘morphological ergativity’’ (where also the highest argument can be the pivot in syntactic constructions, as in Walmatjari), stated by Dixon (1994, 160, 172), exactly points to the fact that for many languages, there is no unique notion of subject. 4. LDG deviates from Kiparsky’s proposals mainly for reasons of markedness. Kiparsky’s [þHR] ‘‘the highest role’’ is equivalent to [hr] in LDG, and his [þLR] ‘‘the lowest role’’ is equivalent to [lr]. In Kiparsky’s account, the intransitive subject is the most marked ([þHR,þLR]), while it is least marked in LDG ([hr,lr]). Complementarily, the medial argument of ditransitive verbs is the least marked ([HR, LR]) in Kiparsky’s system, while it is most marked in LDG ([þhr,þlr]). Consequently, the morphological Cases are marked for negative features in Kiparsky’s system but for positive features in LDG. (See also Kiparsky 2001b.) 5. This concept of input is similar to that of Grimshaw (1997), who considered a predicateargument structure with an identical Logical Form to constitute the input. A more detailed
620
Dieter Wunderlich
characterization of input, including person and number features, is given in Wunderlich (2001). 6. The actual ranking in (15) can be determined only if all Case patterns of German and Icelandic are considered. Both Uniqueness and Max(þhr,þlr) would su‰ce to exclude all candidates but the first one in (15); that both constraints are necessary cannot be shown on the basis of just one type of verb. In Wunderlich (2000), I argue that German and Icelandic do not di¤er with respect to Case assignment, that is, the constraints used in (15) are ranked alike for these two languages. 7. Note that Max(þhr,þlr) would su‰ce to preserve Dative in the passive. In order to also preserve Genitive, Max(lexF) is necessary. 8. It is obvious that some constraints are irrelevant for a specific type of verb. I have included all constraints to illustrate that there is one consistent constraint ranking that is responsible for all Case patterns occurring with any type of verb. 9. Woolford (2001) arrives at a similar result with a di¤erent set of constraints. Note that Woolford does not factorize the Dative into two features, which, however, allows a more detailed analysis for other types of Case patterns (Stiebels 2000; Wunderlich 2003). 10. Some speakers seem to always prefer agreement with the Nominative argument. 11. There are other verbs of Icelandic with both arguments being lexically marked, for instance, vanta ‘lack’, which shows an acc-acc or dat-acc pattern (Yip, Maling, and Jackendo¤ 1987). 12. BarDdal (1997, 1998) argues that these alternating dat-nom verbs were already present in Old Scandinavian. This explains why in a later period, some verbs were reanalyzed as canonical nom-acc verbs, while others were fixed as nonalternating dat-nom verbs. I will not go into any details here. 13. Among the various proposals in the literature to account for absolute ungrammaticality, the versions of Prince and Smolensky (1993) and Ackema and Neeleman (1998) seem the most adequate ones for the kind of syntactic problems considered here. 14. Some sentences are improved with another verb or a more heavy Dative DP. Compare (i) with (59d) and (ii) with (59e). (i) Ich sah ihm den Braten behagen. I saw him.dat the.acc roast meat please (ii) ?Ich sah diesen Film den meisten Kindern gefallen I saw this film the most children liked ‘I saw this film be liked by most of the children.’ 15. The fact that (59d) is better judged than (59e) can also be explained by means of Parallel Movement, which requires a given order of constituents to be preserved (Mu¨ller 2001); see also note 3. 16. The same is true of alternating inverted dat-nom verbs such as henta ‘please’, illustrated in (36). 17. The same is true of the two IP structures of alternating inverted dat-nom verbs.
26
Constraints on Source/Goal Co-occurrence in Carrier
William J. Poser
26.1
Introduction
This chapter addresses a puzzling restriction on the combination of goal and source motion in Carrier verbs, o¤ering a solution in terms of argument structure. 26.1.1
Restrictions on Source/Goal Combinations
In most languages it is possible to express the source and goal of motion in the same clause, as in the English I walked from home to school. In Carrier,1 an Athabaskan language of the Central Interior of British Columbia, this is not the case; two clauses are required. They may be joined by the complementizer hoh ‘while’,2 as in (1)–(3), or by the conjunction ink’ez ‘and’, as in (4)–(7).3 (1) Tacˇe ts’i hasya hoh Bincˇe ts’i łgad sya. Tachie pp I.walked.from comp Pinchie pp I.arrived.walking ‘I walked from Tachie to Pinchie.’ (2) Ł k b tgag tnabayoh tin skai hoh daibayoh ts’i fish kitchen I.carried.containerful.from comp dining room pp xeskai. I.carried.containerful ‘I carried the fish from the kitchen to the dining room.’ (3) Skehhod leh n lehne nat Saik’ z et xedig s hoh nts’ n school bus twice Saik’ z there he.sets.out.driving comp downstream k’ nad g s dzen tots k. he.drives.back.and.forth day each ‘The school bus goes back and forth between Stoney Creek and town twice a day.’ (4) Yak’ z x ladetn k ink’ez ndi y n k’einya. heaven he.left and this world he.came.onto ‘He came to earth from heaven.’
622
William J. Poser
(5) Yoh tilgai ink’ez b nk’ t et talgai. house he.ran.from and lake there he.ran.into.water ‘He ran out of the house into the lake.’ k’ nad g s dzen tots k (6) Skehhod leh n lehne nat nts’ n school bus twice downstream he.drives.back.and.forth day each ink’ez Saik’ z et xedig s. and Saik’ z there he.sets.out.driving ‘The school bus goes back and forth between Stoney Creek and town twice a day.’ (7) Y sk’ t n kes i be d ni tintoh ts’i hayang z snowmobile it by moose bush pp he.dragged.it.out ink’ez ti k’eyang z. and road he.dragged.it.onto ‘He dragged a moose out of the bush to the road with a snowmobile.’ It is not necessary for the source and goal to be overt. The same phenomenon is observed in verbs in which one directional argument is expressed by means of an adverbial prefix. In (8) the goal is expressed by the prefix ta ‘into liquid’. The same verb may instead have a source argument, as in (9), but as (10) shows source and goal may not co-occur. (8) D t’aiya talts’ t. little.bird it.fell.into.liquid ‘The little bird fell into the water.’ (9) D t’aiya d t’o behalts’ t. little.bird its.own.nest it.fell.out.of ‘The little bird fell out of its nest.’ (10) *D t’aiya d t’o behatalts’ t. little.bird its.own.nest it.fell.out.of.into.water ‘The little bird fell out of its nest into the water.’ (Other orderings of the prefixes, viz. tabehalts’t and betahalts’t, do not improve this example.) 26.1.2
Specificity to Source/Goal Combinations
The constraint is specifically on the combination of source and goal. Other thematic roles are freely combined, as in (11)–(19).
Constraints on Source/Goal Co-occurrence in Carrier
(11) Source/Instrumental/Theme Y sk’ t n kes i be d ni tintoh ts’i hayang z. snowmobile it by moose bush pp he.dragged.it.out ‘He dragged a moose out of the bush with a snowmobile.’ (12) Goal/Instrumental ˇ nlak ts’i tekeł. D ts’i be C his.own.canoe by Chunlac pp he.will.go.by.boat ‘He is going to go to Ghunlac in his canoe.’ (13) Comitative/Benefactive S łt s s ł tai ba id t’en. my.sister with.me uncle for we.2.are.working ‘My sister is working with me for our uncle.’ (14) Comitative/Locative Fabian ink’ez Ernie b ł t’enbayoh et h t’en. Fabian and Ernie with.him shed there they.are.working ‘Fabian is working with Ernie in the shed.’ (15) Comitative/Testamentary Fabian ink’ez Ernie b ł nahnał h t’en. Fabian and Ernie with.him before.us.2 they.are.working ‘Fabian is working with Ernie in our presence.’ (16) Comitative/Instrumental D g sbeyat k be b ł yast k. telephone by with.him I.spoke ‘I spoke to him on the telephone.’ (17) Theme/Benefactive/Instrumental Benen d ka be sba n nainezkai. sewing.machine by for.me she.sewed.it ‘She sewed it (torn shirt) for me with a sewing machine.’ (18) Instrumental/Testamentary/Theme Nenał sk id ne be l z be dadent’az yayany z. before.us boy hammer by window he.broke.it ‘The boy broke the window with a hammer in our presence.’ (19) Adversative/Benefactive John scˇ’a h ba yatełt k. John against.me for.them he.will.speak ‘John will speak against me, for them.’
623
624
William J. Poser
The impossibility of combining source and goal is reminiscent of the restrictions on the expression of motion and path and motion and manner pointed out by Talmy (1985). 26.2
A Possible Morphological Account
Since many of the verbs involved contain adverbial prefixes that specify a source or goal, a plausible hypothesis is that the constraint is morphological in nature. Specifically, we might suppose that the constraint is simply the result of the inability of more than one adverbial prefix to occupy the same slot. Such a proposal would be problematic since there is actually more than one position available for adverbial prefixes. This is illustrated by such examples as (9), where there are two adverbial prefixes, be- ‘inside’ and ha- ‘out’. In any case, the constraint is observed even in examples in which adverbial prefixes are not involved. Consider (20), which has no monoclausal equivalent. (20) D bez ike yiłcˇut ink’ez y tezg z her.mother.in.law her.feet she.took.it and she.began.to.drag.it hoh ny k h x t’i et ni:ning z. comp over.there they.reside there she.dragged.it.to.a.terminus ‘She took her mother-in-law’s feet and dragged her over there to where they were staying.’ (20) comes from a story about a prophet named Boba. While camped with her son, daughter, and daughter-in-law, Boba died. Her body was wrapped up in preparation for cremation and put across the fire from their camp. One day, after she had been dead for a week, her daughter-in-law noticed her body moving, brought the body back to their camp, and revived her. Later, they took her back to Nadleh (Nautley) village where she remained alive for a month and made many prophecies. Some are interpreted as predicting the coming of horses, trains, white people, noodles, and old age pensions. Whereas ni:ningz contains the adverbial prefix ni- ‘to a terminus’, ytezgz contains no adverbial prefix. Its components are y()- ‘3s disjoint object’, t- ‘inceptive’, ez ‘perfective’, and gz ‘drag’. The source in the situation results from the interaction of the meaning of the verb ‘drag’ and the inceptive. The constraint is therefore not on morphological structure. 26.3
Level of Representation
The evidence we have so far considered does not specify the level of representation to which the constraint applies. In particular, it leaves open the question of whether it applies to a true semantic representation or to argument structure. There is evidence that the latter is the case.
Constraints on Source/Goal Co-occurrence in Carrier
625
The constraint applies only to thematic roles that are syntactically visible. Sources that are incorporated into the verb do not count for the purpose of this constraint. Some body parts may be incorporated into the verb. Such incorporated nouns may not be modified and are otherwise not available to the syntax. In (21) nak’e ‘eye socket’ is incorporated into the verb ‘to flow’. An additional piece of evidence that it is incorporated is that the form used in contexts other than such incorporations is nak’et.5 (21) X z snak’ehaindli. pus it.eye.socket.flows.out.of.me ‘Pus is flowing out of my eyes.’ An example in which such an incorporated source co-occurs with an explicit goal is (22). uzahayandli. (22) Y n s bal k’ t uzek carpet onto his.saliva it.mouth.flows.out.him ‘He is drooling onto the carpet.’ The expression ‘he is drooling’ literally means ‘his saliva is flowing out of his mouth’. The verb is ‘to flow’, with the incorporated postposition ha ‘from, out’, and the incorporated noun za ‘mouth’. In this case the object marker is the third person singular u. One piece of evidence that ‘mouth’ is incorporated is that za is not the normal form of mouth. Except in such incorporations, ‘mouth’ is -zek. The reason that the goal may be mentioned in (22) is the source is not syntactically visible, just as in English, a source ‘mouth’ is implicit in ‘to drool’ but is neither overtly nor syntactically present. Notice that in (20) we had to appeal to an implicit source, which appears to contradict the claim just made about (22). My suggestion is that there is a di¤erence between ‘drag’ and ‘drool’. In ‘drag’, although the source may not be explicitly expressed, it is part of the argument structure of the verb. That is, there is an agent that does the dragging, a theme that is dragged, a source from which it is dragged, and a goal to which it is dragged. In contrast, ‘drool’ has no source role at the appropriate level of representation. 26.4
The Thematic Structure of Carrier Motion Verbs
I suggest that the restriction on co-occurence of source and goal in Carrier verbs is not arbitrary but is the result of a general property of the thematic structure of Carrier motion verbs. My proposal is that Carrier motion verbs are never neutral; they always describe motion with respect to some reference point. By default, this is a goal. Morphology may further specify this goal or may change it into a source, but it may not add reference points.
626
William J. Poser
Evidence for this proposal may be found by examining the simplest, most neutral possible Carrier motion verbs, namely verbs with no meaningful thematic prefixes in the progressive aspect. Such progressive forms are used to describe what at first glance appears to be pure motion with no reference point. For example, if you call someone on your cell phone from a boat and he or she asks you how you are traveling, you would say skeł ‘I am going by boat’. This consists of the progressive aspect stem keł and the first person singular subject prefix s; the is epenthetic.6 Nonetheless, even such verbs have an intrinsic orientation to a destination. Thus, (23), in which the verb is combined with a destination, is grammatical, but (24), with an origin rather than a destination, is ungrammatical. (23) Tacˇe ts’i hekeł. Tachie pp they.are.going.by.boat ‘They are en route to Tachie by boat.’ (24) *Tacˇe x ts’ n hekeł. Tachie from they.are.going.by.boat ‘They are en route from Tachie by boat.’ A similar observation may be made regarding the ‘‘handling verbs.’’ The handling verbs are a set of verbs used to describe handling of objects of various types. They form the basis for one of the several systems of noun classification in Carrier. Handling verb roots merely indicate the type of object handled. The precise way the object is handled is determined by the prefixes added to this root. In (25) we have a number of examples of verbs for handling two-dimensional flexible objects. In each case, comparable verbs exist for handling other classes of objects. (25) Some handling verbs Example
Translation
behanaitełcˇ s daid tełcˇ s d gaid tełcˇ s huk’eitałcˇ s huk’enayitałcˇ s hanaitełcˇ s sgaitełcˇ s n yitełcˇ s atełcˇ s tatełcˇ s natełcˇ s yaiy tełcˇ s
he he he he he he he he he he he he
is going is going is going is going is going is going is going is going is going is going is going is going
to to to to to to to to to to to to
take it out hold it up hang it up put it on (the table) put it back on (the table) bring it back give it to me carry it around bury submerge put on the ground bring it ashore
Constraints on Source/Goal Co-occurrence in Carrier
627
When a handling verb root is used with no thematic prefixes, it means ‘carry object to’, as exemplified in (26). (26) Tsetseł ts zbayoh ts’i yeał. axe woodshed pp he.is.bringing.it ‘He is bringing the axe to the woodshed.’ Here the verb yeał is the progressive aspect of the verb for handling single default objects. It contains two overt prefixes. One is the singular disjoint reference third person object marker y. The other is the progressive aspect marker e. The third person singular subject marker is merely a requirement that there be a vowel in a certain position; this requirement is satisfied by the progressive aspect marker. The verb contains no prefixes that indicate the direction of motion, and as we have seen, the postposition ts’i is directionally neutral. The fact that the motion is to the woodshed must therefore be the result of the default thematic role. Further evidence for a default directional role comes from the system of directional prefixes. Carrier has a number of prefixes that are usually translated as specifying motion in a certain direction. These include the following. (27) Directional prefixes Prefix
Meaning
Example
Translation
a be da na ta t’ ts tsa ya
into a hole inside (container) inside (via portal) to the ground into liquid into a pocket into fire into mouth ashore
alts’ t beinlis daninya nalts’ t talts’ t t’ ł cˇuz ts danla tsanad t’aih yas lat
he fell into a hole pee into it he entered he fell down he fell into the water I pocketed (2-D flexible) he put (plural) into fire he is snacking it floated ashore
A striking fact is that there are no comparable prefixes with opposite meanings, such as ‘out of a hole’, ‘out of liquid’, ‘away from the ground’. Instead, these meanings are expressed by combining one of the ‘into’ prefixes with the prefix ha ‘from, out of ’. This is exemplified by (28)–(33). (28) t’ desd tan ‘I put (long rigid object) into my pocket’ (29) t’ hadesd tan ‘I took (long rigid object) out of my pocket’
628
William J. Poser
(30) tayantan ‘he put it (long rigid object) into liquid’ (31) tahayantan ‘he took it (long rigid object) out of liquid’ (32) dadelge ‘he inserted his finger’ (33) dahadelge ‘he withdrew his finger’ This suggests that there is no inherent directionality to these prefixes. The appearance thereof results rather from the fact that the direction argument of motion verbs defaults to a goal. The prefix ha, which explicitly marks the argument as a source, may be used to override and reverse the default. Further evidence of the lack of intrinsic directionality is provided by (34)–(36), all of which refer to wearing clothing on the torso, such as pants, shirts, dresses, and skirts. (34) cˇaim ntl’as s bes sda. jeans I.am.wearing ‘I am wearing jeans.’ (35) cˇaim ntl’as s benasja. jeans I.put.on ‘I put on jeans.’ (36) cˇaim ntl’as s behanasja. jeans I.took.o¤ ‘I took o¤ jeans.’ (34) literally means ‘I am sitting within’ and contains the postposition be, the incorporated counterpart of the independent postposition bet. Here it clearly has no directionality to it. (35) literally means ‘I walked within’. It consists of the same incorporated postposition with the underlying verb ya ‘for one person to walk on one pair of limbs’. be is now associated with a goal argument. (36) literally means ‘I walked out of ’. It has the same components as (35) with the addition of the prefix ha, which has the e¤ect of converting the goal argument into a source. These examples show that be has no intrinsic directionality. When part of a stative verb, it has no associated directionality. When part of a motion verb, it becomes associated with a goal unless the default is overridden by the use of ha, in which case the goal is replaced by a source. The other clothing verbs, the dual and plural verbs for wearing clothing on the torso and the verbs for wearing clothing on the hands, the feet, and the head, all have the same structure and make the same point about the semantics of be.
Constraints on Source/Goal Co-occurrence in Carrier
26.5
629
Conclusion
In Carrier, it is impossible for the source and goal of motion to be specified in the same clause. This constraint applies at the level of argument structure. This results from the fact that Carrier motion verbs have a single directional argument position. This is a goal by default but may be changed into a source by suitable morphology. Notes 1. The data in this chapter are in the Saik’z (Stoney Creek) dialect, a member of the FraserNechako subgroup of the southern branch of Carrier. Although I have not checked every detail, the facts do not appear to be materially di¤erent in the Nak’albn/Dzingbn (Stuart/ Trembleur Lake) dialect, to which most of the literature on Carrier is devoted. I am grateful to Dr. Mary John, Sr., C.M., for the bulk of the data, and to the late Veronica George for example (20). The transcription used is phonemic North American IPA. Underscores distinguish laminodental fricatives and a¤ricates from apicoalveolars. 2. The gloss ‘while’ is only approximate. This complementizer is used in a broad range of situations in which the events of the two clauses overlap in time. 3. In several of the following examples, the word ts’i is glossed ‘pp’. Although most often it can be glossed ‘to’, this is an underspecified directional postposition and when the verb calls for it may translate English ‘from’. Other abbreviations are comp ¼ complementizer, 2 ¼ 2nd person. 4. The / / is epenthetic and so, not properly speaking part of the prefix. 5. Since body parts are inalienably possessed one cannot really speak of a free form. However, ‘eye socket’ actually does have a free form of sorts since it also means ‘lenticel on birch bark’, which is not inalienably possessed. 6. In this form the progressive aspect is marked only by the stem. The progressive aspect prefix e seen below in the third person plural form hekeł is realized only when preceded by a prefix belonging to the class known to Athabaskanists as conjunct prefixes.