Restrictiveness in Case Theory (Cambridge Studies in Linguistics)

Henry Smith here develops a theory of syntactic case and examines its synchronic and diachronic consequences. Within a unification-based framework, the book draws out pervasive patterns in the relationship between morphosyntax ("linking") and grammatical function. The theory proposed consists of three ordered constraints on the association of NPs and arguments, based on the central notion of "restrictiveness." Beginning with a detailed study of dative substitution in Icelandic, the author moves on to examine a wide array of synchronic and diachronic data and to construct a typology of case. Theoretically innovative and sophisticated, and descriptively wideranging, this book will appeal to all those interested in the crosslinguistic marking of case and the ways in which case systems may change over time.

CAMBRIDGE STUDIES IN LINGUISTICS General Editors: s. R. ANDERSON, J. BRESNAN, B. COMRIE, W. DRESSLER, R. HUDDLESTON, R. LASS, D. LIGHTFOOT, J. LYONS, P. H. MATTHEWS, R. POSNER, S. ROMAINE, N. V. SMITH, N. VINCENT

Restrictiveness in case theory

In this series

52

53 54 55 56

MICHAEL s. ROCHEMONT and PETER w. CULICOVER: English focus constructions and the

theory of grammar PHILIP CARR: Linguistic realities: an autonomist metatheory for the generative enterprise EVE SWEETSER: From etymology to pragmatics: metaphorical and cultural aspects of semantic structure REGINA BLASS: Relevance relations in discourse: a study with special reference to Sissala ANDREW CHESTERMAN: On definiteness: a study with special reference to English and Finnish

57

ALESSANDRA GIORGI and GIUSEPPE LONGOBARDI: The syntax of noun phrases: configura-

62

STEPHEN R. ANDERSON: A-Morphous Morphology

69

R. M. w. DIXON: Ergativity

74 75 76

ALICE c. HARRIS and LYLE CAMPBELL: Historical syntax in cross-linguistic perspective LILIANE HAEGEMAN: The syntax of negation PAUL GORRELL: Syntax and parsing

tion, parameters and empty categories 58 MONIK CHARETTE: Conditions on phonological government 59 M. H. KLAIMAN: Grammatical voice 60 SARAH M. B. FAGAN: The syntax and semantics of middle constructions: a study with special reference to German 61 ANJUM p. SALEEMI: Universal Grammar and language learnability

63 LESLEY STIRLING: Switch reference and discourse representation 64 HENK J. VERKUYL: A theory of aspectuality: the interaction between temporal and atemporal structure 65 EVE v. CLARK: The lexicon in acquisition 66 ANTHONY R. WARNER: English auxiliaries: structure and history 67 p. H. MATTHEWS: Grammatical theory in the United States from Bloomfield to Chomsky 68 LJILJANA PROGOVAC: Negative and positive polarity: a binding approach 70 YAN HUANG: The syntax and pragmatics of anaphora 71 KNUD LAMBRECHT: Information structure and sentence form: Topic, focus, and the mental representations of discourse referents 72 LUIGI BURZIO: Principles of English stress 73 JOHN A. HAWKINS: A performance theory of order and constituency

77 GUGLIELMO CINQUE: Italian syntax and Universal Grammar 78 HENRY SMITH: Restrictiveness in case theory

Supplementary volumes MICHEAL o. SIADHAIL: Modern Irish: grammatical structure and dialectal variation ANNICK DE HOUWER: The acquisition of two languages from birth: a case study LILIANE HAEGEMAN: Theory and description in generative syntax: A case study in West Flemish A. E. BACKHOUSE: The lexical field of taste: A semantic study of Japanese taste terms NIKOLAUS RITT: Quantity adjustment: Vowel lengthening and shortening in early Middle English

Earlier issues not listed are also available

RESTRICTIVENESS IN CASE THEORY HENRY SMITH

CAMBRIDGE

UNIVERSITY PRESS

CAMBRIDGE UNIVERSITY PRESS Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, Sao Paulo Cambridge University Press The Edinburgh Building, Cambridge CB2 2RU, UK Published in the United States of America by Cambridge University Press, New York www.cambridge.org Information on this title: www.cambridge.org/9780521462877 © Cambridge University Press 1996 This publication is in copyright. Subject to statutory exception and to the provisions of relevant collective licensing agreements, no reproduction of any part may take place without the written permission of Cambridge University Press. First published 1996 This digitally printed first paperback version 2006 A catalogue record for this publication is available from the British Library Library of Congress Cataloguing in Publication data Smith, Henry. Restrictiveness in case theory / Henry Smith. p. cm - (Cambridge studies in linguistics) Includes bibliographical references. ISBN 0 521 46287 8 (hardback) I. Grammar, Comparative and general - case. I. Title. II. Series. P240.6.S64 1996 415-dc20 95-32776 CIP ISBN-13 978-0-521-46287-7 hardback ISBN-10 0-521-46287-8 hardback ISBN-13 978-0-521-02655-0 paperback ISBN-10 0-521-02655-5 paperback

For Sun-Joo

Contents

A cknowledgemen ts

page xi

1 1.1 1.2 1.3

Introduction Preliminaries Theoretical framework Overview

1 1 3 20

2 2.1 2.2 2.3 2.4 2.5 2.6

Argument case and case alternations Case alternations and the problem of case Case, argument structure, and interpretation Basic clause structure Case alternations and restrictiveness Why argument structure? Conclusion

22 22 31 48 51 60 69

3 3.1 3.2 3.3 3.4 3.5

A typology of case systems Two accusatives A typology of case Four types of case system Split ergativity Three-case systems and quasi-split ergativity

70 70 74 86 115 118

4 4.1 4.2 4.3 4.4

Linker interactions Other elsewhere patterns Passive and antipassive Word order Obliqueness and compatibility of case and configuration Comnatibilitv of linkers

130 130 146 154

4.5

158 166

x

Contents

5 5.1 5.2 5.3 5.4 5.5

Icelandic Case, agreement, and word order Case preservation and nominative objects Is Icelandic symmetric or asymmetric? Coordination More on the lexical accusative

171 171 180 194 208 213

6 6.1 6.2 6.3 6.4 6.5 6.6

Changes in linking Previous work on linking change Diachronic aspects of case alternations A theory of analogical change in linking Acquisition of termhood Scandinavian Conclusions

221 221 226 233 236 245 255

7 7.1 7.2 7.3 7.4 7.5

Case semi-preservation Case (non-)preservation "Scattered cases" and semi-preservation Further implications Diachronic predictions Conclusion

259 259 274 282 287 288

8

Conclusions

289

Notes References Index

293 307 317

Acknowledgements

The present study grew out of my Ph.D. dissertation in Linguistics at Stanford University. I would like to thank Paul Kiparsky, Rob Robinson, Elizabeth Traugott, and Tom Wasow for their support and valuable advice throughout this project. Without the encouragement of Elizabeth Traugott and Tom Wasow the dissertation would never have turned into this book. I also have many people to thank for their useful suggestions, including Thora Arnadottir, John Ole Askedal, Joan Bresnan, Michael Barnes, Chris Culy, Cathrine Fabricius-Hansen, Vjacheslav Ivanov, Tracy King, Paul Kroeger, Andreas Ludwig, Joan Maling, Chris Pinon, Peter Sells, Hoskuldur Thrainsson, Steve Wechsler, and Annie Zaenen. Thanks also to Ted Andersson for his inspiring courses on Germanic languages and to Michelle Murray, Gertrud Pacheco, Emma Pease, and Gina Wein for their generous assistance during graduate school. I would also like to give thanks to two anonymous reviewers from Cambridge University Press for their valuable comments and to Judith Ayling, Hilary Gaskin, Kay McKechnie, and Joanna West for their help in the publication of this book. I am grateful to Stanford University, the Fulbright Foundation, and the Center for the Study of Language and Information for financial support and to my students and colleagues at Indiana University, Bloomington, for a congenial setting for continuing this project. I also wish to thank my parents. Most of all, I thank my wife, Sun-Joo, for being the best possible emotional and intellectual companion. Material from my article "'Dative Sickness' in Icelandic," Natural Language and Linguistic Theory 12 (1994) 675-736, appears in Section 1.2 and scattered throughout Chapter 2.

XI

1

1.1

Introduction

Preliminaries

The study of case has many facets, and the word "case" itself has come to mean many things. This study aims at a better understanding of the way case functions in syntax by providing a new theory of the syntactic functioning of case and other morphosyntactic devices. Because the focus is on case as it functions in syntax rather than case as a morphological category, I will make certain assumptions which will figure prominently below. First, I will assume that morphological case and syntactic case do not always coincide. Morphological case will simply be the paradigms of affixes which may carry other information as well, e.g. gender and number. For example, the "dative" case as a morphological category is a set of endings. My primary concern is with syntactic case, and here the distinctions are not quite so visible. A syntactic case will be defined by its basic distribution and its interaction with other cases. For instance, if we observe that dative marks the goals of various verbs - what might be termed "indirect objects" - we have a situation very common in the world's languages: I(NOM) gave you(DAT) the book(ACc). But quite often we then notice that the dative marks an NP that is an experiencer in a sentence based on a two-place predicate: I(DAT) like the book(NOM). At this point we have a choice. We can assert that the " I " NP in such a sentence is the "same" in some sense as the "VOU(DAT)" in the sentence with the three-place predicate. The "sameness" might be identity of "grammatical relation," for example indirect object (or 3 in Relational Grammar). Another ("localist") approach would be to seek some shared semantic characteristics, making them both the endpoint of some motion, either physical or metaphorical. Either way, the appearance of a morphological dative is coextensive with some more abstract category. 1

2

Introduction

The problem becomes more difficult when we find isolated verbs taking this same morphological case, where we expect another case, e.g. accusative: I(NOM) lost the book(DAT). We can try to save the match between dative and our abstract categories like indirect object or experiencer, but only at a price. If we say that such a verb (Icelandic has many such verbs) selects an indirect object, then we are well on our way to reducing indirect object to the category of morphological dative. Saying that dative is assigned to indirect objects approaches being a tautology. However, if the distinction between case as a morphological category and case as a syntactic category is made, then we have another option, taken in this work. One morphological case can correspond to multiple syntactic cases. This is like saying that we have various "rules" for assigning dative, one for "indirect objects" (or whatever category is appropriate for datives in ditransitives), and one for "experiencers". And verbs can assign dative in their lexical entries, the limiting case of a rule that applies to one item. This distinction is an interesting one to the extent that it leads to the discovery of new phenomena. One of the primary motivations of this study is to show that this is so. Furthermore, if the definition of syntactic cases involves, as I will argue, a consideration of their basic distributions and effects on the syntax of the clause, then it is proper to consider other morphosyntactic devices as forming similar syntactic categories. It is well-known that agreement and/or a particular position, e.g. sentence-initial position, correlate with what we might want to call subjecthood. That is, NPs that are special in a sense to be described shortly, are often the target of agreement or occur in a special position. For languages that have such behavior, I will treat agreement and word order as syntactic "case" categories which do not correspond to a morphological case. All such syntactic "cases" will be called "linkers." I will use "case" and "linker" interchangeably, except that I will use "linker" as the general term whenever I wish to draw attention to the distinction between case proper and linkers generally. What they all have in common is that they "link" mediate between in a sense to be described shortly - arguments on the one hand and surface NPs on the other. Predicting surface form based on the number and type of NP complements a verb selects is not the only goal. I will argue, in the tradition of direct linking (about which more in the next section), that grammatical relations like "subject o f and "object o f can be defined on the order of the verb's complements and the nature of the linking they undergo. To

Theoretical framework 3 this end one distinction often made is crucial here. Subjecthood properties come in two varieties, coding and behavioral (Keenan (1976)). Coding properties are nominative case in some languages, preverbal position in some languages, etc. Behavioral subjecthood properties are the ability to participate in certain syntactic constructions in certain ways, for example to be the target of control, to be the antecedent of certain kinds of reflexive pronouns, etc. One of the main claims in the following is that the second type of property, the participation in characteristic constructions, is predictable on the basis of linking, i.e. coding properties. From the argument structure of a verb and from linkers like nominative and dative and their nature, we will be able to derive the grammatical function behavior of the various NPs. A final goal of the present study is to provide a theory that can begin to account for another type of case behavior, the change of cases over time. Once again, I will focus on changes in linking rather than purely morphological changes in sets of endings. Of course I will not take the history of linking in any language to be synchronically represented as part of any speaker's competence. I will, however, try to show that the theory developed to account for syntactic case behavior has diachronic implications which are supported by a wide range of data.

1.2

Theoretical framework

In the following chapters I will develop a theory of case within a framework based on two traditions, direct linking and the extended categorial framework. Since the second is more widely known, this section will focus on direct linking and on which aspects of it will be adopted below. 1.2.1 What is linking? Generally speaking, a theory of linking deals with generalizations involving argument structure (possibly including thematic structure), grammatical functions (subject, object, oblique), and morphosyntactic expression (case marking, word order, and agreement). A very simple linking theory for English based on traditional grammar would be as follows: (1)

a. NPs that denote agents bear the grammatical relation subject, b. NPs that denote patients bear the grammatical relation direct object.

4

Introduction c. NPs that denote beneficiaries or recipients bear the grammatical relation indirect object.

(2)

a. Subjects normally occur in preverbal position. b. Indirect objects occur immediately after the verb. c. Direct objects occur within the verb phrase.

Latin, on the other hand, being a language with the nominative-dativeaccusative system alluded to in the previous section, would share the rules in (1) with English but would have (2') instead of (2), which would account for the distribution of cases in a ditransitive like (3): (2')

a. Subjects are marked nominative and the verb agrees with the subject. b. Direct objects are marked accusative. c. Indirect objects are marked dative

(3)

Magister pueris librum dat. teacher(NOM) boys(DAT) book(Acc) gives 'The teacher gives the boys the book.'

The rules in (1) express generalizations about the semantic content of NPs, and the rules in (2) and (2 ; ) state generalizations about the expression of NPs. Notice that mediating between notions like agent and patient on the one hand and morphosyntactic categories like nominative and dative on the other are the grammatical relations. Morphosyntactic expression is assigned to grammatical relations which in turn are correlated with thematic relations. The above toy theory does of course contain many inadequacies. Under (1), we have two problems. First, the rules are too strong and too weak. It is not true that agents are always subjects: for example, in the passive the patient is the subject and the agent is expressed in a byphrase, if at all. Nor is it true that subjects are always agents, as shown by sentences like John died. So just reversing the direction of the generalizations (Subjects are agents, etc.) above would not work. The second problem is the nature of the thematic relations themselves. Providing a satisfactory definition of "agent" or "patient" has been very difficult indeed (see Dowty (1991) for a thorough discussion of this problem). At this point we are faced with two choices. One is to use thematic relations provisionally in the hope of discovering an adequate definition

Theoretical framework

5

later. The second is to try to do without them. This is the route taken in the following work (see §1.2.3 and Chapter 2 below). The term "linking" is applied to a theory that deals with generalizations like those in (1) or those in (2). A theory of linking is said to be "direct" in the sense of Kiparsky (1987) if the mapping between arguments (for Kiparsky, thematic roles) and morphosyntactic form is direct rather than mediated, as in (1) and (2), by grammatical functions like subject and object. Rather, the mapping between arguments and morphosyntactic expression is primary, and grammatical relations can be defined as the result or output of the linking itself. Very roughly, mediated and direct linking can be schematized as in (4):1 (4)

mediated linking: arguments —> grammatical functions —• morphosyntax direct linking: arguments —> morphosyntax —• grammatical functions

In mediated linking, some rule (or the equivalent) assigns grammatical functions to arguments and then case is assigned to grammatical functions. In direct linking, rules assign case and other morphosyntactic categories to arguments directly and grammatical functions can be assigned to the resulting case-marked NP. In fact, as we will see, this assignment is hardly an assignment at all if one defines grammatical functions in terms of the result of the linking (the assignment of morphosyntactic expression to the arguments). 7.2.2 Background to this study: direct linking The direct linking approach has its antecedents in Panini's theory of karakas and, more recently, in Case Grammar (Fillmore (1968)) and in the theory of Ostler (1979). Since I will be using the direct linking framework and since the theory to be proposed shortly is most closely related to that of Kiparsky (1987), I will present the basics of Kiparsky's Linking Theory (henceforth KLT) in this subsection. In the next subsection I will outline which aspects of direct linking will be incorporated into the present work. Since it is a direct linking theory, KLT assumes that morphosyntactic expression is assigned directly to arguments. In Kiparsky (1987) these arguments are taken to be thematic roles, i.e. arguments with labels that reflect lexical semantic classifications. Arguments are classified into thematic roles by the entailments that the argument is involved in

6

Introduction

(Ladusaw and Dowty (1987), Dowty (1989)). Thus an argument is an experiencer if a positive sentence with the verb entails that the entity corresponding to that argument is sentient and aware of something. Leaving aside for now the problem that such definitions pose, the resulting thematic roles are then organized into a hierarchy (many versions exist; Jackendoff (1972) is an especially famous one). KLT assumes the partial hierarchy in (5): (5)

( agent/source ( goal (instrument (theme (locative ( verb ))))))

The argument structures of individual verbs will be a list of thematic roles obeying the order in (5), i.e. the argument structure of a given verb is a portion of this hierarchy. For example the verb kick will have the argument structure in (6): (6)

( agent (theme ( kick )))

Roles lower on the hierarchy are closer to the verb (combine with the verb last), as indicated by the complex bracketing in (5). Thus in (6), the claim is that the theme is more closely related to the verb than the agent. Kiparsky (1987) claims support for this view from (i) unmarked word order and (ii) interpretational dependencies. First, if the thematic hierarchy defines the order in which arguments are semantically combined with their predicates (Kiparsky (1987: 34)), then the "unmarked" word order should be determined by the hierarchy. The idea is that the order that reflects the combining of the arguments with the verb will show up on the surface, provided that no grammatical constraints or extragrammatical factors intervene. Language-specific grammatical factors do indeed intervene in strict word-order languages like English, but the hypothesis seems to be supported in some well-known languages like German and Japanese. However, the proviso concerning grammatical or extra-grammatical factors is a major one, making it very unclear whether there is a cross-linguistic unmarked word order, as Kiparsky (1987: 34) admits. The second piece of evidence for the thematic hierarchy is from interpretational dependencies between the arguments of a verb. The claim is that the interpretation of higher thematic roles often depends on the result of composing the predicate with lower roles but not vice versa. Take the verb keep as an example. When the agent is combined, the presence or absence of a source (e.g. from Bill) determines whether the agent is entailed to "possess" the theme or not. So in John kept the

Theoretical framework

7

money, John is entailed to have the money, whereas in John kept the money from Bill, John is not entailed to have the money. A related piece of evidence for the hierarchy comes from idioms. Marantz (1984) argued for a logical subject/logical object asymmetry on the basis of the greater frequency of object idioms (idioms with a "frozen" object, e.g. kick the bucket) relative to the frequency of subject idioms (idioms with a "frozen" subject, e.g. the devil take . . . ). Kiparsky (1987: 35f.) makes the stronger claim that frequency of idiom type correlates with how far down the thematic hierarchy the idiom type's "frozen" thematic role is. So we should expect most verb-locative idioms (throw X to the wolves, etc.) and fewest verb-agent idioms. This is difficult to show, even for one language, and the list in Kiparsky (1987: 36) shows that verb-locative, verb-theme and verb-locative-theme idioms are common, while verb-theme-goal, verb-goal, verb-agent, etc. are scarce. This is a first step, but all this really shows is a major divide between locative and theme on the one hand and goal and agent on the other. The evidence from idioms is suggestive, but it cannot be taken as a justification of distinctions as fine-grained as those of even the partial thematic hierarchy above. Regardless of whether one accepts the thematic roles approach to argument structure, KLT is a theory of linking and so covers argument relations and morphosyntactic expression but does not extend to WhMovement. This makes sense in terms of Government and Binding Theory (GB) (Chomsky (1981)) as well, since Wh-Movement is never movement to get Case. Rather, linking is concerned with what in GB would be handled through the lexicon, Theta Theory, and Case Theory. Since KLT deals with "NP-Movement" phenomena in a very different way from the GB treatment (KLT has no NP-Movement, as we will see), and Wh-Movement falls outside the area of linking, Kiparsky (1987) adopts the lexical structure of Hale (1983) as the locus of linking (lexical structure is similar to NP-Structure of Riemsdijk and Williams (1981)). Lexical structure consists of trees, and (like NP-Structure in Riemsdijk and Williams (1981)) it corresponds to what in GB would be the structure after NP-Movement and before Wh-Movement (this is not a level in conventional GB: it would be an intermediate stage of the derivation from D-Structure to S-Structure). Since Kiparsky makes these assumptions, his definition of linking is:

8 Introduction (7)

Linking is a three-place relation between a 0-role, an argument (at lexical structure), and a morphosyntactic linker.

A linker is a morphosyntactic device such as a case, a word-order position or an agreement marker, i.e. a "syntactic case," as described in the last section. Linking rules will establish relations of the type in (7). So for instance, the rule "goal gets dative" will establish that an NP argument at lexical structure marked dative will be associated with the goal 0-role of the verb of the clause: (8) < agent NOM


1

DAT 1

Magister

1

N

puerls N

NP

NP

< theme < d a t > » 1 1 ACC

librum

dat

N

1

V

NP

This is the result of linking for the Latin clause in (3) above. Note that the morphologically dative argument at lexical structure is linked via the dative morphosyntactic category (the dative linker) to the goal at thematic structure. Note further the caret sign next to the goal. This denotes "demoted," i.e. syntactically oblique.2 All this means is that the indirect object in Latin is a syntactical oblique, i.e. it does not show any of the behavioral subjecthood or objecthood properties mentioned in the last section and examined more closely in the next chapter. To get this complete result though, two other linkings (for the agent and theme) had to be completed. To get (8) it will be necessary to go into more detail about linking in KLT. Kiparsky (1987: 37), following Ostler (1979), distinguishes two kinds of linking, grammatical and semantic: (9)

a. Grammatical linking: agreement morphology, "grammatical" cases or adpositions, case, or position under verb government. Grammatical linking operates on the projection of (non-

Theoretical framework

9

demoted) roles in the 0-structure. Grammatical linking is of two types, unrestricted and semantically restricted (i.e. limited to a given 0-role). b. Semantic linking: role assigners such as semantic ("oblique," "adverbial") cases or adpositions, and serial verbs. Any 0-role is eligible for semantic linking. Roughly speaking, the distinction between grammatical and semantic linking corresponds to the term/non-term distinction in Relational Grammar, since the significance of the grammatical/semantic linking distinction lies in the definition of grammatical functions, which, it should be recalled, are defined on the result of linking. The definitions are as follows: (10)

a. b. c. d.

Term: any grammatically linked argument3 Oblique: any non-term Subject: highest term of a predicate 4 Object: non-subject term of a predicate

Thus in the Latin in (8) above, the NP magister 'teacher' is linked by a grammatical linker, verb agreement, so this NP is said to be grammatically linked. It is the thematically highest grammatically linked NP and so, since it is the highest non-oblique, it is by definition the subject. An oblique like the dative NP in (8) is linked to a demoted role. Since the role is demoted, only semantic linking is possible. And, since the NP is semantically rather than grammatically linked, it is an oblique rather than a term. The significance of the definitions is even clearer in the passive. KLT, like many current theories (including GB and the Lexical Mapping Theory of recent Lexical Functional Grammar (Bresnan and Kanerva (1989)), takes passive to result from the demotion of the "logical subject," i.e. the thematically highest argument. For the verb do 'give' in (3) and (8) above, that argument is the agent. For the passive in (11), the passive rule in (12) is used to derive the argument structure in (13): (11)

Liber pueris a magistro datur. book(NOM) boys(DAT) by teacher is-given 'The book is given to the students by the teacher.'

(12)

(9 ...) -> (9* ...)

(13)

(agentA (goalA (theme ( datur ))))

10

Introduction

Again, the caret sign denotes "demoted," i.e. unavailable for grammatical linking (oblique). Thus if the passive verb is the basis for linking, we cannot use any grammatical linker, including verb agreement, to link the agent as we did in the active version above. The demoted agent will be available only for semantic linking, by the preposition meaning 'by': (14)

< agentA
DAT

Liber

pueris

< theme


AGR,N0M

a

magistro

datur

N N

N

NP

NP

P

NP PP

Note that the goal, if it had been grammatically linked, would have been the new subject, and, for example, in English where the goal is nonoblique, it could be the subject (The students were given the books). In Latin, though, the goal is demoted, so that the highest role available for grammatical linking in the passive is the theme. It is linked by agreement, a grammatical linker, and so is the thematically highest grammatically linked 9-role, i.e. the subject. Thus in both active and passive subjecthood falls out of the linking itself. So far we have not been explicit about how the linking is done, i.e. which NP arguments and morphosyntactic devices to match up with which arguments. This will be the major topic of this book, and, although our conclusions will differ significantly from those of Kiparsky (1987), it will be useful to spell out the positions taken in Kiparsky (1987) beforehand. First of all, linking is subject to several constraints, which are the rough analogues in Kiparsky (1987: 37) of the Case Theory and Theta Theory of GB: (15)

a. One-to-one: Every linker must link a unique argument to at most one unique 6-role.

Theoretical framework

11

b. Obligatoriness:

(i) Lexically realized arguments must be linked, (ii) Terms must be linked (parametric). c. Priority clause: (i) Grammatical linking has priority over semantic linking. (ii) Inner 6-roles have priority over outer 9roles.

The condition in (15a) disallows one NP expressing more than one 9-role and disallows a 9-role's being expressed by more than one NP (cf. the Biuniqueness Condition in GB). The Obligatoriness Condition Part (i) disallows unlinked NPs, i.e. linking is a precondition on the occurrence of an NP at the surface. This is similar in effect to the Case Theory of GB which disallows phonetically realized NPs from being Caseless. The Priority Clause deals with the relative priority of the linking of certain 9-roles over other 9-roles (ii) and some linkers over other linkers (i). In other words, if a given 9-role is eligible for both a grammatical linker and a semantic linker, the 9-role must be grammatically linked. We have already seen a case where this clause operates: in the active clause in (8) above, the logical subject, since it is not demoted, is eligible for the grammatical agreement linker and, since it is an agent, it is always eligible for the preposition a 'by' (this is a semantic linker for agents). In this case the agreement linker has priority by (15c)(i). In the passive, however, the agent is demoted and so is available only for semantic linking. Finally, in any direct linking theory, the inventory of linkers and the conditions on their distribution are paramount. Since the proposals of Kiparsky (1987) will be taken as a starting point for developing an alternative, I begin with Kiparsky's (1987, 1989) analysis of English. "Grammatical relations," whether primitive or derived, correlate with word-order position in English. This effect can be achieved in direct linking by means of the rules he proposes for English: (16)

Grammatical Linkers (Kiparsky (1989)) a. Agreement and _V": unrestricted b. V_ : unrestricted c. V'_ : Theme

(17)

Semantic Linkers (Partial List, Kiparsky (1987: 14)) a. by_ : highest 9-role5 b. to_ : Goal, Allative

12

Introduction c. d. e. f. g.

into_ : Illative in_ : Inessive at_ : Adessive with_ : Instrumental, Comitative, Non-affected Theme V_ : Benefactive, Experiencer

The rules (linkers) in (16) and (17) are in a format where a symbol or symbols representing the morphosyntactic category is given to the left of a colon, and the restrictions, if any, on which 0-roles the linker may link are given to the right of the colon. If multiple roles appear here, the rule is to be read disjunctively, e.g. (17b) is "Goal or Allative." Roles like allative, illative, inessive, and adessive are subvarieties of locatives. The linkers are somewhat like abstract Cases in GB with surface form more fully specified (in GB it is possible for NOM to stand for no morphological ending) and its assignment possibilities spelled out.6 Since V" is the second verbal projection ( = VP), the symbol " _ v " " means VP-external position (cf. NP-of-S). The symbol "V_" means sister of the verb, and the symbol " V i " means sister of the first projection above the verb, all in accordance with the phrase structure rules in (18): (18)

a. S - • V" N " P " b . V" -> V N " P " c. V' -> V N " (Prt)

Note that sister-of-V is both a grammatical linker for any role (16c) and a semantic linker for benefactives and experiencers (17g). Although it is not reflected in these rules he proposes, Kiparsky asserts that agreement always links the highest available (non-demoted) role, making agreement an exception to the Priority Clause Part (ii) in (15c)(ii). To support this, Kiparsky cites the diachronic development of agreement from anaphors, and the fact that anaphors often require a subject antecedent of some sort. Kiparsky assumes that the anaphoric property of "subject orientation" is synchronically still present in agreement markers. Since he asserts that agreement "inherently" seeks the highest available role, it is not stipulated in the agreement rule.7 Kiparsky further assumes that the VP-external position is associated with agreement and so is also highest-role-oriented. To take a simple example, consider the following ditransitive: (19)

John gave me the books.

Theoretical framework

13

The 9-structure of give is: (20)

( agent ( goal ( theme ( give ))))

By the Priority Clause (15c) we should start with lower roles and with grammatical linkers. The theme is lowest and is eligible for linking by the sister-of-V position (16c) or the sister-of-V' position (16c) (agreement and VP-external position (16a) are not available, since there are higher nondemoted roles, the agent and the goal). Since sister-of-V' (16c), although a grammatical linker, is semantically restricted to theme, and the lowest 0-role here is a theme, this 6-role is linked by (16c): (21)

( agent ( goal ( theme ( give ))))

Moving on to the next lowest role - in accordance with the Priority Clause (ii) - the goal is eligible for linking by the sister-of-V linker only (goal is not the highest available role for grammatical linking and so agreement and VP-external position may not apply). Goal is linked by (16b): (22)

( agent ( goal ( theme ( give ))))

I

I

V_

V'_

All that remains is the agent, and it is linked by agreement and VPexternal position (16a): (23)

( agent

I

agr, _V"

( goal ( theme ( give ))))

I

V_

I

V'_

All G-roles are linked, and so the NPs at lexical structure pass the Obligatoriness Condition. Kiparsky's analysis captures the other possibilities for ditransitives, including those involving a particle, like send up, since he assumes that goals in ditransitives are optionally demoted: (24)

( agent ( goal(A) ( theme ( give ))))

14

Introduction

The possibilities for linking correspond to acceptability judgments: (25)

V"

V a. b. c. d. e. f. g. h.

V

N"

N'

Prt

sent a drink sent sent John *sent John *sent a drink *sent *sent *sent

P"

up up a drink up a drink a drink up John up John up to John up a drink

to John to John up a drink a drink John

In all of the versions in (25), the agent is linked by agreement and VPexternal position as already noted. The first possibility (25a) is based on a demoted goal linked by means of the semantic linker to. The theme is linked by sister-of-V position. In (25b) the goal is similarly linked by to, but this time the theme is linked by sister-of-V'. This is possible, since in KLT neither of the two possible positional linkers for the theme has priority over the other. In (25c) the goal is not demoted. Only the theme qualifies for the sister-of-V'; so if both goal and theme are to be positionally linked, it will be theme by sister-of-V' position and goal by the inner position. In the remaining VPs in (25d-h), something has gone wrong. In (25d) the sister-of-V position has linked two 6-roles (goal and theme), which violates the One-to-One Constraint in (15a) above. The same is true of the outer verbal position in (25f); here it links both the goal and the theme. In (25e), the outer verbal position has linked a non-theme, contrary to that linker's own distributional constraint. Finally, in (25g) and (25h), one NP, a drink and John, respectively, is not linked at all: in each, the offending NP is in a position that corresponds to no linker and neither has an appropriate prepositional linker. Note that (25h) could be saved by linking the final NP, John, with the preposition to. This, in fact, would yield the acceptable (25b).

Theoretical framework

15

Direct linking, by deriving grammatical functions, makes several claims. Most obviously, the more that can be derived rather than taken as primitive the better. However, a defender of grammatical functions as primitives might counter that the linkers themselves would not be necessary if grammatical functions are taken as primitives. The real questions are rather which theory is simpler overall and which theory is empirically better supported. On the first score, the elegance of the definitions of grammatical functions should carry some weight. More importantly, direct linking makes a range of empirical claims which stem from the elimination of grammatical functions as primitives. First of all, direct linking does not provide for "deep grammatical relations" like those in Relational Grammar (RG) (see especially Perlmutter (1982)). Notions of subject in RG include initial 1, working 1, acting 1, and final 1 (and correspondingly for other grammatical relations). In direct linking there are no intermediate strata and no grammatical relations at all until linking is complete. In fact, direct linking allows no use of grammatical relations in capturing morphosyntactic generalizations like the distribution of nominative and accusative cases. And, since only final grammatical relations are defined, the most that other processes like binding can mention is logical subject (highest argument in argument structure) and (final, surface) subject. For areas other than morphosyntax, then, notions like working 1 and acting 1 are not available. Finally, it should be noted that logical subject and logical object are not the same as initial 1 and initial 2 in RG, which are also unavailable in direct linking. In a theory like RG based on grammatical relations, there are normally rules based on grammatical relations. Passive is an example, since on the RG view it turns a subject into an oblique and an object into a subject. Once again this is a morphosyntactic process and direct linking cannot capture it using grammatical relations (as we have already seen, direct linking holds passive to be the demotion of the logical subject). In RG, the rules that change grammatical relations can furthermore be used to make possible entire levels that are more abstract than would otherwise be possible. A typical example is the RG approach to unaccusativity (also essentially adopted by GB). RG has an unaccusative rule for so-called unaccusative verbs like die. Roughly speaking, unaccusative verbs like die have non-agentive subjects. RG assumes that, unlike unergative verbs (which do have agentive subjects, e.g. work), unaccusative

16

Introduction

verbs have underlying initial objects rather than subjects. This object is transformed into a subject before the surface. In direct linking, however, such an approach is not possible. The single argument of all intransitives, whether unaccusative or unergative, is a logical subject. To the extent that unaccusativity is really a syntactic phenomenon, then, it must be captured in some other fashion, e.g. by making a distinction between external and internal arguments (i.e. the sole argument of an unaccusative would be an internal argument, but the sole argument of an unergative would be an external argument). In the next subsection I will discuss how the proposals of this book differ from those of KLT. Before doing so I want to emphasize both what is common to direct linking theories and what places the theory I will develop firmly in the class of direct linking theories. First, a direct linking theory relates argument structure to surface form "directly." This relationship is not wholly direct in that we have a distinction between surface form versus syntactic linkers which "assign" that form, i.e. the morphological vs. syntactic case distinction mentioned earlier. However, relating arguments to surface form by syntactic linkers is more "direct" than relating argument structure to surface form via grammatical relations. The reason is that our syntactic linkers will have strong limits on how abstract they can be. If an NP is dative in one sentence and is accusative in another sentence, traditional theories would allow us to claim that both NPs are objects but that the objects get different case marking. The category object relating the arguments to cases is abstract in the sense that the cases do not constrain our choice of grammatical function analysis for the NPs Linkers are syntactic "cases" in the sense that they do place a limit on this kind of abstractness. If one NP is dative and another accusative, they cannot be said to be linked by the same linker.8 More generally: (26)

If two NPs are linked by the same linker, they must show the same corresponding surface form. Equivalently, if two NPs show different surface form, they cannot be linked by the same linker (s).

By "surface form" I mean the surface correlate of a morphosyntactic category. So, to return to the hypothetical example, if one NP is dative, another is accusative, and we have already analyzed the language as having a dative and an accusative linker, it would not be possible to say that both NPs are linked by the dative linker, but one just appears

Theoretical framework

17

accusative. This is quite unlike the situation with grammatical functions, where it is perfectly fine to say that two NPs are direct objects but one appears dative and the other accusative. The same is true for other types of linkers as well. Take the situation where, based on other facts, our analysis of a language has some combination of word-order positions as the linkers. Where an NP appears in the phrase structure (the result of another kind of analysis) will put limits on which linker we could claim to be linking the NP. An NP that appears outside V' cannot be said to be linked by sister-of-V or VP-external position. So in direct linking the surface form - case, configuration, and/or agreement - does limit the possibilities for morphosyntacic analysis. A fundamental assumption of direct linking is that different morphological categories imply different linkers.9 Morphology strongly constrains the theory of linking. The general question about direct linking of any kind is whether it captures morphosyntactic generalizations better than an approach using grammatical functions (or their equivalents, such as structural positions in GB). 10 There are three ways that direct linking or mediated linking could be better at capturing such generalizations. First, we should ask whether one or the other leads to a confirmed generalization that the other does not. Second, the generalizations that one approach can capture may be more accurate than those made possible by the other one. Third, it will be important to find out whether the generalizations on one approach are more general and the set of generalizations is more elegant. This last criterion can be seen as the criterion of explanatory value. 1.2.3 Notation and basics The theory to be presented below is one of direct linking, and, although KLT is its point of departure, it will be useful to have an overview of how the proposals of the following chapters differ from KLT. The differences are both substantial and notational. First, I will not assume thematic roles or thematic role labels for arguments. I will propose a distinction between argument types and argument categories and will use only the latter in linking. Argument types are classes of arguments based on the entailments in which they participate (see Dowty (1989)). Argument types are basically thematic roles. Argument categories, however, are defined as syntactic classes of arguments that bear only a looser connection to corresponding argument types. For example, in some languages which we will examine, many

18

Introduction

"experiencer" arguments are dative, but many are nominative. If the 0role experiencer is used in linking we get a rule like "Some experiencers are dative" or "Experiencers are dative unless otherwise indicated by exception feature." In the next chapter I will suggest that we instead adopt a syntactic category EXP whose members must be interpreted as experiencers in the semantic sense. The implication is from syntactic experiencerhood to semantic experiencerhood. But the implication is only one way. An argument can be interpreted as an experiencer without belonging to this EXP category. Analogously, in the case of the category auxiliary (AUX), the members of the category AUX must express tense, aspect, or modality, but not all words (or even verbs) that express tense, aspect, or modality belong to this AUX category (cf. must vs. have to - must inverts, but have to does not, even though they both express necessity). If this way around the problem of thematic roles is adopted, we can avoid having to complicate the semantic definition of an argument type just to make the definition cover syntactic facts, like subject selection, binding, etc. Subject selection is another area in which 0-roles have been invoked, and if one is to do without them, a word about subject selection is in order. One of the virtues of the hierarchy in (5) above is that arguments are ordered with much less stipulation than if one simply ordered arguments for each individual verb. The hierarchy embodies a claim that the syntactic "eligibility" for subjecthood (modulo demotion, as in, e.g. passive) is completely predictable on the basis of lexical semantics. This is a strong claim, and in many cases it works. For example, verbs which express the inverse relation of hit, where theme would be subject and agent would be object (a lexicalized version of got hit by) are rare or non-existent. But such predictions are subject to some obvious counterexamples. One such is strike in English as in John strikes me as brilliant, where the theme rather than the experiencer (a subvariety of goal in KLT) is the subject. This is the opposite of what we expect from the hierarchy which has goal higher than theme. I will follow Dowty (1991) in taking generalizations about the semantic basis for subject selection to be defaults, possibly to do with language acquisition - rather than to do with synchronic grammar. That is, a verb with a "theme" subject and "agent" object would be harder to acquire or more costly (on a redundancy rule approach, e.g. Jackendoff (1975)) than verbs the other way around. Thus in the theory to be proposed below, argument structure consists of ordered arguments, but the ordering will be by default -

Theoretical framework

19

or in the worst case, by stipulation, as in standard ordered argument theories like most versions of Categorial Grammar. No thematic role labels are involved. I will also propose that one piece of functional information, obliqueness versus non-obliqueness, be part of a linker's specifications. In Kiparsky (1987) if, say, a dative experiencer, although the logical subject, is not the behavioral subject (i.e. it is an oblique), this is handled by including lexically stipulated demotion on the experiencer 0-role in the verb's lexical entry: (27)

Mir gefallt me(DAT) pleases 'I like the essay.'

der the(NOM)

(28)

( goalA ( theme ( gefallen ))

Aufsatz. essay

The demotion is lexically specified here as it was for certain goals in some of the ditransitives above in (25). There are several disadvantages to this. One is that this makes it very easy to write a grammar in which some dative experiencers are syntactically oblique but some are not. Nonsyntactically oblique (behavioral subject) experiencers are found in Icelandic and some other languages (see the next chapter for references and discussion), but in these languages all dative experiencers behave as subjects, i.e. are not syntactically oblique. Although dative subjects are well-known, it seems that no language has been found with both dative behavioral subjects and behaviorally oblique dative logical subjects. Therefore in the following I will assume that linkers themselves are specified for obliqueness and obliqueness is inherited (across-the-board) by the NPs they link. So in German the experiencer dative is both morphologically and syntactically oblique, whereas in Icelandic the experiencer dative is morphologically oblique but syntactically non-oblique. The definitions in (10) will still work, but the source of (syntactic) obliqueness will be the linkers themselves rather than features stipulated at verbal lexical entries. I will further assume that linkers come in three kinds or types: case, agreement, and word order. But I will not assume that agreement has any special synchronic "subject orientation." Rather it is proposed that orientation towards highest (or lowest) argument will be taken care of by constraints otherwise needed for cases like nominative. Nominative case, and, in English, VP-external position normally link the highest argument, and however this is handled can be extended to agreement.

20

Introduction

Likewise I will not assume a priority of grammatical linking over semantic linking, and it will not even be necessary to assume this particular distinction. Priority of one linker over another will be handled by means of constraints stated in terms of a notion of restrictiveness. The intuitive idea behind restrictiveness is relative information content. Cases (and other devices) of greater information content will have priority. This theory will provide a way of predicting on the basis of the complements that a verb selects what cases, agreement markers, and positions those complements will show, and, from there, of predicting their characteristic grammatical functionhood behavior. To do this, several universal principles will interact with language-specific inventories of morphosyntactic devices to yield the correct syntactic "case" (linker) frames. Neither part of the Priority Clause (15c) will be assumed: any priority of one linker over another or one argument over another should follow from general constraints and minimal assumptions about individual linkers. Since the basic notion behind restrictiveness is intuitively simple and general, many of the proposals made below are compatible with more than one theoretical framework. I have found it quite convenient, though, to couch the main constraints which constitute the theory in a unification-based framework. The basic notion of restrictiveness will rely on the notion of subsumption. If linker A is more specific than linker B in terms of linking features, A will be the more restrictive linker and will have priority over B. The linking features over which specificity will be determined will be features and values which will refer to basic distribution, upon which a theory of syntactic - not morphological - case rests. So for example, assume a language has a special dative case for arguments belonging to the syntactic category EXP (syntactic analogue of experiencer) and a nominative case that in principle marks any argument. When it comes to linking a given EXP argument, the dative will have priority because its environmental specification, "EXP arguments," is more specific than the environmental specification for nominative ("arguments"). This is the basic idea, and the details are a large part of what this book is about. 1.3

Overview

This book is organized as follows. Chapter 2 presents the basics of the theory after examining the phenomenon of case alternations. The focus will be on Icelandic and its Germanic relatives. I will present there the

Overview 21 three main constraints which together constitute the theory of linking. The theory will first be shown to cover the data from the Germanic languages, and in Chapter 3 this theory will be generalized by adding two parameters. The resulting typology of two-linker systems will predict the existence of unexpected language types which do exist. The behavior of superficially quite different case systems will be shown to have a surprising unity, which the theory captures. Chapter 4 will examine the interaction of multiple morphosyntactic processes, such as passive and case marking. The implications of the theory for the interaction of case, position, and obliqueness will be drawn out. Chapter 5 will continue the discussion of case and word-order position with a more detailed analysis of Modern Icelandic. In Chapter 6, it will be shown that, in conjunction with certain assumptions about change, the theory accounts for a wide range of non-parametric changes in linking. Finally, Chapter 7 will provide an account of some new facts about case preservation and passive not accounted for by present theories of case but which follow from a minimal extension of the theory.

2

Argument case and case alternations

In this chapter I introduce the outlines of a theory of case. As a way into this problem and for concreteness I will first examine a problem that highlights the need for a constrained theory of case. This is the problem of case alternations, where two cases alternate with little or no change in meaning. After presenting this problem, I lay out the fundamentals of a theory of linking and some of its most basic consequences. Later I return to the case alternations and show that the theory just developed turns out to handle the alternations with no extra machinery beyond what is motivated for the most basic phenomena. Finally, I summarize the implications of this theory of argument case for the notion of argument structure and the distinction between arguments and adjuncts.

2.1

Case alternations and the problem of case

2.1.1 Basic facts Recent work in syntax has touched upon a case alternation in Icelandic known to traditional Icelandic grammar as Dative Sickness (pdgufallssyki) or Dative Substitution (henceforth DS). At first approximation it can be described as the tendency for accusative subjects to become dative, as with the verb langa 'long' in (1): (1)

a. Mig langar ad me(ACc) longs to 'I long to go.' b. Mer langar ad me(DAT) longs to

fara. go fara. go

The accusative variant (la) is original in Icelandic, the dative (lb) an innovation. Icelandic also shows a further alternation of impersonal (1) with personal constructions like (2):1 22

Case alternations and the problem of case 23 (2)

Eg

langa ad fara. long to go 'I long to go.'

I(NOM)

For parallelism, I will call this second alternation Nominative Sickness or Substitution (NS). Diachronically, there are two tendencies, one from accusative to dative and another from either accusative or dative to nominative. Both "substitutions" have a synchronic aspect as variation and a diachronic aspect as trends. This section will establish that DS and NS have to do with case only, not grammatical function. However, as we will also see, DS and NS differ in two respects: (i) DS but not NS is limited to experiencer NPs and (ii) DS (not NS) is limited to the "replacement" of accusative case. 2.1.2 Case alternations and subjecthood It is necessary first of all to inquire whether the case alternations depend on grammatical functions. That they do not can be easily established by comparing Icelandic and German, two languages with different grammatical subjecthood properties but which share DS and NS. The alternations in (1-2) in Icelandic involve the "thematically most prominent" argument or "logical subject." It is one of the most interesting results of recent work on Icelandic syntax (Andrews (1976), Thrainsson (1979), Andrews (1982), and Zaenen, Maling, and Thrainsson (1985)) that all of the initial NPs in sentences like (1-2) also behave as subjects grammatically. On a wide variety of tests which distinguish subjects from topicalized NPs, the lexically case-marked NPs pattern with the subjects, e.g. lexically case-marked NPs can be the target of control, like subjects (and unlike topicalized NPs, cf. Andrews (1982: 55)):2 (3)

a. Eg

vonast til ad vanta ekki efni i ritgerdina. hope for to lack not material(NOM) for thesis-the 'I hope not to lack material for the thesis.' b. Barnid vonast til ad batna veikin. child-the(NOM) hopes for to recover-from disease-the(NOM) T h e child hopes to recover from the disease.' I(NOM)

And, like subjects, lexically case-marked NPs undergo Raising: (4)

a. Hana virdist vanta peninga. her(ACc) seems to-lack money(Acc) 'She seems to lack money.'

24

Argument case and case alternations b. Barninu virdist hafa batnad veikin. child-the(DAT) seems to-have recovered-from disease-the(NOM) T h e child seems to have recovered from the disease.'

Note that the lexical case is preserved under Raising, showing that the subjecthood behavior could not be dependent on some kind of case absorption. In Modern Icelandic generally, behavioral subjects are normally nominative, but may be in any of the other cases. Which case is possible depends on the verb and, for some cases, is partially predictable from the verb's meaning. Some verbs take experiencer arguments (i.e. whose referents are entailed to be aware of something and to be sentient beings), which are often dative: (5)

a. Mer kolnar. me(DAT) is-getting-cold 'I am getting cold.' b. Barninu batnadi veikin. child-the(DAT) recovered-from disease-the(NOM) T h e child recovered from the disease.' c. Mer bydur vid setningafraedi. me(DAT) is-nauseated at syntax 'I abhor syntax.'

The variant in (lb) above with the verb langa 'long' also shows a dative experiencer. Another group of experiencer subject verbs take nominative only, a typical example being sjd 'see':

(6)

Eg

se

I(NOM) see

frig. you(Acc)

'I see you.' Note, too, that the NS variant of langa 'long' in (2) likewise shows a nominative experiencer (for those speakers that get (2)). The experiencer subjects of a small class of verbs is accusative. Example (la) with langa 'long' is one, as are the following: (7)

a. Mig kelur. me(ACc) is-freezing 'I am freezing.'

Case alternations and the problem of case 25 b. Mig me(Acc) 'I lack a c. Mig me(Acc) 'I abhor

vantar hnif lacks knife(ACc) knife.' velgir vid setningafraedi. is-nauseated at syntax syntax.'

Andrews (1982: 463) examines this accusative-subject class and concludes that the accusative here is probably idiosyncratic, since there seems to be no semantic criterion that would distinguish these verbs from nominativesubject verbs and dative-subject verbs. Andrews does notice certain tendencies, which also feature prominently in the handbooks. Accusatives tend to appear with physiological states, while the datives appear with psychological states. This and other generalizations are just tendencies, and pairs like (5c) and (7c) seem to be truth-conditionally equivalent. There are also dative and accusative subjects which are not experiencers, making the correlation between experiencerhood and morphological case even less direct.3 German also has lexical cases for experiencer arguments. Unlike in Icelandic, experiencer NPs in German are behavioral obliques, not subjects or objects (Cole et al. (1980), Reis (1982)). Depending on the verb, an experiencer may be nominative (8), dative (9) or accusative (10): (8)

Der Mann sieht den Tisch. the(NOM) man sees the(Acc) table T h e man sees the table.'

(9)

Mir ist iibel. me(DAT) is sick 'I am sick.'

(10)

Mich hungert. me(ACc) hungers 'I am hungry.'

(cf. Icelandic (6))

The latter two classes, dative and accusative experiencers, do not behave as subjects in contrast to their counterparts in Icelandic, e.g. they cannot be the target of control: (11)

*Ich hoffe iibel zu sein. I(NOM)hope-lsg sick to be 'I hope to be sick.'

26 (12)

Argument case and case alternations *Ubel zu sein ist unangenehm. sick to be is unpleasant 'To be sick is unpleasant.'

On the other hand, nominative experiencers do pattern with normal subjects: (13)

Er hofft mich dort zu sehen. he(NOM) hopes me(Acc) there to see 'He hopes to see me there.'

The experiencer subject of sehen 'see' can be controlled, but the experiencers of dative verbs cannot. This situation contrasts with that in Icelandic, where lexically case-marked NPs do behave as subjects. Very interestingly, German has precisely the same types of case alternations as Icelandic, despite the radically different functional status of the NPs in question. German shows alternation of accusative with dative on experiencer NPs, although, as we have seen, these NPs are obliques. Almost all of these verbs can take a dative (and sometimes nominative) as well. The accusative in (14) varies with the dative in (15): (14)

a. Mich ekelt. me(ACc) is-disgusted 'I am disgusted.' b. Mich schaudert vor Baren. me(ACc) shudders at bears 'I shudder at bears.'

(15)

a. Mir me(DAT) b. Mir me(DAT)

ekelt. is-disgusted schaudert vor Baren. shudders at bears

Any condition on DS that the NP involved must be a subject, while true for Icelandic, would not generalize to German DS. For a truly general explanation of the DS phenomenon, we will have to avoid any appeal to the special status of oblique case in Modern Icelandic. Many of the accusative and/or dative experiencer verbs in German can alternatively take nominative, i.e. they can show NS: (16)

Ich

hungere.

I(NOM) am-hungry

'I am hungry.'

Case alternations and the problem of case 27 (17)

Ich

schaudere vor Baren. shudder at bears 'I shudder at bears.'

I(NOM)

The nominative in (16) varies with the accusative in (10), and the nominative in (17) varies with the the accusative in (14b) and the dative in (15b). Therefore, NS is not dependent on DS, since an accusative-taking verb that does not undergo DS for some speakers, e.g. hungern 'hunger,' can undergo NS for those speakers. For many further examples, see Curme (1922: §219) and Dal (1966: §124). 2.1.3 Systematic limitations on the case alternations I now turn to some systematic limits on DS. DS is variously described as a replacement by dative, an analogy based on a default dative, or an analogy based on experiencer dative (Barnes (1986)). I will first show that only this last view is correct, because only those accusative subjects that might be called experiencers can undergo DS. We can first ask whether a non-experiencer argument can take dative by DS. In (18a) the sole NP is accusative. It nonetheless cannot become dative (19) but can become nominative (18b): (18)

a. Kjarkinn courage-the(Acc) T h e courage was b. Kjarkurinn courage-the(NOM)

brast. was-lacking lacking.' brast. was-lacking

(19)

*Krafti brestur. strength(DAT) lacks 'Strength is lacking.'

The accusative in (18a) is best taken as lexical since sole NPs otherwise get nominative, but it cannot show DS, so DS cannot just be the replacement of accusative with dative on subjects. Other non-experiencer subjects likewise do not show DS. Thus the result of replacing the accusative on the subjects in (20) with dative is ill-formed (accusative examples are from Svavarsdottir (1982: 22)): (20)

Bdtinn/*Bdtinum boat-the(Acc)/boat-the(DAT) 'The boat drifted to sea.'

rak til hafs. drifted to sea

28 (21)

Argument case and case alternations Manninn/ *Manninum bar vid baeinn. man-the(ACc)/man-the(DAT) could be seen against house-the 'The man could be seen against the house.'

What is common to these verbs as opposed to the DS verbs is that their subjects are not experiencers, i.e. none of these sentences entails that the individual denoted by the subject NP is aware of anything or sentient. DS is not the tendency for all accusative subjects to become dative but rather the tendency for accusative experiencer subjects to become dative. Several previous explanations (e.g. van Valin (1991)) have relied on analyzing the dative as the default lexical case in Icelandic in order to explain DS. This leads to an expectation that genitive subjects could become dative. This does not happen: (22)

Verkjanna/*Verkjunum gaetir pains-the(GEN)/pains-the(DAT) be-noticeable T h e pains are not noticeable.'

ekki. not

This is all the more surprising given the rarity of genitive subjects. Despite their low numbers and lack of semantic motivation they cannot participate in DS. Note that the genitive subjects are not experiencers either, forming another class of non-experiencers which do not undergo DS. Finally, DS is often closely associated with NS on another argument. For the accusative subject verbs in (1) and (7a) above with only one argument, DS is simply the replacement of the sole accusative by a sole dative. However, for verbs like vanta lack' (7b) with more than one argument, there is another accusative NP which can also alternate, but with the nominative. Yip, Maling, and Jackendoff (1987: 23If.) mention in passing that this NP is treated differently depending on the verb. With verbs of lacking, the second NP becomes nominative except for vanta 'lack' (23c ( = 7b)), where it remains accusative: (23)

a. Mig brestur kjark me(Acc) lacks courage(Acc) 'I lack courage.' b. Mennina Jsrytur mat the-men(Acc) lacks food(ACc) 'The men lack food.'

Case alternations and the problem of case 29 c. Mig vantar hnif me(ACc) lacks knife(ACc) 'I lack a knife.' (24)

a. Mer brestur me(DAT) lacks b. Honum J)raut him(DAT) lacked c. Mer vantar me(DAT) lacks

kjarkur courage(NOM) J)rottur strength(NOM) hnif knife(ACc)

Other case combinations occur for other speakers and so are possible in principle; e.g. the verb bresta 'lack' appears not only in the ACC-ACC frame (23a) and the DAT-NOM frame (24a) but also in the DAT-ACC frame (25) and occasionally in the ACC-NOM frame (Halldorsson (1982: 168)): (25)

Mer brestur kjark me(DAT) lacks courage(ACc) 'I lack courage.'

(26)

l>a brast £>olinmaedin hann sem snoggvast og then lacked patience(NOM) him(ACc) for a moment and hann sagdi he(NOM) said 'Then he ran out of patience for a moment and he said . . . '

Likewise vanta 'lack' is not confined to ACC-ACC and DAT-ACC, since ACCNOM examples occur (Halldorsson (1982: 178)). Previous theoretical discussion of DS has largely ignored the case alternation on the nonexperiencer NP and why it becomes nominative. Note, too, that no dialect seems to allow NS on both arguments: (27)

*k>eir vanta kjarkur. They(NOM) lack courage(NOM) 'They lack courage.'

Thus, with verbs like vanta, accusative, dative, and nominative vary on the first argument (DS, NS), and accusative and nominative (NS) vary on the second argument independently, but when the first argument suffers NS the second argument cannot show nominative. Finally, it is worth pointing out that, unlike DS, NS is extremely common in the world's languages. The Relational Grammar (RG) litera-

30 Argument case and case alternations ture presents examples of languages with what is termed optional inversion. This means that the logical subject (RG initial 1) of psych verbs will have the option of being a final 3 or a final 1. In RG, analyses are generally set up so that final relations map directly onto surface morphology. So final 3 translates into other frameworks as dative and final 1 translates as nominative. The case alternation in optional inversion is between dative and nominative, a type of NS. Examples are Kannada (Sridhar (1976), Dryer (1982)) and Udi (Harris (1984)). That some verbs in Korean alternate between dative and nominative for experiencers is well known, for example with the verb chohta 'to like': (28)

a. na-ege chaek-i chohta. me-(DAT) book-NOM like 'I like the book.' b. ne-ga chaek-i chohta. I-NOM book-NOM like 'I like the book.'

In our terms, Korean shows NS with these verbs. Tentatively speaking, the NS phenomenon is quite common cross-linguistically, in contrast to DS which (so far) seems to be confined to Germanic. The fact that some verbs allow DS but not NS while others allow NS but not DS clearly means that one cannot be the precondition for the other. When considered diachronically, the change of which each alternation is a slice must be able to begin independently. I will take up the diachronic aspect of the two alternations in Chapter 6. In this section it has been shown that, although DS and NS are case alternations, DS is limited to experiencers and to accusative NPs, unlike NS. DS and NS are usually thought of as either replacement rules or variation conditioned only by sociolinguistic factors. The problem with replacement rules is that they are too powerful; they do not come with any limits on what can be replaced with what. And many replacements should be made difficult or impossible to capture, such as the replacement of dative with genitive. On the other hand, it should be emphasized that labelling these alternations "sociolinguistic variation" does not answer the question why DS systematically shows these two restrictions and NS does not. At the very least, it must be acknowledged that the variations are grammatically conditioned. 4 The rest of this chapter will be an attempt to show what those grammatical conditions are.

Case, argument structure, and interpretation 2.2

31

Case, argument structure, and interpretation

This section will briefly develop a theory of morphological and syntactic case, which will be able to solve these problems. The general framework is unification-based, and, although I will not invoke one theoretical framework, these proposals are most compatible with Generalized Phrase Structure Grammar (GPSG) (Gazdar et al. (1985)) and its relatives. I will start from a standard ordered argument theory and present a general way, given a verb and its argument structure, of building up case frames. This theory is based on three main constraints, introduced in §§2.2.1-3. 2.2.1 Linking features One way to separate out the universal aspect of linking is to generate case frames and subject them to universal well-formedness conditions. I will propose three such constraints. Before turning to the constraints, it is essential to assume two notions, subsumption and unification. As mentioned earlier, linkers and the NPs to be linked will be defined as feature structures5 and these will be presented in Attribute-Value Matrix (AVM) notation familiar from LFG. An AVM is a directed acyclic graph, and an important property for us will be that a feature can take either an atomic (e.g. + ) or a structured (e.g. [NUM [SING]]) feature structure as a value. An important instance of this is that, as in LFG (subsequently borrowed in other frameworks), a verb "takes" arguments in the sense that it has a feature whose value is a feature structure made up of the feature structures of its arguments: (29)

CAT [v] CAT [ N P ] LINK [ARG [CAT [EXP [PLACE [1]]]]] SUBCAT

. REAL [

]

CAT [ N P ] LINK [ARG [CAT [DIR [PLACE [2]]]]] - REAL [

]

This feature structure includes many of the features we will be using. The feature SUBCAT is from GPSG and its descendants and is for subcategorization. LINK is for linking specifications and will take as a value a feature structure indicating the properties relevant for linking which the category in question (usually an NP) has. The feature ARG is

32 Argument case and case alternations for argument, CAT is for argument category, EXP is for experiencer, and DIR is for direct (see the next subsection). The PLACE feature is just a way of indexing arguments to get an ordering of arguments. The logical subject will by definition have [PLACE [1]]. REAL is a feature for realization and takes as values various morphosyntactic categories. I will take linking to be the filling in of values for REAL (SO far unspecified in (29)) on the basis of the value of LINK. In other words we are deriving case frames of verbs from properties like argumenthood and experiencerhood in the NP the verb selects. A path is a sequence of features, and, by way of extension, we can say that a path takes as a value the value that the last feature in the path takes. Thus in (29) the path SUBCAT LINK ARG CAT DIR PLACE has the value [2]. The main constraint below states that the more specific the linker, the higher priority it gets. In the unification-based framework this is handled through subsumption. Feature structure A subsumes feature structure B if B is no more informative than A: (30)

a. A: [ ] b. B: [ARG] c. C:

[ARG [EXP]]

d. D:

[ARG [EXP

[2]]]

A subsumes B which subsumes C which subsumes D. Note too, that A subsumes C and D, B subsumes D and that each feature structure subsumes itself.6 We will say that A properly subsumes B if A subsumes B but B does not subsume A. In this case A is less informative than B. All of the non-reflexive subsumptions in (30) are proper subsumptions. Unification will also be important in what follows. Basically, to take the unification of A and B, pool the information of A and B. If the information is incompatible, unification is undefined and is said to "fail." 7 To take a couple of examples like those we will see later, consider (31). NPs have certain properties like argumenthood, experiencerhood, etc., captured in the LINK value. Linkers have these same properties and a large part of constraining linking will be to "match" the linker to an NP with appropriate properties, REAL is a feature that indicates the morphosyntactic expression that the linker "assigns." It is like the "operation" that - by way of the linker - comes to be associated with a given NP. So for instance, the unification of (31a) and (31b) is (31c):

Case, argument structure, and interpretation (31)

a.

33

[CAT[NP] LINK [ARG [CAT [DIR [PLACE [2]]]]] .REAL [

b.

]

[ LINK [ARG] REAL [NOM

C.

CAT [ N P ]

I

LINK [ARG [CAT [DIR [PLACE [2]]]]] REAL [NOM]

J

In other words, an NP is linked by unifying with the linker. Note that if unification fails no linking can occur, as with an experiencer dative and a direct argument : (32)

a.

• CAT [ N P ] LINK [ARG [CAT [DIR [PLACE [2]]]]] . REAL [

b.

]

LINK [ARG [EXP]] REAL [DAT]

No unification can take place because the dative is specified for experiencers whereas the NP target is a direct argument. The experiencer dative must link experiencer NPs. 2.2.2 How semantic is "semantic" case? If one takes a closer look at this type of case, a problem emerges immediately. Some of these inherent cases seem to be partially predictable from a verb's meaning, while others are not. First, a case can specially mark a particular argument of a particular verb. When this happens, the verb is generally said to assign case idiosyncratically to that argument. One Icelandic verb that assigns dative to its second argument is tyna 'lose' (33), and one that assigns genitive is sakna 'miss' (34): (33)

(34)

Barnid tyndi bokinni. child-the(NOM) lost book-the(DAT) 'The child lost the book.' Eg

sakna J)in. miss you(GEN) 'I miss you.'

I(NOM)

Such idiosyncratic cases are located in the lexical entries of the verbs involved.

34 Argument case and case alternations On the other hand, some lexical cases do seem to have some thematic basis (see, e.g. Zaenen, Maling, and Throinsson (1985)). For example, as we have seen above, it is a loose generalization that experiencers - arguments whose referents are entailed to be sentient and aware of something - take dative in Icelandic and German. But this generalization is only partial. Some experiencers are nominative and a few accusative. Nevertheless, the experiencer dative is fairly robust and productive (see also Chapter 6). At this point, we are faced with at least two choices in handling such partial generalizations. We could list exceptions to the experiencer-dative rule. We already noted that this leads to a very large number of exceptions. This approach faces two further problems, the first being that thematic roles are notoriously difficult to define in the first place, so that an exception-ridden rule referring to them is on very shaky ground. Another difficulty is that if the nominative-experiencer class is an exception to the regular dative pattern, we might expect this class to show marked behavior, including being less frequent. However, evidence for this is lacking. Icelandic and German have quite a number of nominative experiencer verbs as well, and they show no marked behavior. These partial "thematic" statements can, however, be replaced by generalizations about the interpretation of certain NPs. If we ask in what sense a dative experiencer subject is more predictable than a genitive subject, the answer must be not that experiencers are dative (too many exceptions) but rather that dative logical subjects are almost always experiencers. The predictability runs from the syntax (argument structure) to the semantics, not from the semantics to the syntax as usually assumed for "thematic case," and so we might derive an interpretation from the argument structure rather than vice versa. Thematic role theory also assumes that lexical cases "select" for a given thematic role (or piece of a lexical decomposition). However, as we have seen, the generalizations such a move is meant to capture are partial at best, with the predictability of information running rather from the syntax to the semantics. One way to capture this insight is to posit syntactic classifications of arguments such as experiencer or goal, as I will here: (35)

a. Argument categories: DIR, EXP, GOAL b. Default: [CAT [

]] —> [CAT [DIR]]

In this, I am following the strategy for thematic relations proposed in Dowty (1991) of positing just those aspects of thematic roles as necessary

Case, argument structure, and interpretation

35

for a given area. I will take "thematic roles" to be argument types picked out by sets of entailments (Chierchia (1984, 1989), Dowty (1985, 1989), Ladusaw and Dowty (1987)). But I will not assume that these argument types are directly accessible to case marking. For the area of case, arguments are organized into (syntactic) argument categories which are distinct from (semantic) argument types. A syntactic experiencer is necessarily interpreted as a semantic experiencer, i.e. as a being who is aware of something and is sentient. The first argument of langa 'long' will be a syntactic experiencer and so bears Exp(eriencer): (36)

langa ( X|EXP, y VP ) 'long'

But the reverse is not necessarily true: a participant entailed to be aware of something and sentient does not necessarily correspond to a syntactic experiencer: (37)

If an argument belongs to the syntactic category experiencer (X|EXP), then its referent is entailed to be aware of something and be a sentient being, i.e. a semantic experiencer (syntactic experiencer —» semantic experiencer).

There is thus a one-way implication: if an argument is a syntactic experiencer (it bears EXP) it must be a semantic experiencer (the entailment holds). The converse does not hold: it is not the case that every argument corresponding to the subjects of the entailments in (37) bears the value EXP, i.e. it is not the case that all semantic experiencers correspond to syntactic experiencers. The verb sjd 'see' is an example of this: (38)

sja ( x, y ) 'see'

The first argument, although a semantic experiencer, is direct, the default argument category which places no additional constraints on the interpretation of an argument. In the abbreviated forms of lexical entries like (38) I will supress DIR. Similarly, a syntactic goal (GOAL) is necessarily interpreted as an entity to which something moves. As with the EXP category, the converse is not necessarily true: a participant entailed to be an entity to which something moves is not necessarily a syntactic goal. In a fashion parallel to (37) for EXP, the following will constrain the interpretation of syntactic goals:

36 Argument case and case alternations (39)

If an argument of a verb belongs to the syntactic category goal (X|GOAL), then a positive sentence containing that verb entails that something moves to the referent of that argument, i.e. a semantic goal (syntactic goal —> semantic goal).

Once again we have a one-way implication: if an argument is a syntactic goal (it bears GOAL) it must be a semantic goal (the entailment holds). However, not all semantic goals are syntactic goals. The well-known problem of commercial transaction verbs (see, e.g. Jackendoff (1987: 381) and Dowty (1991: 556)) is avoided this way. These verbs are a problem for theories of thematic roles which claim that each of a verb's arguments receives a unique and distinctive thematic role from that verb. Take the following Icelandic sentences: (40)

Jon seldi Olafi bilinn. John(NOM) sold Olaf(DAT) car-the(Acc) 'John sold Olaf the car.'

(41)

Olafur kaupti bilinn fra Olaf(NOM) bought car-the(ACc) from 'Olaf bought the car from John.'

Joni. John(DAT)

Both sentences describe the same situation; they have the same truth conditions. Both human participants are entailed to act agentively. There are two entities which change hands, the car and the means of payment. Further, there are two goals: both verbs entail that something moves to each human participant. The car goes to Olaf, and the money or other means of payment goes to John. However, the participant that is lexicalized as subject in (40) is expressed by a PP in (41). The problem is that both entail that something goes to each of the two human participants. Yet only one of the NPs gets dative. In our terms, although the verb selja 'sell' has two semantic goals, it has only one syntactic goal: (42)

selja ( x, y|GOAL, z ) 'sell'

The fact that the implication is one-way - a syntactic goal must be a semantic goal - but not vice versa is supported by such otherwise problematic examples. One might ask why we do not get a verb with two datives. In other words, why do we not get two syntactic goals? This might be stipulated, as thematic uniqueness usually is. But we will see later that the uniqueness of syntactic argument categories follows from

Case, argument structure, and interpretation

37

the general constraint on linking called the "Linking Dissimilation Constraint." In several important respects, argument category features such as EXP and GOAL are like the feature AUX + (Steele et al. (1981), Gazdar et al. (1985) and references there). They define subcategories of arguments just as AUX + defines a subcategory of verb. And both EXP and AUX + bear the same relation to interpretation, AUX + by definition expresses tense, aspect, and/or modality, but it is not the case that any time one of these notions is expressed, we get auxiliary behavior (e.g. have to in English expresses obligation but is not an auxiliary, i.e. it does not invert or take a bare verb complement). So too each argument category like EXP is constrained to be interpreted in a specific way (e.g. (37), (39)), but it is not the case that every time an argument is interpreted in such a way, we necessarily get the corresponding argument category. In each case the syntax implies an interpretation but not vice versa. Traditional thematic roles, on the other hand, are "bidirectional." We will see that the unidirectional approach is ideal for case marking. The claim is not that syntactic cases like the dative in Icelandic reflect thematic roles but simply that they receive a predictable, redundancy-decreasing interpretation. One immediate advantage of the present approach becomes apparent if we ask what syntactic rules in general can mention. As we have seen, the thematic roles approach requires exception features to allow for exceptions to case-marking rules like dative-to-experiencers. Thus such an approach gives us access to the thematic roles themselves and the thematic roles minus instances with one or more exception features. Now note that an argument category like EXP in Icelandic has the same extension as the conjunction of the thematic role experiencer and the exception feature which allows an argument like the logical subject of sjd 'see' to be nominative. The thematic roles approach allows mention of both the equivalent of our argument categories and the pure thematic relations in addition to all sorts of categories defined by the conjunction of a thematic role and various combinations of exception features. Thus the present approach is just positing exactly what is needed for case marking while avoiding the claim that the entities involved in case marking should be universal and wholly semantically defined. This fits in with the general picture that many generalizations that were thought to make mention of thematic role turn out not to be best stated in terms of thematic roles (Dowty (1991)).8

38 Argument case and case alternations 2.2.3 The range of cases In this section I motivate some basic distributional features for the cases. The discussion proceeds from the most specific to the most general types of case. Later I will make this specific-to-general order more precise, terming it restrictiveness. I will assume that syntactic cases are specified for distribution, and this specification will be captured by means of the value of the feature LINK in the entries of the various cases (and other linkers). First, as we have seen, a case can mark a single argument of a single verb. When this happens, the verb is generally said to assign case to that argument as in (33) and (34) above. Such idiosyncratic cases are located in the lexical entries of the verbs involved. The case assigned will have distributional specifications as follows: (43)

a.

[DAT [LINK [ARG [CAT [DIR [PLACE

[2]]]]]]] (in lexical entry for

tyna) b.

[GEN [LINK [ARG [CAT [DIR [PLACE

[2]]]]]]] (in lexical entry for

sakna) The first entry is for the dative with the verb tyna 'lose.' Here the specifications are that the dative is assigned to an NP which is furthermore an argument - modelled by the the feature ARG - which is of category DiR(ect) and is the second argument. For convenience, the LINK values of cases will be abbreviated and placed after a colon as in (44): (44)

a DAT: [ARG|2] (in the lexical entry for tyna 'lose') b. GEN: [ARG 12] (in the lexical entry for sakna 'miss')

The fact that the "dative" argument is the second argument of just this verb is handled by the location of this dative within the lexical entry for the verb tyna. It has scope only over that entry. The genitive with sakna 'miss' works similarly. One of the prime motivations for introducing syntactic argument categories in the last section was in order to formulate so-called thematic cases like the "experiencer" dative in Icelandic. This is the kind of case commonly thought to require mention of thematic relations by syntax. In the present scheme, such cases refer to argument categories, e.g. (syntactic) experiencer dative:

Case, argument structure, and interpretation (45)

39

DAT: [ARG|EXP]

The question is what the range of such argument categories is. Since argument category case is formulated in terms of argument categories, this question reduces to the range of possible argument cases, i.e. what range of predictable interpretations an argument case may have. Most interesting is whether cases need to refer to the most difficult to define thematic roles, agent and theme/patient, or to argument categories correlating with them. I turn now to a way to formulate syntactic case without "agent," "patient" or argument categories loosely correlating with them (so no AGT or PAT). Some cases seem to mark subjects and objects regardless of thematic relations or argument categories. Such cases have traditionally been analyzed as being dependent on grammatical functions. Case rules then take the form of statements such as "Subject gets nominative" or "Highest available grammatical function gets nominative." For the Germanic languages and many others, the two cases which correlate with grammatical functions are nominative and accusative. I will, however, argue that a characterization of these cases in terms of the notion of argument removes the need for functional information other than obliqueness in capturing case marking. The only information needed for the environmental specification of these syntactic cases is whether an NP is an argument or not. In languages like the ones I am considering in this chapter, the nominative marks NPs only if they are arguments and the accusative marks NPs without restriction: 9 (46)

a. b.

NOM: [ARG] ACC:

[ ]

Ever since Jakobson's (1936) pioneering study of Russian case, nominative has usually been taken to be the lack of case. 10 Since Jakobson, the further assumption has been made that nominative is also the Elsewhere in the domain of NPs that must receive case, but a notable exception is Gaedicke's (1880) analysis of Vedic Sanskrit, in which he argues that there is no core function of the accusative but rather the accusative fulfills whatever functions are not fulfilled by the other cases.11 He thus unifies the long and disparate list of uses of the accusative under a notion of accusative-as-default similar in many respects to the one proposed here. Delbriick (1893: §§70, 176) follows this approach

40

Argument case and case alternations

in his analysis of Indo-European accusatives. Whitney (1892: 286) objects to this view on the grounds that any other case could be formulated as the case that fulfills those functions not fulfilled by any other case. This misses the point, however, spelled out at length by Gaedicke, that the greater difficulty in unifying the more disparate functions of the accusative than in unifying the functions of the other cases is the advantage to having accusative as a default. We gain more by way of economy of rules (or their equivalents) by postulating a default accusative than a default dative, a default nominative, etc. Nevertheless the majority view these days is that nominative is the Elsewhere as assumed in most versions of the Principles and Parameters framework, where NPs that do not receive Case by some other source are forced to move to receive nominative (Chomsky (1981)). However, (46) distinguishes morphological zero or default case from default case for linking.12 I now present some arguments for why accusative in the languages in this chapter should be analyzed as the ultimate Elsewhere for linking. The important difference between nominative and accusative is that accusative can mark non-arguments, NPs that Jakobson, following Peskovsky, termed weakly governed accusatives. These are accusatives that bear a looser relation to the verb, i.e. in our terms, they are not subcategorized for and so are non-arguments. They occur freely, leading to multiple accusatives per clause. Nominative never marks nonarguments. The first class of weakly governed accusatives are the accusatives of path. Both Icelandic and German have this construction, as demonstrated for Icelandic by the following (Zaenen, Maling and Thrainsson (1985)):

(47)

a. Hann keyrdi bilinn fcessa he(NOM) drove car-the(ACc) this(ACc) route(ACc) 'He drove the car this route.' b. Bilinn var keyrdur J>essa leid. car-the(NOM) was driven(NOM.m.sg) this(ACc) route(ACc) 'The car was driven this route.' c. *J)essi leid var keyrd bilinn. this(NOM) route(NOM) was driven car-the(ACc) 'This route was driven the car.'

Case, argument structure, and interpretation

41

The semantic contribution of the accusative of path is like that of an adverbial, not that of an argument. Nor does it normally passivize, leading to the view that it is not an argument.13 Icelandic exhibits other nonargument accusatives, such as the accusative of time (48) and duration (49) adverbials (see also Kress (1982: 224)): (48)

t>ad rigndi vikuna sem leid. It rained week-the(Acc) REL passed 'It rained last week.' (Einarsson (1949: 106))

(49)

Eg

var t>ar tvo daga. was there two(Acc) days(Acc) 'I was there two days.'

I(NOM)

(Einarsson (1949: 106))

German has similar constructions (Dal (1966: §13)). What is predicted by the case specifications in (46) above is that accusative will show Elsewhere behavior. This it does in the languages under consideration. I conclude that the accusative is the ultimate default case in the languages discussed.14 This has the immediate advantage of unifying the treatment of many uses of accusative. Instead of having separate rules for accusative of object, accusative of path, cognate accusative, and adverbial accusatives, accusative is simply the least restrictive case.15 We are now in a position to ask about the syntactic nature of the various cases involved in DS and NS. As we have seen, the dative is for syntactic experiencers and the nominative for arguments. The accusative is used for a much smaller class of experiencer verbs. This is best treated as an idiosyncratic case, because there is no semantic generalization which distinguishes it from the other cases, and the class is very small. I will return to the lexical accusative and alternative ways of analyzing it below.

2.2.4 Three constraints Every NP has to be linked, but this does not tell us what argument any NP will be linked to, nor how the linking will be realized (as a case, a particular configuration, or agreement). These two aspects of linking can be expressed formally using the attribute-value matrix notation introduced above. The problem, then, is to devise a substantive theory of which NP is linked to which argument and how the linking is realized. Three constraints are now offered that together constitute such a theory.

42

Argument case and case alternations

2.2.4.1 Linking and the applicability constraint It will be assumed that case frames are generated freely subject to constraints, which will rule out many frames. I will assume that linkers are like rules in that they apply wherever they can: (50)

Linking Condition: All NPs must be linked, and an NP is not linked unless there is no linker which has not applied but whose application would not violate any other constraint.

The Linking Condition ensures that every NP has a non-null value for 16 REAL, i.e. that it will be linked. So far, any linker can link any NP. But we want to narrow down the class of licit linkings on the basis of the LINK value of the NP, i.e. on the basis of properties such as argumenthood. Three constraints will suffice to narrow the class of case frames to the ones we want by constraining the relationship of an NP's linking properties (LINK) and the linking it actually undergoes (REAL). The first constraint is the Applicability Constraint (AC). This simply says that the specifications of the linker must unify with those of the NP. Otherwise the case would not be applicable. Some NPs are arguments of verbs. This is represented by having that NP contain the specification ARG for LINK, [ LINK [ ARG ]], and, if the argument is direct (the default argument category), it will be [ LINK [ ARG [ CAT [ DIR ]]]]. Cases also will have LINK values to determine their distribution. For example, I have argued that nominative should be a case applicable in principle to any argument and so should have the LINK value [ARG]. SO to an NP direct argument the nominative is applicable, but a case specified [ LINK [ ARG [ CAT [ EXP ]]]], e.g. one of the types of dative in Icelandic, will not unify with that NP, since EXP and DIR will clash. The AC requires the case and the NP it links to have unifiable LINK specifications: (51)

Linking is the unification of an NP and a linker.

A corollary to this is the Applicability Constraint: (52)

Applicability Constraint: An NP and its linker must unify.

I will say that a linker is applicable to an NP if and only if it can unify with the NP. Thus if the linker's LINK value contains some distributional feature incompatible with the LINK value of the NP, then the result of using that linker to link the NP is ill-formed. This is what is meant by LINK'S being distributional information for the linkers. The LINK value in the lexical entry of the linker will, in conjunction with the AC, determine

Case, argument structure, and interpretation

43

its distribution. Take nominative as an example. The NP argument and the nominative linker unify, resulting in an argument that is specified for morphosyntactic category. The verb gefallen 'please' in (53) has an abbreviated entry like (54) and a full entry like (55). Unifying the verb's entry and that of the nominative (56) gives (57): (53)

Mir gefallt der Aufsatz. me(DAT) pleases the(NOM) essay 'I like the essay.'

(54)

gefallen (

(55)

[V]

y ) 'please'

X|EXP

CAT [ N P ] LINK [ARG [CAT [EXP [PLACE [1]]]]] SUBCAT

REAL [

]

CAT [ N P ] LINK [ARG [CAT [DIR [PLACE [2]]]]] L REAL [

(56)

]

Entry for Nominative LINK [ARG] 1 REAL [NOM] J

(57)

CAT [V] CAT [ N P ] LINK [ARG [CAT [EXP [PLACE [1]]]]] SUBCAT

REAL [

]

CAT [ N P ] LINK [ARG [CAT [DIR [PLACE [2]]]]] REAL [NOM]

Note that the nominative would fail to unify with the first argument of this verb, which will have to be linked by dative. 2.2.4.2 Restrictiveness and the Restrictiveness Constraint So far we have only provided necessary conditions for the appearance of any given case, not sufficient ones. Nominative does not in fact appear on all arguments, nor does accusative appear on all NPs. In this section and the next, I will present two further constraints on case which will make case frames fully predictable.

44

Argument case and case alternations

These two constraints use the independently needed distributional information from the last sections, the necessary conditions on case distribution (the value of the feature LINK). The fact that the remaining constraints converge on the same set of features provides considerable evidence for their identity as a class. In this section, I introduce the most important constraint of the system. Intuitively, it says that a linker that carries information that properly contains the information carried by another linker will have priority in places where they both might apply. This is a more precise version for cases of the specific-blocks-general principle, originally due to Panini, and an extension of the Elsewhere Condition in phonology (Anderson (1969), Kiparsky (1973), Koutsoudas, Sanders, and Noll (1974)). On the present view, there are no case rules and hence no rule ordering for cases. The ordering effect is achieved by means of the notion of restrictiveness: (58)

A linker X is said to be more restrictive than a linker Y if and only if X's LINK value (linking specifications) is properly subsumed by Y's.

Restrictiveness is based on subsumption and the already motivated value of LINK, and transitivity is built in. Looking back at some Icelandic cases discussed so far, the idiosyncratic cases and argument category cases (which both fit the schema [ARG|X], X ^ 0 ) are more restrictive than nominative or accusative, since their linking specifications are properly subsumed by those of the latter two (grammatical) cases. Furthermore, nominative is more restrictive than accusative, since [ARG] is properly subsumed by [ ]. The Restrictiveness Constraint (RC) uses this notion of restrictiveness to derive the priority of cases. In general, more specific cases have priority over more general ones: (59)

Restrictiveness Constraint: If an NP is linked by a linker A, and there is a linker B that is more restrictive than A and violates no other constraint, then linking the NP with A is ill-formed.

This relation replaces any notion of rule ordering. What the Restrictiveness Constraint says is that if any linker applies (and at least one must by the Linking Condition (50)), then that linker will be the most restrictive one available. It is now clear in what sense the non-lexical accusative is the default case. Since the accusative's specifications sub-

Case, argument structure, and interpretation

45

sume those of all the other cases, the accusative must be less restrictive than any other case, hence its default status for linking. The RC does not by itself rule out multiple linking. In most but not all cases, only one linker is used per argument. I will assume for present purposes that case and position are always compatible but that two cases will always be incompatible.17 Case and position are linkers of different types. The RC imposes 'disjunctive ordering' only when the linkers are incompatible, but linkers of different types are in general compatible. One property of Icelandic is that the subject, no matter what its case, must in non-inverted sentences come before the verb. This can be captured in the present system by adding to the inventory of Icelandic linkers a second nominative-like linker: (60)

_V:

[ARG]

The symbol "_V" is for a morphosyntactic feature for preverbal position and will indicate that the preferred position of this NP is before the verb and, in inverted sentences, just after it.18 The identity of its specifications to those of NOM will cause the preverbal configuration and nominative to behave similarly with respect to the RC. The fact that they belong to different morphosyntactic types - one is a case, the other a position will cause them to interact with other linkers differently. An example will illustrate. Take the sentence in (5b), repeated here: (61)

Barninu batnadi veikin. child-the(DAT) recovered-from disease-the(NOM) T h e child recovered from the disease.'

The following may apply to the first argument: (62)

a. DAT: [ARG|EXP]

b. NOM: [ARG] c. _V: [ARG] d. ACC: [ ] By the RC, the dative must apply since it is the most restrictive. Nominative and accusative cannot apply, because they are of the same type as the dative. As we will see next, nominative (62b) is left for the second argument, to which it correctly applies. However, the preverbal positional linker may also apply to the first argument, leading to a dative subject in preverbal position.

46 Argument case and case alternations 2.2.4.3 A dissimilation constraint on linkers The third and last constraint based on LINK values limits the use of all but the default case. For instance, nominative does not mark all arguments, but rather as Zaenen, Maling, and Thrainsson (1985) note, it marks only the highest available grammatical function. Our last constraint will make this follow. It will be easier to see how this third constraint works if we consider more closely what it is supposed to do. Intuitively, we can think of case marking in a language as working top-down according to the following procedure: (i) Link the highest argument with the linker that passes both the AC and the RC, i.e. the most restrictive case available, (ii) Move down the list to the next argument, (iii) List all the linkers that could apply to the argument, (iv) Delete from the list any linker that could already have been used, i.e. delete a linker (e.g. a case) if it is the most restrictive applicable (AC) linker of that type (e.g. case) to any "earlier" (higher) argument, (v) Use the most restrictive case remaining on the list, (vi) If another argument remains on the argument list, go to (ii); otherwise stop. 19 In other words, for arguments lower on the argument list, avoid the use of linkers that would otherwise be applicable "in principle" to a higher argument. Assuming that linking is "top-down" then this amounts to saying, very roughly, do not use a linker if it could already have been used earlier. So, to take nominative as an example, one could not use it on the second argument in (63) ( = (6)), because it "could" apply (and does) to the higher argument: (63)

Eg

se frig.

I(NOM) see you(Acc)

'I see you.' Rather the second argument gets the remaining applicable case, the accusative, since the nominative has been preempted by the highest argument. As mentioned earlier, I will assume that case frames are freely generated and subjected to constraints, rather than assuming that linking "proceeds" top-down (or bottom-up). Therefore I have formulated the constraint in (64) and its graphical version in (65) to say that no two arguments of the same predicate may get the same case if there is another alternative for the lower one, and similarly for word order and agreement:

Case, argument structure, and interpretation

Al

(64)

Linking Dissimilation Constraint: Given two co-arguments, x and y, x higher than y, if y is linked by a linker B that has the same LINK value as the most restrictive linker of that type applicable by the AC to x, call it A, and there is a less restrictive linker C, applicable by the AC to y, then the linking is ill-formed.

(65)

*verb (... x A: [a]

... y ... ) B: [b] C:[c] where A and B are of the same type, a = b, and C is applicable by the AC to y (c unifies with y's LINK value)

In (65) and henceforth, applying cases are noted in italics. Note that this constraint is asymmetric in that it "questions" the linking of the lower argument y but not the higher argument x. It is this asymmetry that gives the system its "directional" procedural interpretation, which will become clear with the next example. As noted before, an alternative to (64-65) would be to have linking proceed top-down. The higher arguments would then correctly have priority over the lower ones. This means that on the present approach the nominative is correctly predicted to be used only once in languages like Icelandic and German, on the assumption that nominative is not also a lexical case. A "second" use of the nominative would violate the LDC. This is borne out: in Icelandic, a sentence like (63) above or (66) is well-formed only with the accusative on the second argument, not with the nominative:20 (66)

Hann lamdi mig/*eg. he(NOM) hit me(ACc)/me(NOM) 'He hit me.'

First, the verb lemja 'hit' has the following argument structure: (67)

lemja ( x y ) 'hit'

Then let all the cases and other linkers apply that can: (68)

lemja ( x

y

NOM

V ACC

) 'hit'

*NOM

*_V ACC

Use of either of the two linkers with an asterisk would lead to a violation of the LDC, and here and in the future I will indicate with an asterisk

48

Argument case and case alternations

linkers whose use at an argument would violate the LDC. I list the cases in the order that restrictiveness supplies, an order that is, of course, not stipulated. The case that actually applies, indicated by italics, is the most restrictive one. Note the order in which the constraints apply. The AC is an absolute constraint, and one can tell if a structure violates it without reference to any other constraint. The LDC depends on the AC, since it asks if there is a less restrictive alternative under the AC for the lower argument. In other words, it does not rule out a structure unless there is another linker that passes the AC on the lower argument. The RC depends on both the other constraints. It says that of all the applicable linkers the most restrictive must apply. Applicability here depends on both the AC and the LDC. Linkers that do not pass the AC for an argument are not favored by the RC whether they are more restrictive than the applicable linkers or not. And a less restrictive linker can be used on lower arguments if the more restrictive linker(s) would violate the LDC. The last example illustrates this nicely: the RC does not force the LDC-violating nominative on the lower argument in (66), despite the fact that it is more restrictive than the accusative.

2.3

Basic clause structure

In the last section, I illustrated how the three constraints narrow in on the case frame for a basic transitive. This section will demonstrate that these constraints handle basic clause structure generally. Turning first to the passive of a simple transitive like that from the last section, we find that things work similarly. The passive counterpart of (66) is (69): (69)

Eg

var

I(NOM) was

laminn (afhonum). hit

by him(DAT)

'I was hit (by him).' As mentioned in the previous chapter, I adopt a suppression view of the passive (Kiparsky (1987), Bresnan and Kanerva (1989), cf. Chomsky (1981) and see Chapter 4). The top argument in the lexical entry in (67) is suppressed, i.e. it is no longer treated as being on the argument list:21 (70)

lemja-PASS ( xA y ) 'hit'

Basic clause structure

49

Now let all the linkers apply to this derived structure: (71)

lemja-PASS ( xA

y

) 'hit'

NOM

ACC

Since in the passive the top argument is suppressed, neither the nominative nor the accusative may apply to it, leaving the two syntactic cases to compete for the only remaining argument. The nominative, as the more restrictive, wins. The account of the passive also relies on the nominative's being more restrictive than the accusative. Since the analysis of the passive relies on the different status of the nominative and the accusative, it is another confirming instance for the analysis of Icelandic nominative as the argument case and the accusative as the default. The priority of lexical case follows directly from the RC. Consider again the verb tyna 'lose' and its passive: (72)

a. Barnid tyndi bokinni. child-the(NOM) lost book-the(DAT) 'The child lost the book.' b. Bokinni var tynt (af barninu). book-the(DAT) was lost by child-the(DAT) 'The book was lost (by the child).'

In these two sentences, the following linkers are applicable: (73)

a DAT: [ARG|2] b. NOM: [ARG] c. _VP: [ARG] d. ACC: [ ]

(part of the lexical entry for tyna 'lose')

Taking the active in (72a) first, we get the linking in (74): (74)

tyna ( x

y)

'lose'

NOM

DAT

_ VP

*NOM

ACC

*_

VP

ACC

The only difference from the active clause previously considered is the applicability of the dative case in (73a) for the second argument. Here the dative blocks the accusative, since the environment in (73a) properly

50 Argument case and case alternations includes the environment in (73d). Restrictiveness automatically captures the priority of lexical case. In the passive in (72b), the dative likewise blocks the nominative as well as the accusative: (75)

tyna-PASS ( xA

y)

lose'

DAT NOM

_VP ACC

The dative in (73a) is more restrictive than the nominative in (73b) (and the accusative in (73d)) and so wins. The theory correctly predicts a sentence containing only a dative. Thus the effects of all types of rule ordering including the prevalent "pre-linking" of lexical case (Yip, Maling, and Jackendoff (1987)) come without any stipulations beyond those independently necessary to capture the distribution of cases in the first place. The essential feature of this account is that it takes the values which comprise the value of LINK and which are necessary to get cases to be applicable in the right places and uses them to derive their "order" of priority. As there is no a priori reason why the distributional features of a linker should define linker priority, the result is all the more unexpected. In this section and in many examples to follow, it is clear that one special type of priority, "case preservation," follows directly from the restrictiveness-based theory. The priority of idiosyncratic case depends on the relation of the idiosyncratic case to the syntactic cases. All idiosyncratic cases are more restrictive than the syntactic cases and so have priority. This of course generalizes to the other type of lexical case, argument category case. An argument category case like the dative of goal is more restrictive than both the nominative and the accusative and so will be preserved in Icelandic, a fact that I will discuss in Chapter 5. Finally, the LDC also provides for so-called "nominative objects."22 To illustrate, we now provide the full derivation for the nominative object example from the last section: (76)

batna

(X|EXP

y

DAT: [ARG/EXP]

NOM: [ARG]

NOM: [ARG]

*_V:

_ V: ACC:

[

[ARG]

]

) 'recover-from'

ACC:

[ARG]

[

]

Case alternations and restrictiveness

51

Here nominative is fine on the lower argument, because nominative is not the most restrictive case applicable to any higher argument. Dative, not nominative, is the most restrictive case applicable to the first argument. Nominative objects are correctly allowed just when there is some more restrictive case for the higher argument.

2.4

Case alternations and restrictiveness

It turns out that this restrictiveness-based theory of case handles the problems of DS and NS from §2.1 automatically. Both alternations will follow from the optionality of the original case. When that case (accusative in DS, accusative and/or dative in NS) cannot apply, the next most restrictive case is predicted to apply. In DS, the next most restrictive case is the experiencer dative, and in NS the next most restrictive case is the argument nominative. The failure of the original case to apply forces the replacement case to apply. This is accomplished without any actual replacement. The mechanism inherent in the restrictivenessbased theory provides exactly what is needed to capture the alternations. The limitations on DS follow automatically from how the theory is set up.

2.4.1 Non-substitutions DS involves the lexical accusative and the lexical dative. Experiencers of some verbs get accusative, and experiencers generally get dative. In the present theory, these two cases are as follows: (77)

ACC: [ARG|EXP|1]

(part of lexical entry of langa 'long,' of bresta

'lack,' etc.) (78)

DAT: [ARG|EXP]

The accusative applies to the experiencer logical subjects of various verbs. It is an idiosyncratic case analogous to the dative with tyna 'lose' and the genitive with sakna 'miss' of §2.1.2. The method chosen to represent this is not as important as the fact that this accusative mentions something more specific than the dative. Any other way of representing the information that reflects this fact would work just as well.23 It is clear that anything to which accusative in (77) can apply is something to which the dative (78) could apply but not vice versa. By the RC, the accusative must

52 Argument case and case alternations apply, and, since both cases may not apply, the dative (along with other less restrictive cases such as nominative) is blocked. This correctly describes the state of affairs with most verbs in earliest Old English and Old Icelandic (see Chapter 6) and non-DS accusative psych verbs in German and Icelandic. It equally follows from the RC that, if the accusative is unavailable for any reason, the dative will apply (and block the rest of the cases). Therefore DS can be completely captured as the optionality of the lexical accusative. DS results from a quadruple case competition for the first argument of the verbs in §2.1, combined with the optionality of the most restrictive case. When the most restrictive case, the lexical accusative, becomes optional, then the next most restrictive case, the dative, should be able to apply: (79)

a.

ACC: [ARG|EXP|1]

(part of lexical entry of langa 'long,' bresta

'lack,' etc.) b. DAT: [ARG|EXP]

c. NOM: [ARG] d. _V: [ARG] e. ACC: [ ] The simplest situation is DS with one-NP verbs like langa 'long' for speakers who do not accept NS with this verb (i.e. find (2) above ungrammatical). Such speakers get both (la) and its DS variant (lb), repeated here: (80)

Mig/Mer langar ad fara. me(Acc)/me(DAT) longs to go 'I long to go.'

The lexical entry for such speakers is as follows: (81)

langa

(X|EXP (ACC: [ARG|EXP|1])

y V p) 'long'

The lexical accusative in (81) is optional, as indicated by the parentheses. As usual, this device for capturing optionality means that what is contained in the parentheses is an optional part of the entry. It is an abbreviatory convention for collapsing what amount to two lexical entries, one with the accusative, one without. It is important, however, that we not be able to collapse any two possible entries in this way. In fact, entries of the type (81) can relate

Case alternations and restrictiveness 53 only the disjunction of something and its absence, not something and something else, as would be possible if we used full-fledged disjunction (curly brackets). Thus, DS on this account only involves the power to relate accusative to its absence. It does not allow us to relate accusative directly to dative which would carry with it the undesirable power to relate accusative to anything else, e.g. genitive. The linking for langa 'long' on the non-NS dialect is straightforward. The lexical entry can be written with or without the accusative, leading to two linkings. First, if the accusative option is taken, we get the following: (82)

langa ( X|EXP

yVp

ACC:

) iong'

[ARG/EXP/1]

DAT: [ARG|EXP] NOM: [ARG]

_ V:

[ARG]

ACC:

[ ]

This is the linking for the accusative variant. By the RC, the accusative must apply. No other case can be compatible, but the position is compatible, so it also applies. This argument is the highest non-obliquely linked argument and so is the subject. The non-accusative option is as follows: (83)

langa ( X|EXP

y V p) 'long'

DAT: [ARG/EXP] NOM: [ARG]

_ V:

[ARG]

ACC:

[ ]

The RC mandates that the most restrictive case must apply, which here is dative. So the dative is the case that applies. This yields the DS variant. NS follows in a similar fashion. Recall that many semantic experiencers get nominative. This followed from their NPs' not bearing the syntactic value EXP. Thus one view of NS is as the variation of a verb between the one class and the other, i.e. variation between the presence and absence of EXP. For speakers who accept (2) (i.e. have NS with langa 'long'), the first argument is optionally a syntactic experiencer. Generally, for any verb, if EXP is optional the NS effects follow. So, to the lexical entry in (81) above, we add the optionality of EXP: (84)

langa (

X|EXP)

y V p) 'long'

(ACC: [ARG I EXP 11])

54 Argument case and case alternations This does not necessarily mean any semantic difference, since, even if the argument is not constrained to be interpreted as an experiencer by the feature, it may nonetheless be interpreted this way on account of the verb's meaning. Note that, if the EXP option is not taken, neither accusative nor dative can apply, leaving nominative as the next most restrictive case. By the RC, nominative must apply, yielding NS. Note again that optionality allows us to relate an entry with EXP and one without. It does not allow one to relate EXP to anything else but its absence. The NS verbs merely vary between two independently existing classes, the EXP class and the direct, nominative class, e.g. sjd 'see.' To demonstrate concretely how this works, consider both options. The option with EXP will, of course, be the same as the last two examples. What is new for this dialect is the option without: (85)

langa ( x

y V P ) long'

NOM: [ARG]

_V:

[ARG]

ACC:

[ ]

The most restrictive applicable linkers are nominative and preverbal position, and they are both compatible, so by the RC they apply. Note that this variant of langa 'long' is just like sjd 'see.' The non-EXP option puts the verb in the nominative experiencer class. It was noted in §2.1 above that for accusative verbs DS is not a necessary condition for NS. A verb can fail to show DS but still show NS, as is true of German hungern 'hunger' which takes accusative or nominative but not dative (for some speakers): (86)

(87)

Mich hungert. me(ACc) hungers 'I am hungry.' Ich

hungere.

I(NOM) am-hungry

The present system is equipped to handle this, because the two uses of optionality in lexical entries are independent. One can construct a lexical entry where the accusative is not optional, but the EXP classification is: (88)

hungern ( X|EXP ACC:

) 'hunger' [ ARG EXP 1 [OBL

Case alternations and restrictiveness

55

If the non-EXP option is taken, neither the accusative nor the dative can apply, since they are both specified for EXP. The nominative is the most restrictive remaining case. This example is also interesting because it shows how we can capture the independence of DS and NS from the oblique subject phenomenon of Icelandic. Take German ekeln 'be disgusted,' which varies between accusative, dative, and nominative: (89)

ekeln (

) 'be disgusted'

X|(EXP) [OBL

The three linkings are as follows: (90)

ekeln ( X|EXP

) 'be disgusted' /

ACC:

ARG/EXP/1

I OBL

/

J

|~ARG|EXP]

DAT:

[OBL

NOM: [ARG] ACC:

(91)

[ ] 'be disgusted'

ekeln ( X|EXP DAT:

ARG/EXP

OBL NOM: [ARG] ACC:

(92)

[ ]

ekeln ( x

) 'be disgusted'

NOM: [ARG] ACC:

[ ]

Note that, if the accusative or dative applies, the first argument will be oblique, since the German lexical cases are oblique. If neither of these cases apply, leaving the nominative to apply, then (and only then) will the first argument be the highest non-obliquely linking argument and so be the subject. In German, both DS and NS are compatible with oblique lexical case. In Icelandic, which also shows DS and NS, the lexical cases are not oblique. The present theory, by making obliqueness dependent on the cases, captures the independence of the oblique or non-oblique status of lexically case-marked NPs on the one hand and the case alternations on the other.

56 Argument case and case alternations 2.4.2 Further predictions It was noted in §1 that transitive DS verbs show an interesting interaction of DS and NS. On any explanation of DS and NS it should follow that the second argument can show NS but not if the first argument does. The present approach to DS and NS extends naturally to the transitive verbs. The only additional factor is the second argument and its accusative: (93)

ACC: [ARG|2]

(part of lexical entry of bresta 'lack,' etc.)

An accusative specified as in (93) is present in the lexical entry of each of the ACC-ACC verbs. As with the first accusative, this second accusative is optional, as illustrated for bresta 'lack': (94)

bresta ( X|EXP (ACC)

y

) 'lack'

(ACC)

The optionality of this second accusative in conjunction with the theory of case above gives all the correct frames for the most liberal speakers. In the following, for each demonstrated linking the corresponding example is given on the right, as well as the case frame it produces. First, the ACC-ACC version is given in (95), where both accusative options are taken: (95)

bresta bresta (( X|EXP

y

ACC

ACC

DAT

NOM

NOM

_v

) 'lack'

ACC-ACC

(23a)

*_v ACC

ACC

If neither accusative option is taken, we get the (96)

bresta ( X|EXP

y

DAT

NOM

NOM

*_y

_V

ACC

) 'lack'

DAT-NOM

variant:

DAT-NOM

(24a)

ACC

As before with langa 'long,' the next most restrictive case for the first argument is dative, since it qualifies as a syntactic experiencer. The second argument does not, and so the next most restrictive case for the second argument is nominative, not dative. This accounts for the ungrammaticality of DS with this argument (see (19) above). It is also clear why DS should be equally ungrammatical even when this argument

Case alternations and restrictiveness 57 is the only overtly expressed one (and so subject). Either way, it does not qualify for dative and so cannot show DS. Nothing in the present theory makes the choice of the two accusatives dependent on each other. Rather they are logically independent. If the no-Ace option is taken on the first argument, but the ACC option is taken on the second one, then we get the DAT-ACC frame: (97)

bresta ( X|EXP DAT

NOM

y

) lack'

DAT-ACC

(25)

ACC

NOM

y

*_v

ACC

ACC

Contrariwise, taking the ACC option on the first argument and the no-Acc option on the second gives the ACC-NOM frame: (98)

bresta ( x EXP ACC DAT NOM

y

'lack'

ACC-NOM

(26)

NOM

*_v

ACC

_v ACC

However, the one frame that is not allowed is NOM-NOM, since it would violate the RC (and even if the logical subject were not EXP, such a frame would be ruled out by the LDC under the assumption that NOM is not a lexical case), and indeed such a frame leads to unacceptability (27). Quite generally, the simple optionality of the lexical case automatically leads to the application of the correct case. Most interestingly, the application of DS or NS and their interaction follow automatically as consequences of the optionality of the accusatives. What the accusative alternates with need not be specified but follows from general principles and the otherwise motivated grammar of Icelandic. Further, the transitive verbs can show NS for some speakers. A small number of Svavarsdottir's (1982: 3Iff.) informants gave nominative for the experiencers of some normally accusative and dative verbs on her questionnaire. Likewise, two out of the seventeen participants in Andrews' (1990: 204) survey found (99b) fine, and seven more found it marginal (unsurprisingly, all seventeen found (99a) grammatical):

58 Argument case and case alternations (99)

a. Stulkuna vantar efni girl-the(ACc) lacks material(NOM/Acc) 'The girl lacks material for the paper.' b. Stulkan vantar efni girl-the(NOM) lacks material(NOM/ACc)

i ritgerdina in paper-the(Acc) i ritgerdina. in paper-the(ACc)

The neuter noun efni 'material' is ambiguous between nominative, accusative, and dative, but in (99b) it is likely to be accusative (double nominative never occurs in languages like Icelandic - see Chapter 3 for the other type - and unambiguous dative is not possible here). The accusative on the second argument in (99b) is ambiguous between a lexical accusative and a syntactic accusative. I tentatively assume for vanta 'lack' that the second argument's accusative is lexical, because its accusative seems not to be optional for most speakers (see (24) above), but of course this may not be true of Andrews' two speakers: (100)

vanta ( xX

yy

) 'lack'

NOM

ACC

_v

*NOM

ACC

NOM-ACC

(99b)

*_v ACC

Verbs of this class show clear instances in the intransitive use (cf. (18)) of what on most theories must be lexical accusative, and, on the present proposal, the normal ACC-ACC and the DAT-ACC uses must involve a lexical accusative for the logical object. Thus the simplest assumption is that this accusative is lexical here, but the other assumption would lead to no further problems. Note, finally, that this theory is equipped to capture the full range of speaker variation as to which subset of the case frames above are acceptable. Less liberal dialects just have optionality in fewer places. Furthermore, this can vary from verb to verb. So, for example, Yip, Maling, and Jackendoff (1987: 23If.) report that vanta 'lack' usually can show DS on the first argument but not NS on the second. A lexical entry for vanta may be set up with only the first accusative optional for such speakers: (101)

vanta (

X|(EXP)

y

(ACC)

ACC

) 'lack'

Case alternations and restrictiveness

59

This entry will lead to frames with accusative only for the second argument, whether or not it is the only overtly expressed argument. Once again, this follows from the cases and the RC. Being non-optional, the accusative will always compete for the second argument. As the most restrictive competing case, it will always block the others. Other variations by speaker and lexical item can be handled in a similar way. This account also correctly predicts when DS is not an option. Some verbs take a non-experiencer accusative on their subject. One such verb was reka 'drift' (20). This verb can have accusative or nominative but not dative: (102)

Batinn/Baturinn/*Batinum rak til hafs boat-the(Acc)/boat-the(NOM)/boat-the(DAT) drifted to sea 'The boat drifted to sea.'

The lexical entry for this verb includes an idiosyncratic accusative: (103)

reka ( x

y ) 'drift'

(ACC: [ARG|1])

When the accusative option is taken, that case is the most restrictive and will apply. When the accusative option is not taken, the nominative applies as the next most restrictive applicable case. Since reka 'drift' has no syntactic (or semantic) experiencer, the dative is out. The impossibility of DS and possibility of NS with non-experiencers both follow. The present approach makes the solution to DS and NS an automatic consequence of the optionality of some element. Because everything follows from simple optionality, the descriptive and explanatory problems are both addressed. First, as for the descriptive problem, it was emphasized that DS shows many restrictions that NS does not. These were the restrictions to accusatives and to experiencers. These limitations on DS follow from the subsumption relation of the lexical accusative of the DS verbs and the dative of the argument category experiencer: accusative mentions everything dative does (argumenthood, experiencerhood) and some verbspecific information besides, making the lexical accusative more restrictive than the experiencer dative. For there to be alternation with an experiencer dative, the other case must be more restrictive, which means in turn that it must be an idiosyncratic case for experiencers. The only case like this in the languages in question is accusative. The fact that DS arguments must "become" dative follows from the fact that it is the dative

60

Argument case and case alternations

that picks up that argument when its idiosyncratic case option is not taken. This dative is for experiencers, and so the experiencer limitation on DS is built into the dative, where it is needed for independent reasons. As for NS, the fundamental similarity of this alternation with the other one is captured by their both involving optionality. In both alternations, the zero option leads to the application of a less restrictive case. In NS, the take-over by the nominative is also automatic. But since nominative is not limited to any particular argument category, so too NS is automatically not limited, which is empirically correct. The fact that NS is not limited to affecting accusatives follows from the fact that the specification of nominative subsumes the specifications of all the lexical cases.24 Thus, any optional, more restrictive case will automatically show NS when nominative is available. Furthermore, the present approach addresses the explanatory problem of the case alternations. The excessive power to relate any two cases is not inherent in this system. Rather, the AC, RC, and the LDC place strong limits on case frames and case alternations while correctly characterizing the actual alternation types.

2.5

Why argument structure?

The theory of case proposed above and developed in the following chapters is equipped to handle a wide variety of facts in a constrained way. From the argument structure of any verb, the notion of obliqueness as a property of cases, and the three constraints (the AC, the RC, and the LDC), it is possible to derive the functional structure of clauses. Before concluding this chapter, it is natural to ask why argument structure has been chosen for such a fundamental role in the theory. In this section I look at this question from three angles. First, the argument categories proposed above are not only useful in capturing facts like the case alternations, as we have seen, but they also may explain why DS and NS are both very common in Germanic, and NS is common cross-linguistically, but DS is rare to non-existent outside of Germanic. The argument categories depend on a central role for argument structure. Second, the constraints proposed here, especially the RC, would not work if formulated at a level other than argument structure. In particular, an RC defined on some level of thematic or semantic structure will not cover as much ground as the RC in the present theory. Finally, by having the theory

Why argument structure? 61 operate on argument structure, the operation of the three constraints can be used as a diagnostic for arguments and adjuncts.

2.5.1 Argument categories and case alternations By basing the theory of linking on argument structure, and argument categories in particular, it is possible to explain one of the most intriguing differences between the two case alternations addressed in the last sections. One of their most striking similarities is that they are both common in Germanic, and one of their most striking differences is that only NS is common outside of Germanic. The present theory generates a proposal that the cross-linguistic rarity of DS can be derived from the limited access of cases to "thematic" information. In fact, I know of no other example cross-linguistically of what I will call Generalized DS. This would be the alternation of two cases which are both more restrictive than the nominative, i.e. variation between (104a,b): (104)

a. [ARG|X|Y] b. [ARG|X]

c. [ARG] d. [ ]

(NOM in Icelandic-type languages) (NOM in "type 2" languages - see Chapter 3)

One answer to this problem may come from what can substitute for X and Y in (104a,b). If the features that may substitute for X and Y in (104) are few, we predict that the Generalized DS configuration is unlikely to arise. The DS phenomenon can be regarded as providing evidence for the syntactic argument category of experiencer. Without EXP in the context of the restrictiveness-based theory of case, we would lose our explanation of DS. The success of the explanation lends further support to the assumption it rests on, the positing of EXP. The assumption, more generally, is that EXP and ARG, etc. (parts of the value of LINK) form a class of features in the classical sense. Their identity as class is supported by the fact that various independent criteria select these features. The three constraints, AC, RC, and LDC, are three such criteria. However, it is not so clear what the exact boundaries of this class are. How many argument categories should we admit? For example, do we need argument categories like "lacker" which would be interpreted as the experiencer participant of the lacking relation? One might think so, since all the verbs of lacking in Icelandic take an accusative subject:

62 Argument case and case alternations (105)

Case:

[ARG|EXP|LACKER]

Another way to frame this issue is to ask how much of the semantics has been grammaticalized. The RC and the LDC would work with cases like (105), but they are not necessary to get blocking. If each verb of lacking has its own idiosyncratic case, it will also stand in the correct relation to the experiencer dative. One piece of evidence that verbs of lacking do not have a more general (less restrictive case) like (105) is that the verbs of lacking do not lose the accusative less quickly than other accusative subject psych verbs, as might be expected if one hypothesizes that less restrictive cases are more stable. Another source of evidence that there is no "lacker accusative" would be its lack of productivity: if new verbs of lacking do not take accusative but rather dative or nominative this would be evidence that speakers have not acquired a rule like (105). This would be difficult but not impossible to find out. Generally speaking, the accusative for the various experiencer arguments seems not to be productive (see Chapter 6). The importance of DS lies partly in the evidence it provides that the class of linking features should be as small as possible. If DS rests on the assumption of the category based on EXP, the further assumption that such features are not to be posited unless necessary can explain the actual rarity of DS. If argument categories are rare, then the generalized DS configuration in (104) is rare. NS, on the other hand, does not depend on there being two lexical cases in a subsumption relationship. All that is required is one lexical case on the one hand and an any-argument case (instantiated as nominative in Icelandic) or some other less restrictive syntactic case (see the next chapter) on the other. The any-argument linker is a feature of many if not most languages, so the conditions for NS are not as hard to fulfill. NS, then, is correctly predicted to be crosslinguistically common. The evidence that the rarity of DS provides for a maximally small set of linking features fits in well with the present approach. An important aspect of the restrictiveness-based theory of case is that the more problematic thematic roles agent and patient/theme do not need to be reconstructed here in order to handle grammatical case. The grammatical cases can be defined in terms of the notion argument, the order of arguments and their linking, and three constraints. Functional information can in turn be predicted from this information. As for the argument categories needed, I have focused partiularly on experiencer in DS. I have not

Why argument structure? 63 proved that it is exactly this category that is needed. The X in (104a,b) could be indirect object, the notion 3 in RG, generalized "goal" (which would require definition). Experiencer (EXP) is the smallest category consistent with the facts, and it works. The question of the exact characterization of the set of argument categories must remain open, but the present approach reduces the reliance on thematic role-like entities and makes more modest semantic claims, along the lines of AUX.

2.5.2 The semantic selection alternative All three constraints, the AC, the LDC, and the RC, work well if one assumes the vocabulary of argument structure: [ARG], argument categories, and individual arguments. If we now accept that the constraints express desirable generalizations, the question is how essential argument structure is to these constraints. I would like briefly to consider an alternative. Consider the RC. I have defined the RC on syntactic argument structure which is the input to translation in the standard way. There is no "level" of semantic structure. But there is a common view about case that it selects for Lexical Conceptual Structure (LCS) or some other "lexical semantic structure" (for example, Grimshaw (1990), Jackendoff (1990)). On this view, the RC would be a sort of semantic blocking principle. Cases that select for more of the "semantic structure" would block cases that select for less. So the dative of goal might select the TO function in LCS, where TO is a function from individuals to paths with that individual as an endpoint (Jackendoff (1987, 1990)): (106)

DAT: TO

If this view were correct, then we might try to do without the argument categories at argument structure and just have the Restrictiveness Constraint mandate the use of the case that selects for the most of this semantic structure. This is fine as far as "semantic" case versus "grammatical" case is concerned. The dative'might select the path function TO in LCS whereas nominative and accusative did not. The problem with this semantic selection view is that it can handle neither the least nor the most restrictive linkers. I have argued that, for languages like Icelandic and German, nominative is more restrictive than the accusative, and I will extend this view of syntactic case in the next chapter. But it is not clear that the nominative or the accusative selects

64 Argument case and case alternations for LCS or, if they do, that nominative selects for more than the accusative in Icelandic and German. Of course, LCS may be enriched to reflect information we need for nominative and accusative assignment, but only, it seems, in ways which cause LCS to lose its explanatory potential. For example, Jackendoff (1990: 252ff., 1993: 294) proposes annotations at Lexical Conceptual Structure that indicate which positions in the LCS correspond to arguments in the syntax (versus ones which correspond to adjuncts or no NP at all). Jackendoff takes the entry for "enter" as typical: (107)

/enter/ V [Event GO ([xhing

] A', bath TO ([ P i a c e IN ([xhing

] )])])]

Here, types of positions in such LCS are supposed to define certain theta roles. For example, since "Theme' is shorthand for 'first argument of GO, BE, SAY, or ORIENT'" (Jackendoff (1993: 290)), the argument that satisfies the lefthand open position in the entry is a theme. But the annotations, the subscripts "simply stipulate by means of an A that this constituent must be syntactically realized; (A) stipulates optional syntactic realization" (Jackendoff (1993: 294)). In other words, syntactic information realization as an argument in the syntax - is being explicitly stipulated at LCS. This being so, LCS is not a "semantic" level (whatever that would be), and explanations of argument linking involving LCS cannot be guaranteed to be anything other than syntactic. Surface expression plays a crucial role in determining what the LCS is. But if syntax plays a crucial role in determining LCS, then a theory of LCS is not a purely semantic theory nor one that explains or derives argumenthood. 25 Nominative could be formulated to be applicable to annotated positions. However, once this move is made, the claim that cases "select" LCS becomes meaningless, since the syntactic information we need for case marking is simply being imported into LCS for no other reason except to capture syntactic facts (e.g. subject selection and the argument adjunct distinction). More generally, the adoption of annotations at LCS which are motivated for purely syntactic reasons should call into question how "semantic" LCS actually is. Neither LCS, thematic roles, nor "semantic" explanations for syntactic phenomena in general can claim much significance or superiority over purely syntactic approaches when the "semantic" model is (covertly) syntacticially motivated.

Why argument structure? 65 At the other end of the range of cases, idiosyncratic case is also a problem for the semantic selection view. Two synonymous verbs can take different cases on their logical objects. Say that one of them takes an idiosyncratic dative and the other has normal grammatical case (accusative in the active). Now the dative cannot "select" for any part of the LCS of that verb that is not present in the LCS of the nonidiosyncratic case synonym.26 Thus to maintain that the idiosyncratic case selects for LCS, we would have to make any synonym not assigning idiosyncratic case an exception to the idiosyncratic case. On the other hand, if idiosyncratic case does not select for LCS, then it cannot be brought under the alternative Restrictiveness Constraint which on the LCS-based view refers to amount of LCS selection. So idiosyncratic cases would either have to have exceptions to their exceptional status, or they would fail to be part of the same case system as the other cases. In general, argument structure provides both the right primitives and the right place to carry out the kind of linking envisioned here. 2.5.3 Arguments and adjuncts One consequence of this theory is that it leads to a more precise characterization of argument and adjunct. The theory being developed here is mainly a theory of argument case. The specific-over-general principle has found many applications. Something like this specific-over-general principle seems to be at work in various aspects of grammar and general extra-grammatical behavior. The Elsewhere Condition is most famous in phonology (S. Anderson (1969), Kiparsky (1973), Koutsoudas, Sanders, and Noll (1974)) and more recently in morphology (e.g. S. Anderson (1986) and references there). The Elsewhere Condition is sometimes regarded as a special instance of a more general semantic blocking principle, which disallows paraphrases of lexical items, so that, for instance, the day after today is ungrammatical or at least strange compared to the synonymous lexical item tomorrow, and pink blocks light red. However, as S. Anderson (1986) points out, these semantic Elsewhere-like phenomena differ depending on the area of grammar. Thus pink blocks light red, but light red does not lead to ungrammaticality. The Gricean imperative to be brief does induce Elsewhere-like blocking, but it is not a rule of grammar. Violating it does not lead to ungrammaticality. A similar contrast can be observed for Elsewhere effects in arguments and adjuncts. One of the central claims of this chapter and the following

66 Argument case and case alternations is that argument linking operates according to the specific Elsewhere-like constraint I have called the Restrictiveness Constraint. If this is correct, the blocking behavior of linkers should be more like the strict one found in phonology and morphology and not like the loose Gricean one. The competition among cases is grammatical and absolute. The competition among adjuncts, however, is the competition between independent words. It is like the competition between pink and light red, and so is expected to be looser. In fact, many complements display exactly this behavior, and this class we will identify as adjuncts. Adjuncts, it is well known, are not identical with the class of optional NPs, since both arguments and adjuncts can be obligatory or optional. Uszkoreit (1987: (27a-d)) illustrates the four-way contrast with German examples like the following. The first sentence (108a) contains a postverbal obligatory argument, den Tisch 'the table,' the second sentence has an optional one, einen Brief 'a letter.' Recall too that the verbs of lacking in Icelandic generally had an optional first argument. The next sentence (108c) contains an obligatory adjunct and the last (108d) an optional one: (108)

a. Der Mann hatte den Tisch geschlagen. the(NOM) man had the(Acc) table hit T h e man had hit the table.' b. Der Spion schrieb (einen Brief). the(NOM) spy wrote a(Acc) letter T h e spy wrote (a letter).' c. Er wohnt in Frankreich in Paris in einem Vorort he(NOM) lives in France in Paris in a suburb in einer alten Villa . . . in an old villa 'He lives in France in Paris, in a suburb, in an old villa d. Der Spion traf ihn (in Berlin) (in einem Cafe). the(NOM) spy met him(ACc) in Berlin in a(DAT) cafe T h e spy met him (in Berlin), (in a cafe).'

The adjuncts are not distinguished by being optional, since some adjuncts are obligatory, and some arguments are optional. However, these adjuncts, unlike the arguments, are repeatable, as (108c,d) demonstrate. However, adding another NP in (108a) does not allow that second NP to be interpreted as another entity hit by the man. This will follow from the fact that argument linking requires an argument to be linked to, whereas adjuncts can be added freely as long as they are interpretable. So verbs

Why argument structure? 67 must have argument structures but also be able to syntactically require an adjunct: (109)

a. b. c. d.

schlagen ( x, y ) 'strike' schreiben ( x, (y) ) 'write' wohnen ( x ) NP 'live' treffen ( x, y ) 'meet'

The extra NP in (109c) will require at least one adjunct. The difference between this NP and the NP arguments is that additional (adjunct) NPs can be added in (108c) without ungrammaticality. Whether or not we have a level of argument structure, this difference must be encoded somewhere. Note that the present theory provides evidence for the argumentadjunct distinction and a diagnostic for distinguishing them. Consider the LDC and the sentences in (108). It is clear that the repeated adjuncts can be linked by the same preposition, e.g. in 'in.' If the LDC applied to these NPs, they would clearly violate it, since they are linked by the same preposition and no other linkers. But the LDC is explicitly a constraint on argument linking and so should not rule these out. Repeatability of adjuncts provides confirmation for the LDC as a constraint on argument linking. We can take this line of argumentation further, and ask what this means for argument categories. The LDC is not applicable to adjuncts, but it does seem to be sensitive to argument category features like EXP and GOAL, as evidenced by the dative-nominative sentence in (5b) above and the following ditransitive passive and its linking (for more on Icelandic ditransitives, see Chapter 5): (110)

Konunginum voru gefnar ambattir. king-the(DAT) were given(f.pl) slaves(NOM.f.pl) 'The king was given maidservants.' (Zaenen, Maling, and Thrainsson (1985:(44)))

(111)

gefa (x A

y|GOAL

z) 'give'

DAT: [ARG/GOAL]

NOM: [ARG]

NOM: [ARG]

_VP

[ARG]

ACC:

[ ]

*_VP: [ARG] [ ]

ACC:

Here the feature GOAL is the only thing preventing the LDC from ruling out nominative on the second argument, since it is all that distinguishes

68

Argument case and case alternations

the nominative from the most restrictive linker of that type (case) applicable to the higher argument. The LDC must not rule this structure out, because the sentence is good. However, if the LDC has to refer to an argument category, but it does not refer to adjuncts, then they evidently are argument categories and not of one piece with the adjuncts. To unify the goal of transfer and the locational goal in one semantic supercategory like the thematic role goal would miss an important aspect of explanation. It would not allow the LDC to be consistently formulated in terms of argument structure, which it needs to be. Another major difference between arguments and adjuncts that also follows from the constraints is their different mode of "blocking." The RC is an absolute blocking constraint, and only argument cases seem to be linked in accord with this constraint. Thus while goal dative must block default accusative, we get free variation for the adjuncts: (112)

a. The bird flew to the tree. b. The bird flew onto the tree.

The proposition that (112b) denotes entails the one denoted by (112a). The sentence in (112b) carries more information than the sentence in (112a), and in that sense the preposition onto in (112b) is more "specific" than the preposition to in (112a). But there is no question of onto "blocking" to. At most, if one knows that (112b) describes a state of affairs but uses (112a) one is violating the Gricean imperative of informativeness. The choice of onto is far from being obligatory, but the choice of dative over accusative in (110) is obligatory. The argument cases in DS competed with each other by the RC. But the adjuncts' LINK value I will take to be non-null, atomic, and different from the arguments' LINK values. This is accomplished by the feature ADJUNCT. Thus for (112), the following compete: (113)

Adverb (NP) to: [ADJUNCT], onto: ACC: [ ]

[ADJUNCT],

...

Either of the prepositions blocks the accusative, but neither blocks the other. Both sentences in (112) are generated by Phrase Structure. Both are interpretable. The use of the accusative is not admissible, because under any interpretation it might get, there is a more restrictive linker, to or onto, etc. The RC rules out accusative, but it does not decide between prepositions. The non-blocking of adjunct linkers follows from

Conclusion 69 their not having features like the argument categories. Once again, it is essential that the present theory treat argument categories as fundamentally distinct from the semantics, unlike in thematic role theory. 2.6

Conclusion

The restrictiveness-based theory of case captures the systematic similarities and differences between DS and NS without introducing unneeded devices into case theory. The solution to these particular problems was a by-product of a solution to the independent problem of generating case frames and predicting grammatical functions from argument structure. The uniform treatment of case, a novel feature of this theory, allows DS and NS to receive a unified account. In each alternation, the optionality of one case automatically leads to take-over by the "replacement" case. One key aspect of the account was the analysis of nominative as the anyargument case and accusative as the default case in Icelandic and some of its relatives. This will turn out to be a special case of a more general scheme of syntactic case which will be taken up in the next chapter.

3

A typology of case systems

This chapter will explore some consequences of the theory introduced in the previous chapter. There each case was associated with certain syntactic features, implemented there as the value of the feature LINK, which determine that case's syntactic behavior. In this chapter, I will extend this treatment of case to cover languages of different types and, in the next, to languages that make extensive use of linkers other than morphological cases. By "languages of different types" I mean languages whose cases show superficial differences from the system introduced in the last chapter, which had a nominative for arguments and an accusative as default. Some languages have a default nominative rather than the default accusative. And in some languages, the case of the object of a transitive verb is the same as the case of the subject of an intransitive verb, and a different case marks the subject of a transitive verb. The former case would be called "absolutive," the latter "ergative." This chapter extends the particular system of Chapter 2 to cover all these types. The main result of this chapter is that a generalization of the theory of the last chapter involving just two parameters correctly predicts four main types of two-case system. This four-way typology will lead to a range of predictions about the behavior of case in individual languages. In the next two sections, I will generalize the theory of linking and case frames already formulated and spell out the predictions it makes. Then I will present examples of each of the major predicted case systems and how the predictions fare in each of the languages considered. 3.1

Two accusatives

In the last chapter I argued that accusative in Icelandic and some related languages is the default in the domain of linking, rather than nominative as commonly assumed. The accusative could potentially link any NP 70

Two accusatives

71

whether argument or adjunct, and does so if there is no more restrictive linker available. This correctly predicts an Elsewhere pattern for linking by accusative in languages like Icelandic, and such languages will belong to one type (type 1) in the typology below. But some languages which we call "nominative-accusative" show the reverse pattern. Their nominative can appear on both arguments and adjuncts, whereas their accusative is highly limited. Nevertheless, they are nominative-accusative in the sense that the case which marks the subject of a transitive is the same as that which marks the subject of an intransitive - this is called "nominative" - and this case is distinct from the case which marks the object of a transitive, called "accusative". I will first generalize the system of Chapter 2 to handle this second type of nominative-accusative language. One of the most studied of nominative-accusative languages of this second, default-nominative, type is Japanese. In Japanese the accusative case marker -o marks the direct objects of transitive clauses: (1)

Taroo ga kodomo o sikatta. John NOM child ACC scold-PAST 'Taro scolded the child.'

In general, the Japanese accusative marks only arguments but never the highest argument. The evidence for this view will be presented at length in 3.3.2. The Japanese accusative will be associated with the feature specification [-H(ighest)A(rgument)], meaning "cannot appear on the highest argument": 1 (2)

ACC2: ^

The subscript 2 on accusative distinguishes the accusative here from the accusative in the last chapter and refers to type 2. Later I will show that an accusative like (2) is characteristic of type 2 languages in the typology of this chapter. The accusative entry in (2) makes Japanese -o like Icelandic nominative (both restricted to arguments) but with the addition of the specification [—HA]. Japanese nominative (-ga) will have the same LINK value as the accusative in Icelandic, i.e. it is the linking default: (3)

NOM2: [

]

The linking for (1) will be as follows:

72 (4)

A typology of case systems sikar- (X|+HA NOM2: [

y|-HA ) 'scold' ]

ACC2'.

\ARG 7

[-HAJ NOM2: [

]

The accusative is applicable to the second argument, because it is not the highest. It is not applicable to the highest argument precisely because it is highest. From now on, I will omit the redundant [HA] specifications from the argument structure, since it is immediately clear which arguments are highest and which are not. One immediate correct consequence of the use of [—HA] for the LINK value of the Japanese accusative is that it predicts that accusative will not be preserved in the passive. Wherever accusative and nominative compete in Japanese, the accusative should win, because it is the most restrictive. One desirable effect of [-HA] on the accusative is that accusative does not compete on the first argument of the passive. Thus, the linking for the passive in (5) is as in (6): (5)

Kodomo ga Taroo ni sikar-are-ta. child NOM Taroo by SCOW-PAST T h e child was scolded by Taro.'

(6)

sikar- (xA

y) 'scold' NOM: [ ]

In the passive the second argument becomes the highest, ruling out the use of accusative, which is prohibited on highest arguments. The passive correctly gets nominative. One question that immediately arises is whether this specification for [—HA] belongs in the same class as the features introduced thus far. Should the [—HA] specification, if present, be a part of the value of LINK or not? Certainly, since this feature does have a great deal to do with the linking properties of its case, the natural assumption is that it is part of the LINK value. This tentative assumption leads to interesting and correct consequences for the analysis of Japanese. Consider the situation that would arise with idiosyncratic case in a language like Japanese. The idiosyncratic case will either have the specification [-HA] or it will not (the only two logical possibilities). If the idiosyncratic case is not specified [—HA], then neither the idiosyncratic case nor the accusative is more restrictive than the other:

Two accusatives 73 (7)

a.

DAT: [ARG|2]

b. ACC2:

(part of the lexical entry of soodan suru 'consult,' etc.)

[ ARG —HA

Neither case's LINK value subsumes the other's, so blocking between the two cases is not defined, an undesirable state of affairs. When the logical object of an idiosyncratic dative verb needs to be linked, neither case will have priority. But the dative should have priority, since only dative (not accusative) is grammatical with a dative verb like soodan suru 'consult' (Kuno (1973), Marantz (1984: 78)): (8)

John ga Mary ni soodansita. John NOM Mary DAT consult-PAST 'John consulted Mary.'

To get this result, either [—HA] is not a part of LINK (the less natural assumption), or the other possibility above must be considered: the dative is [—HA].

If [-HA] is indeed part of the LINK value of the idiosyncratic case, then we get two predictions. First, the idiosyncratic case is correctly predicted to block the accusative, since its LINK value now does properly contain that of the accusative: (9)

a. DAT2:

ARG 12 (part of the lexical entry of soodan suru —HA J 'consult,' etc.)

b. ACC2:

ARG -HA

The dative blocks the accusative as in (10): (10)

soodan suru (x NOM2: f

y) 'consult' J

DAT2:

ARG/2 —HA

ACC 2 :

I ARG

I -HA NOM 2 : [

]

The accusative is correctly blocked despite its being potentially applicable by the AC to the second argument in (10). Only the dative is grammatical. If idiosyncratic datives generally have [-HA] in Japanese, then dative is more restrictive than accusative, and there is no need to revise

74 A typology of case systems the definition of restrictiveness or make [-HA] an exception to it. Nonnominative cases in Japanese generally need to bear [—HA]. The second prediction is that the idiosyncratic cases, like the accusative, should not be preserved in the passive, since, like accusative, they bear [—HA]. It turns out that idiosyncratic case behaves in exactly this way in Japanese (Kuno (1973), Marantz (1984)), as demonstrated by the passive corresponding to (8): (11)

Mary ga John ni soodans-(r) are-ta. Mary NOM John DAT/by consult-PASS-PAST 'Mary was consulted by John.'

In the passive, the second argument (Mary) is the highest, and the dative, being by hypothesis [—HA], is inapplicable. The accusative also is inapplicable, and so this leaves nominative, correctly predicting the active-passive pattern: (12)

soodan suru-PASS (xA

y) 'consult' NOM2: [

]

Nominative correctly applies in the passive of a verb with idiosyncratic case in the active. In this section I have extended the restrictiveness-based theory to a second type of nominative-accusative type of language. I will return to this type of nominative-accusative system in §3.3.2 after deriving the full typology, to which I now turn. 3.2

A typology of case

The extension of the theory to the second type of nominative-accusative language is the first of two steps on the way to a four-way typology. The hypothesis to be proposed is that there are two basic parameters at work in case theory and that, from these two parameters alone, four types of two-case system are predicted. The first parameter, the Limitation Parameter, is based on the distinction already introduced between the two types of nominative-accusative system. Some languages have a limitation like [—HA] on their any-argument linker (e.g. Japanese accusative), and some do not (e.g. Icelandic nominative). The second parameter, the Preference Parameter, is a choice of whether this feature refers to highest or lowest argument and simultaneously whether the LDC is a constraint on possible case for the higher or lower of two arguments.

A typology of case 75 Crossing these two parameters provides a four-way typology. In this section the two parameters will be formulated as will the predictions that follow from them. These will be tested against data from a variety of languages in §3.3.

3.2.1 The two-parameter hypothesis The hypothesis to be explored here is that two parameters derive a correct four-way typology. I start out from the assumption that each two-case system has an any-argument case and a default case. This yields the general system in (13): (13)

a. Case A: [ARG] b. Case B: [ ]

Each of the languages considered so far has two cases like these, although their labels differ. Icelandic's Case A is nominative, and its Case B is accusative. In Japanese Case A is accusative, and Case B is nominative. The difference between them consists so far in the extra limitation on Case A in Japanese. This first difference will be the basis of the first parameter, the Limitation Parameter. This parameter provides that some languages have a limitation on their any-argument case, their Case A: (14)

Limitation Parameter (preliminary): Languages can have the limitation [—HA] in the LINK value of their any-argument case.

The parameter yields the following possibilities: (15)

a. Case A: [ARG] , ~ A [ARG b. Case A: I—HA

Icelandic has (15a), Japanese (15b). The feature [—HA] can be present on the any-argument case (and lexical cases as in Japanese). I leave it open whether the default case might be limited to non-highest arguments. For typographical ease I will often abbreviate (15) further as in (16), but we have reason to reconsider the relation of [ARG] to [-HA] at several points below: (16)

a. Case A: [ARG] b. Case A: [ARG|— HA]

76 A typology of case systems Since

is all or part of the value of [ARG], its presence depends on [ARG]. I have already presented a little evidence for this parameter in the last section and will present more in §3.3.2 and §3.3.3. The second parameter, the Preference Parameter, rests on the hypothesis that the way [—HA] works in a language is closely related to the way the Linking Dissimilation Constraint (LDC) of the last chapter works in that language. Recall that the LDC is asymmetric. It considers the linking of pairs of arguments and asks for alternatives for only one of those arguments. In Icelandic, the LDC asks for alternatives on the lower of two arguments, and I repeat the version of the LDC which I gave in Chapter 2: (17)

[—HA]

Linking Dissimilation Constraint: Given two co-arguments, x and y, x higher than y, if y is linked by a linker B that has the same LINK value as the most restrictive linker of that type applicable by the AC to x, call it A, and there is a less restrictive linker C, applicable by the AC to y, then the linking is ill-formed.

In the notation I have been using, this is as follows: (18)

* verb

(...X

A: [a]

y • ••) B: [b] C: [c]

where A and B are of the same type, a = b, and C is applicable by the AC to y (c unifies with y's LINK value) Note that this constraint is asymmetric in that it "questions" the linking of the lower argument y but not the higher argument x. It is this asymmetry that gives the system its "directional" procedural interpretation. An alternative to (17-18) would be to have linking take the argument list top-down in Icelandic. Once the more restrictive linker (e.g. with LINK value [ARG]) is used on the first argument it can apply to, then subsequent (in the procedure) arguments cannot get this case without violating the prohibition on identical most restrictive linkers, which would still be necessary. So "early" use of the nominative on the higher argument precludes its "later" use on a lower argument. It has been noted in the literature that ergative-absolutive languages are in some sense the "reverse" of nominative-accusative languages. This view has found its clearest form in the Case Tier Hypothesis (Yip, Maling, and Jackendoff (1987)). This hypothesis holds that the difference

A typology of case 11 between nominative-accusative languages and ergative-absolutive languages consists in the direction of association of cases on a Case Tier and NPs on their tier. So if we have one or two NPs and two cases (19), we will get the nominative-accusative pattern if we associate cases with arguments rightwards (20), and we will get the ergative-absolutive pattern if we associate leftwards (21): (19) (20)

a.

Case

Case

NP

NP

NOM

ACC

NP b . NOM

(21)

ACC

I

NP

NP

a.

ACC

NOM

b.

ACC

NOM

NP

NP

NP

I

In leftward association, the NOM shows an absolutive pattern (subject of intransitive, object of transitive). The ACC shows an ergative pattern. In terms of the present approach to linking, under the procedural interpretation of the system, the Case Tier Hypothesis translates to the idea that Icelandic is top-down (or that the higher arguments are linked "earlier" than lower ones). Note that this would build directionality into the system. It would not mean that the cases would be on tiers. Specifically, what most versions of the Case Tier Hypothesis assume is that nominative and accusative are basically the same, and that the only difference between them results from directionality. This aspect of the hypothesis is incompatible with what is being developed here, since it is a central claim that the two syntactic cases should be distinguished by one being limited to arguments (the [ARG] case), and the other being the default. I propose to introduce a parameter into the system which has the effect of directionality of linking. Note that what leads to the directional interpretation is the asymmetry built into the LDC. It turns out that the two

78 A typology of case systems nominative-accusative types of languages differ in directionality of linking and that - correspondingly - there are two types of ergative-absolutive language. We can ask now, in the directional spirit of the Case Tier Hypothesis, what it would be like in each case if this asymmetry in the LDC were the reverse (or, equivalently, if lower arguments are linked "early" and higher arguments are linked "late"). If the LDC considered alternatives to the linking for the higher argument we would get the following: (22)

Linking Dissimilation Constraint: Given two co-arguments, x and y, y higher than x, if y is linked by a linker B that has the same LINK value as the most restrictive linker of that type applicable by the AC to x, and there is a less restrictive linker C, applicable by the AC to y, then the linking is disallowed.

I have placed emphasis on the change in the LDC for the "reverse." Graphically, we get the following: (23)

* verb (...y B: fb] C:[c]

...

x ...) A: [a]

where A and B are of the same type, a = b, and C is applicable to y (c unifies with y's LINK value) This reverse LDC is the same as the original one modulo the substitution of higher for lower in (17-18) above. The system which is exactly like Icelandic but has this reverse LDC is what I will call "type 4" below. An example of such a language is Kabardian, discussed in §3.3.4. The following sketch will make clear what is meant. Consider the case of a transitive in such a type 4 language. The argument structure and the linking will be as follows: (24)

verb ( x *Case A: [ARG] Case B: [ ]

y) Case A: [ARG] Case B: [ ]

The RC causes the most restrictive linker to apply consistent with the LDC. If the LDC is asymmetric in the "reverse" way, then the lower argument, y, is linked by the more restrictive Case A and the LDC has nothing to say, because the reverse LDC only looks for alternatives on the higher of a pair of arguments, which y cannot be. As for the other,

A typology of case 79

higher argument, its most restrictive linker consistent with the "reverse" LDC is Case B. The more restrictive Case A is out, because it would mean both arguments are identically linked where there is an alternative for the higher argument. Now consider an intransitive: (25)

verb ( x ) Case A: [ARGJ Case B: [ ]

The case that links a solitary argument is of course Case A. Now compare the intransitive (25) and the transitive (24). The same linker, Case A, links both the subject of the intransitive in (25) and the object of the transitive in (24) with Case B linking the subject of the transitive. Because of this distribution, Case A would be called "absolutive" and Case B "ergative." Hence such a language would be the ergative-absolutive reverse-wise counterpart to Icelandic. If this last type, our type 4 below, is the reverse of Icelandic (type 1), what is the reverse of Japanese, our type 2? Now that we have two instantiations of the LDC depending on whether the LDC considers alternative linkers for the lower (Icelandic) or higher (Kabardian) of two co-arguments, we can ask which is appropriate for Japanese. Looking back at the demonstrations of the last section, this is not so clear. Since the accusative in Japanese is limited to non-highest arguments, the LDC is not crucial in the linking of transitives (but will be in ditransitives). However, as we will see in Japanese causative constructions, the LDC should consider alternatives for higher arguments, like Kabardian, but unlike Icelandic. Note that, if so, this is the second place in the Japanese case system where the notion "high" is used. It was and will be argued that Japanese has the restriction [—HA] "not allowed on highest arguments" on its [ARG] linker. The second parameter, the Preference Parameter, consists in the claim that these two places where the notion "high" is used are related and can likewise be reversed. So the answer to the question of what the reverse of Japanese is would be a language with the opposite value, i.e. "low" in both places. It would have [-LA] "not allowed on lowest arguments" on its any-argument linker, and its LDC would consider alternatives for lower arguments, just as in Icelandic. Such languages will be called type 3, an example of which is Warlpiri, discussed in §3.3.3. The situation in type 3 is sketched schematically in the following for transitives and intransitives:

80 (26)

(27)

A typology of case systems verb ( x Case A: [ARG/-LA] Case B: [ ] verb ( x ) Case B: [

y) Case B: [

]

]

Case A shows an ergative pattern, Case B an absolutive one. But the cases themselves are different from those in the previous type of ergativeabsolutive language. Here what is called "absolutive" is the default case, whereas in type 4, the default shows the ergative pattern. We get two types of ergative-absolutive language - exactly parallel to the two types of nominative-accusative language. To keep the LDC general, we should factor out the high versus low aspect and leave the rest as part of Universal Grammar: (28)

Linking Dissimilation Constraint: Given two co-arguments, x and y, y high- /low-er than x, if y is linked by a linker B that has the same LINK value as the most restrictive linker of that type applicable by the AC to x, and there is a less restrictive linker C, applicable by the AC to y, then the linking is disallowed.

The choice of low versus high will be set by individual languages. And, similarly, the feature [ - H A / - L A ] must be set by the same parameter. We can say that [-HA] and [-LA] are the same feature in Japanese and Warlpiri but it must be "set" as to whether it prohibits the [ARG] case from linking highest or lowest arguments. So one way to do this is have a feature like the following universally with the choices:2 (29)

[-XA] where x = H if the feature means "prohibited on highest arguments" and x = L if the feature means "prohibited on lowest arguments"

To reflect this generalization of the feature [-HA] to [-XA] with H filled in as only one of two possible values of a parameter, we need the following revised Limitation Parameter: (30)

Limitation Parameter (revised): Languages can have the limitation [—XA] in the LINK value of their any argument case.

The Preference Parameter simply chooses a value for H(igh) versus L(OW) in (28) and (29):

A typology of case 81 (31)

Preference Parameter: Choose one of {high, low} in (28) and (29).

The claim behind the Preference Parameter is that the setting high versus low in the LDC and the argument case limitation feature is one setting. Crossing the two parameters, the Limitation Parameter and the Preference Parameter, gives a four-way typology. I will call Icelandic type 1. This type has no limitation on the [ARG] case and has the value low for the Preference Parameter which induces top-down directionality of linking under a procedural interpretation: (32)

Type 1: Limitation Parameter: No Preference Parameter: Low (Top-Down)

Type 2 differs in both respects: (33)

Type 2: Limitation Parameter: Yes Preference Parameter: High (Bottom-Up)

Type 3 is the reverse of type 2 with respect to the Preference Parameter, so it shares with type 1 the low value of the Preference Parameter: (34)

Type 3: Limitation Parameter: Yes Preference Parameter: Low (Top-Down)

Finally, the type which is the reverse of type 3 in both respects is type 4: (35)

Type 4: Limitation Parameter: No Preference Parameter: High (Bottom-Up)

This last type is thus like type 1 except opposite with respect to the Preference Parameter. Note that the theory as developed in this section provides a crosslinguistic way of defining cases in a syntactic sense. The syntactic behavior of any given linker is determined by the three constraints of Chapter 2, as modified in this chapter. Each of these constraints makes reference to the case not by its name but by its LINK value. Hence the syntactic behavior of any linker is determined by its LINK value. But note now that this gives us an elegant way to identify cases cross-linguistically. In fact,

82

A typology of case systems

this identification is possible precisely because the LINK value of a linker defines it syntactically. The equivalences are given in the following table: (36)

type of information

defined linker

[

ACCi, NOM2, ABS3, ERG4

]

[ARG]

NOM1? ACC2, ERG3, ABS4

[ARG|X], X = EXP, GOAL

argument category case (e.g. dative) idiosyncratic case

Y], Y = 1, 2 in lexical entry

[ARG|...

Argument category cases can be identified across languages to the extent that the feature GOAL in one language can be identified with the feature GOAL in another language. At any rate, argument category case and idiosyncratic case receive cross-linguistic definitions based on the type of LINK information they contain, i.e. information besides [ARG] (and [-XA]). More importantly for the typology of two case systems, the two syntactic linkers are defined by the feature [ARG] and its absence. One syntactic case has a null LINK value, and the other has the information that it is limited to arguments. One of the most noteworthy results that emerges from this typology is that the labels for cases are only labels. They do not define cases. The nominative-accusative and ergative-absolutive patterns each come in two radically different varieties, if this scheme is correct. 3.2.2 Predictions These new, more general definitions of cases (and other linkers) allow us to generalize the predictions about case in Chapter 2 about type 1 languages up to their full cross-linguistic generality. Any prediction about nominative in Chapter 2 we can now see was really a prediction about the any-argument case and so is a prediction about type 2 accusative, type 3 ergative, and type 4 absolutive. Similarly, predictions about the accusative in Chapter 2, especially, the prediction that it will show Elsewherelike behavior, is just a special case of a wider prediction about default cases, so about type 2 nominative, type 3 absolutive, and type 4 ergative as well. Before turning to more languages, I summarize these predictions in their most general form. First of all, an [ARG] linker no matter what label it receives is predicted to show LDC effects. If two co-arguments have an [ARG] linker as their most restrictive linker, that linking will be ruled out. This is because there

A typology of case 83 is always the less restrictive default case as an alternative for one of the arguments. Thus we get the following prediction, which is the LDC applied to the [ARG] linker: (37)

Argument Linker Dissimilation Prediction: A linker with the LINK value [ARG] may not link two co-arguments.

This prediction of course follows directly from the LDC and the nature of the any-argument linker. It also follows from the LDC that one of two co-arguments will get the default case, since the LDC only rules out a linking if there is an alternative for a certain one of the co-arguments. So any time in the absence of the LDC we would get two co-arguments linked by the [ARG] linker (or two [ARG] linkers), then the LDC will force the use of the default. So we get the following corollary: (38)

Corollary: If the [ARG] case is the most restrictive linker that passes the Applicability Constraint for two arguments, one of those arguments will show default case.

In Chapter 2, we saw that this means that in a type 1 language, anywhere the nominative is the most restrictive case that passes the AC for two coarguments, that clause will contain an accusative.3 Since type 1 languages have the low-value for the Preference Parameter the accusative will appear on the lower of any such pair of co-arguments. A closely related prediction concerns the repeatability of the default case. The default case has a special status with respect to the LDC, because the LDC only rules out linkings where there is an alternative for a designated member of a pair of similarly linked co-arguments. If two co-arguments are linked by the default case, there can of course be no alternative case, since there is no less restrictive case than the default case, whose LINK value subsumes any LINK value. From this it follows that the LDC will never rule out the multiple use of default cases, regardless of what the language labels it (accusative, nominative, absolutive, or ergative), and we get the following: (39)

Default Linker Non-Dissimilation Prediction: Default (LINK: [ ]) linkers never show LDC effects and so are repeatable.

In fact, it is easy to see that the default case is the only case with this property. For any non-default case, there is always an alternative that will cause the LDC to rule out the linking. So, for nominative, if it is the most restrictive linker linking two co-arguments, the fact that the default

84 A typology of case systems is an alternative will cause that linking to be blocked. This is not just true for nominative or [ARG] linkers generally - as emphasized in the Argument Linker Dissimilation Prediction - but rather is true for any non-default linker. The default is always a less restrictive alternative in all contexts except where it is the only case that passes the AC. There is no less restrictive alternative to the default. From this it follows that all linkers except for the default linker will show LDC effects: (40)

Corollary: Non-default linkers show LDC effects, and so they are not repeat able.

This means that if a linker does not dissimilate we are forced to analyze it as a default. This will further mean that it cannot "block" any other linker by the Restrictiveness Constraint. An analysis of a case as the default makes predictions in all three areas, distribution, blocking, and dissimilation which must all hold true. The last prediction also relates to the default case, and has a somewhat different status than the others. Recall from Chapter 2 that the way a default status for accusative in Icelandic was justified was by showing that it, but not nominative, could mark adjuncts such as time adverbials. From this we can derive a predicted tendency: (41)

Elsewhere Pattern Prediction: Only default case will have an Elsewhere distribution, the distribution that reflects the complement of the conditions on the other linkers combined, and is likely to mark adverbials and other adjuncts.

The second half of the prediction must read "likely," because it is well known that adverbials and other adjuncts may be marked with a case otherwise used for arguments. It is a matter of great controversy whether a dative of time-at-which should be introduced by the same rule as a dative of "indirect object." As I indicated at the end of the last chapter, I am advocating that such datives be kept syntactically distinct. But once we allow for adjunct cases that happen to use the same morpheme as a syntactic case, we admit of the possibility of apparent counterexamples to the claim that there is one default linker and that it alone will be likely to mark a range of adjuncts. We can ask what empirical content the Elsewhere Pattern Prediction has. There are two specific ways in which (41) has a great deal of empirical content. The first relates to the markedness of different options for adverbials. The default in a language is the linker that applies just in case no

A typology of case 85 other linker can. Nothing requires that the default be used. The extent of its applicability is the complement of the extent of applicability of all the other linkers combined. This means that if a language has a specific case or adposition for every adverbial and for various argument types (especially GOAL), it may be true that the default happens never to be allowed to be used more than once per clause. For this reason, we cannot make the stronger prediction that default linker must be used more than once in some sentence of each language. The default of course will be used in the absence of more restrictive linkers that pass the AC and the LDC, but only there. Now, consider an adverbial that has no special case. It costs nothing at all for that adverbial to be linked by the default linker. On the other hand, if the morphological dative or the morphological genitive marks many adverbials, this is just coincidence or historical residue and so should be less likely to happen than a wide range of adverbials bearing default case. Within a given language, certain facts like a uniquely wide and disparate set of uses for one case calls for an analysis in terms of default. Recall in this regard Gaedicke's (1880) justification for treating accusative as default in his grammar of Vedic Sanskrit, discussed in Chapter 2. Thus it is a cross-linguistic prediction that the default case is more likely on a range of adverbials than any other case. This is what is meant in (41). If more languages have some other morphological case marking adverbials, then this would be a major problem for the theory. The second way the theory confronts the data in (41) is that the default case is expected to show Elsewhere behavior in the realm of argument linking. This means that if some case demands to be treated as the Elsewhere since its distribution is so heterogeneous, then this case must have the other attributes of the default, i.e. repeatability, least priority, etc. This brings us to the last and most important point of all. The theory here in conjunction with analyses of particular languages not only makes the predictions above and many others about which case blocks which, but it makes a tightly interconnected set of predictions. Once a case is analyzed as the [ARG] case on the basis of its distribution, it must obey the Argument Linker Dissimilation Prediction. In other words, if a case is the [ARG] case with respect to the AC and the RC, it must be the same case with respect to the LDC. The same goes for the default. Each linker must behave consistently with respect to all three constraints. If a case is repeatable but blocks another case, then this is a clear counterexample

86 A typology of case systems to the theory being developed. The theory makes the predictions for a given case simultaneously and for a group of cases. The theory, in conjunction with the analysis of a particular language's cases in terms of LINK values, is meant to predict every possible case frame in a language and rule out all the impossible ones. Thus each analysis bears on the theory in the strongest way.

3.3

Four types of case system

This section will examine the preliminary evidence for this typology. The first and most important prediction inherent in the typology is that each of the four types should exist. This section will present evidence that each type is indeed attested. For each type, I will examine several languages to test the predictions of the last section. I will also indicate other languages that may fall into each of these types. 3.3.1 Type 1 In Chapter 2 we saw our first example of a type 1 language. The account of Icelandic and related languages there can now be seen as a special case of the four-way typology. I will return to Icelandic in greater detail in Chapter 5. This section examines data from another type 1 language, Classical Greek, and tests the predictions that follow from calling Greek type 1. A type 1 language has no limitation on its argument case and has the value low for the Preference Parameter: (42)

Type 1: Limitation Parameter: No Preference Parameter: Low (Top-Down)

The no value of the Limitation Parameter means that the any-argument case (Case A) will not be prohibited on the highest or lowest argument. The low value of the Preference Parameter means that the LDC will question the linking of the lower of each pair of co-arguments. The result of this is that Case A will have a nominative distribution and Case B an accusative one, with nominative as the any-argument case and accusative as the default: (43)

a. NOMI: [ARG] b. ACCI: [

]

Four types of case system

87

So the predictions of §3.2.2 that referred to the any-argument case are predictions about nominative in Classical Greek (just as they were in Icelandic, another type 1 language), and predictions about the default case are, for Greek, predictions about accusative (once again like Icelandic). I will now confirm that the predictions are borne out in Greek. The first prediction for the argument case is the Argument Linker Dissimilation Prediction, a special case of the LDC. Linking twice by the [ARG] linker will inevitably violate the LDC. For Greek this means that any clause containing more than one argument of the same verb will contain at most one nominative. Further, by the corollary in (38) applied to Greek, we expect that if nominative is the most restrictive linker that passes the AC on two arguments, then the lower will show accusative. Greek seems to be like Icelandic and its relatives considered in Chapter 2 in not allowing two nominatives per clause except in predication. Although this is based on negative evidence from a dead language, the absence of such double nominative sentences is striking, when compared to Japanese (next section). As for the corollary, it is indeed true that the accusative is used when the nominative is the most restrictive linker for more than one argument. The simplest situation is an intransitive (44) which I present with a simple transitive (45) for comparison: (44)

(45)

Proxenos paren Proxenos(NOM) was-present Troxenos was present.'

(X.A.I.2.3)

proseilephe ten ton ponerOn koin^n received the(ACc) the(GEN) vile(GEN) common(ACc) eponumian sukophantes appelation(ACc) informer 'he received the common appelation of the vile, "informer"' (Aeschines 2.99)

First of all, the nominative-accusative pattern is obvious. Where both nominative and accusative may apply, and the LDC is irrelevant, i.e. in the intransitive and for the highest argument of the transitive, the nominative must apply by the RC, and in fact it does. On the lower argument of the transitive, use of the nominative violates the LDC. The only wellformed structure has the less restrictive accusative, fulfilling the prediction.

88

A typology of case systems

If accusative is the default linker, it must not dissimilate, according to the Default Linker Non-Dissimilation Prediction. I know of no sentences where we would wish to have accusative dissimilation. This prediction also says that accusative is in principle repeatable. As noted above, if a language is so rich in more restrictive linkers that the default has little chance to apply by the RC, then it may turn out that the default linker is never actually repeatable. The fact that the default cannot repeat in that language would be an accidental gap. However, cross-linguistically, the default should be likely to repeat. And in each language, if any case can be used for more than one argument per clause, it is the default, which in Greek is accusative. Greek actually shows a wide range of multiple accusatives. First, two arguments per clause may take accusative. In Greek, ditransitive verbs take dative for the goal and accusative for the last argument, but a number of verbs take two accusatives: (46)

Kuros didosin autOi hex men6n misthon Cyrus(NOM) gives him(DAT) six months pay(Acc) 'Cyrus gives him pay for six months.' (X.^4.1.1.10)

(47)

oudeis edidaxe me tauten ten technen nobody(NOM) taught me(ACc) this(ACc) the (ACC) skill(ACc) 'Nobody taught me this skill.' (X.0.319.16)

This can be handled in a fashion very similar to the nominative versus dative experiencer classes in Icelandic (Chapter 2). The second argument of both types of verb are semantic goals, i.e. they correspond to participants who are entailed to have something move to them. The semantic goal of the first class of ditransitives is a syntactic goal as well. The second class will have three direct arguments, one of which is a semantic but not syntactic goal: (48)

didomi ( x

y|DAT z) 'give'

(49)

didasko ( x

y

z) 'teach'

The dative of (syntactic) goals is not applicable to the second argument of the second class. The most restrictive case that passes the AC for that argument is the any-argument case, in Greek the nominative, as shown in (50):

Four types of case system (50)

didasko (x

y

89

z) 'teach'

NOM

*NOM

*NOM

ACC

ACC

ACC

The nominative, although it is the most restrictive case that passes the AC for the second argument, is not the most restrictive applicable case, because its use would violate the LDC. Greek, being type 1 with the low value for the Preference Parameter, questions the linking of lower arguments versus higher arguments by the LDC. The nominative is the most restrictive applicable case that passes the AC for the highest argument. Therefore, by the LDC, use of the nominative on the second (or any lower argument) would violate the LDC's prohibition on identity. So accusative is the only available option, as it is for the last argument as well. The other possible analysis is that the semantic goal in (47) takes a lexical accusative, although Greek has no other lexical accusatives. On the view I am taking, that the accusative is not lexical, one predicts that this accusative will not be preserved in the passive. The goal argument does passivize, and the other argument remains accusative: (51)

ouden allo didasketai anthropos e epistemen nothing(Acc) else teach-PASs man(NOM) than knowledge(ACc) 'Man is taught nothing else except knowledge.' (P.Men.Slc)

This receives the following analysis: (52)

didasko (xA

y

z) 'teach'

NOM

*NOM

ACC

ACC

The "retained" object must get accusative by the LDC. Note that the accusative in these ditransitives shows the distribution we predict, because it applies exactly when no other case is applicable. The ditransitives are also the first evidence for the other facet of the Elsewhere Pattern Prediction: that the default is likely to mark a heterogeneous class of NPs. Already we have seen that the accusative links the logical object of most verbs, the goal of a class of ditransitives and the "retained" object in the ditransitive passives. In addition, Greek shows a wide range of non-argument accusatives (Smyth (1956: 355-63)), which are expected on the Elsewhere Pattern Prediction. In Greek, accusatives of path are common and quite free:

90 (53)

A typology of case systems hopos agein (stratian) e stenas e how to-lead army(ACc) either narrow(ACc) or plateias hodous broad(Acc) roads(Acc) 'how to lead an army over either narrow or broad roads' (Xen.Cyrop. 1.6.43: 10-11)

Another non-argument accusative in Greek is the accusative of extent. Sentence (54a) exemplifies the accusative of extent of time, and (54b) the accusative of extent of space: (54)

a. entautha emeinen hemeras hepta. there remained days(Acc) seven 'There he remained seven days.' (Xen.Anab. 1.2.6: 28) b. apekhei de he Plataia t6n Thebdn be-distant PTCL the Plataea the(GEN) Thebes(GEN) stadious hebdomekonta stades(Acc) seventy 'Plataea is seventy stades distant from Thebes.' (Thuc.2.5.2:

These adverbial accusatives need no special "semantic accusative" or separate rules but follow from the ultimate default nature of type 1 accusative. The linking for (54a) is as follows: (55)

meno (x) 'remain' Adverb (NP) NOM

ACC

ACC

Additionally, the adverb will contribute to the interpretation, but this is not represented at argument structure. A similar analysis can be given to "cognate accusatives": (56)

a. pollen phluarian phluarounta much(Acc) nonsense(Acc) nonsense-talking-PART(ACc) 'talking much nonsense' (P.A. 19c) b. esthenese tauten ten noson fell-ill(3sg) this(ACc) the(ACc) disease(ACc) 'He fell ill of this disease.' (Isoc. 19.24)

The linking for these is essentially the same as in (55). The particular significance of the cognate accusative is that as a non-argument that

Four types of case system

91

normally has no special case, it is one of the best diagnostics for which linker is the ultimate default linker. The "free" use of the accusative in Greek extends to any loose adverbial connection to the rest of the clause. Such accusatives are traditionally known as "accusatives of respect": (57)

deinoi de terrible(NOM.pl) PTCL 'terrible in battle'

makhen battle(ACc) (AQSch.Pers.21)

For other non-argument accusatives in Classical Greek, see Smyth (1920: 355-63). The general characteristic of these constructions is that they contribute semantically like an adverbial, they do not passivize except under special circumstances, and they are compatible with another accusative in the sentence.4 They are as freely addable to clauses as adverbials of the same meaning. By having the accusative as in (43b), we can get all of these uses of the accusative without special rules. However, anytime there is a preposition to mark an adjunct, the preposition is used if it is obligatory. The accusative is the default and so will pick up any NPs that are not marked by more restrictive prepositions or cases. Accusative thus marks a range of non-argument NPs. Nominative, however, never marks adjuncts but rather is limited to arguments. Assuming the general linkers proposed and that Greek is type 1 gets us a range of predictions which are fulfilled, i.e. the universal network of predictions from §3.2.2. In the next chapter I will show that English is type 1, and in Chapter 5 I provide a more detailed analysis of another type 1 language, Icelandic.

3.3.2 Type 2 As the representative of the other main type of nominative-accusative system, type 2, I take Japanese. Section 3.1 showed that there is basic motivation for thinking Japanese accusative to be like type 1 (Icelandic, Classical Greek, etc.) nominative with the addition of a limitation, [—HA]. Further, the assumption of the high value for the Preference Parameter would ensure that this [ARG] accusative would appear as far down in the argument hierarchy as possible, leading to a basically nominative-accusative pattern, which we find. I summarize the parameter values in (58):

92 (58)

A typology of case systems Type 2: Limitation Parameter: Yes Preference Parameter: High (Bottom-Up)

This section will test the various predictions for Japanese and will show that a wide variety of phenomena previously thought to require a special account follow directly from the universal principles and the analysis of Japanese as type 2. 3.3.2.1 The Double -o Constraint In this section I will show that the Argument Linker Dissimilation Prediction, in conjunction with our analysis of accusative in Japanese in §3.1, predicts exactly the behavior that has been said to require a "Double -o Constraint" in Japanese. In other words, the behavior of Japanese, far from requiring any special constraint, will follow from the universal LDC. In Japanese the accusative case marker -o marks the direct objects of transitive clauses: (59)

John ga hon o yondeiru. John NOM book ACC read-PROG-PRES 'John is reading a book.'

One fact about Japanese is that it has two double -o constraints (Harada (1973), Poser (1991)). One is a stylistic surface constraint against two -o marked NPs which varies from speaker to speaker. The other, the deep Double -o Constraint, disallows two NPs that would receive -o even if some other process causes one of the -os not to appear on the surface. In the causative of an intransitive verb, the causee can be marked by the accusative marker -o or by the dative/agentive marker -ni with a slight meaning difference: (60)

a. John ga Mary ni arukaseta. John NOM Mary DAT walk-cAus-PAST 'John let Mary walk.' b. John ga Mary o arukaseta. John NOM Mary ACC walk-CAus-PAST 'John made Mary walk.'

If the causative is of a transitive verb, only the dative/agentive marker is possible and has either the 'let' or 'make' interpretation:

Four types of case system (61)

93

a. *John ga Mary o hon o yomaseta. John NOM Mary ACC book ACC read-CAus-PAST 'John made Mary read a book.' b. John ga Mary ni hon o yomaseta. John NOM Mary DAT book ACC read-CAus-PAST 'John let/made Mary read a book.'

The Double -o Constraint is meant to capture the fact that (61a) is ungrammatical. Poser (1991) argues that the deep Double -o Constraint is not a constraint on surface case, since neither movement nor processes which cause the NP not to receive -o on the surface can affect the ungrammaticality of sentences like (61a). Topicalization normally replaces nominative and accusative on the topicalized NP with the topic marker -wa, but Poser (1991) shows that it cannot improve (61a) despite the surface non-appearance of -o. First, a simple transitive like (62) can have either John or the book NP topicalized: (62)

a. John wa hon o yondeiru. John TOP book ACC read-PROG-PRES 'As for John, he is reading a book.' b. Hon wa John ga yondeiru. book TOP John NOM read-PROG-PRES 'As for the book, John is reading it.'

In the causative, the replacement of -o by -wa does not improve (61a): (63)

a. *Hon wa John ga Mary o yomaseta. book TOP John NOM Mary ACC read-CAus-PAST 'As for the book, John made Mary read it.' b. Hon wa John ga Mary ni yomaseta. book TOP John NOM Mary DAT read-CAus-PAST 'As for the book, John made Mary read it.'

The Double -o Constraint cannot be a constraint on surface case, since such a constraint would not rule out (63a), which one expects to be bad for the same reason as (61a). Another of Poser's (1991) arguments for the non-surface nature of the Double -o Constraint comes from relativization. Here the relativized NP receives no surface case marker at all. Nonetheless, relativization does not improve (61a) either:

94 A typology of case systems (64)

a. *John ga Mary o yomaseta hon John NOM Mary ACC read-CAUs-PAST book 'the book which John made Mary read' b. John ga Mary ni yomaseta hon John NOM Mary DAT read-CAus-PAST book 'the book which John made Mary read'

Once again we want the constraint that rules out (61a) to rule out (64a) as well, but a surface constraint on morphological accusative case markers will not do the job. The same holds true of the phenomenon of Case Marker Drop. Even if the accusative case marker is "dropped" (or at least does not appear on the surface), the sentence is still unacceptable: (65)

a. *John ga Mary o hon yomaseta. John NOM Mary ACC book read-CAus-PAST 'John made Mary read a book.' b. John ga Mary ni hon yomaseta. John NOM Mary DAT book read-CAus-PAST 'John let/made Mary read a book.'

If some constraint is to rule out (65a) and (61a), it cannot be a surface constraint. The Double -o Constraint is a special case of the LDC as applied to argument linkers. In the ungrammatical sentence in (61a), co-arguments are linked by the non-null, non-default accusative, which the LDC rules out:5 (66)

yomaseta ( x NOM

y -ni,

z ) 'caused to read' ACC

NOM

ACC NOM

In the linking in (66) the dative/agentive -ni must be used. Use of -o (whether dropped or not) would lead to a LDC violation. For the other ungrammatical sentences, all that need be said is that the NP that would surface as -o is linked by the same case but that the case is not spelled out as -o. But this also follows from the general condition that all arguments and NPs must be linked. For instance, the realization of the sequence -o-wa (accusative and topic marker) is just -wa. Another alternative would be to define the LDC on surface form. Such a constraint would be as follows:

Four types of case system 95 (67)

Alternative LDC: Do not link co-arguments with the same nondefault linker.

This alternative LDC would have many of the same desirable effects as the one I am proposing. For instance, two nominatives per clause would be ruled out for Icelandic. This alternative solution runs into problems with Japanese, which, as Saito (1983) observes, can have -o or adjacency to V (_V) as its "accusative" linker. We might capture Saito's observations by proposing the following linkers: (68)

a. -o: [ARG|-HA] b. _V: [ARG|— HA]

The LDC as presently formulated says that, all else being equal, two argument NPs of the same predicate linked by -o and _ V would dissimilate. Under the alternative LDC, however, two co-arguments linked only by (68a) and (68b) should not dissimilate, because they are different linkers, and the alternative LDC is defined over linkers not LINK values. It turns out that there is dissimilation and that our LDC can explain some of the scrambling behavior of Japanese, if we assume the additional linker in (68b). In Japanese, objects may appear with (69a,c) or without (69b) the accusative case marker, but, if they appear without it, they must be adjacent to the verb ((69b) versus (69d)) (Saito (1983)): (69)

a. Masao ga hon o katta. Masao NOM book ACC bought 'Masao bought a book.' b. Masao ga hon katta. Masao NOM book bought c. Hon o Masao ga katta. book ACC Masao NOM bought d. *Hon Masao ga katta. book Masao NOM bought

In (69d) Hon is not linked and so violates the Linking Condition that requires all NPs to be linked. Now note that the LDC gets the following correctly ruled out: (70)

*John ga Mary hon o yomaseta. John NOM Mary book ACC read-CAus-PAST 'John made Mary read a book.'

96

A typology of case systems

This sentence violates the double accusative constraint, even though one could say that Mary is linked by -o and hon is linked by _V. Under the alternative LDC, (70) is not ruled out, since -o and _V count as separate syntactic cases because of their form. One way out of this problem is to follow Kiparsky (1987) and have -o or _V as the alternate realizations of one linker. Then the alternative LDC will rule out use of -o and _V for two arguments of the same verb, because those two arguments would be linked by the same linker, which the alternative LDC rules out. The question then arises how to decide what are separate linkers (_VP and nominative in Icelandic) and what are not (-o and _V in Japanese). The alternate realization hypothesis is most suspicious when there is no morphological relation between the two realizations/linkers, e.g. an accusative postposition versus a word-order position.6 This problem is very reminiscent of the problem posed by the case alternations of Chapter 2. There it was noted that the power to relate a case to another case rather than just to its absence carries with it undesirable power. So too here, the alternate realization hypothesis carries the same potential to relate any two cases. This is not necessary under the present account. Linkers of different types (case, agreement, word order) are separate linkers but are syntactically identical and behave as identical for the LDC as formulated above. The result is that such syntactically identical linkers may be used for different arguments as long as some other linker (e.g. dative) provides an additional distinguishing feature, as with the "dative subject-nominative object constructions." This approach allows the tightest connection between (morphological) case and (syntactic, abstract) Case. If, however, a case and a word-order position can be alternate realizations of the same abstract linker, then the richness of the realization component makes the distance between case and Case larger. On the present approach case and Case are as close as possible, constraining heavily the notion of Case. On this account, the ungrammatical sentence in (65a) above actually might have two derivations, but the LDC correctly rules them both out. One involves case marker drop as mentioned above. Case marker drop does not affect linking and so is irrelevant for the LDC. The other derivation of (65a) would involve using both -o and adjacency to the verb (_V) with the accusative on the causee and adjacency to V on the lowest argument:

Four types of case system (71)

*yomaseta ( x NOM

y -ni, -o

z ) 'caused to read' _V

NOM

NOM

97

The LDC disallows this structure since the linking features for the y and z arguments are the same. Syntactically the two linkers, accusative case and adjacency to V, are the same and so show LDC effects. Further confirmation for the theory comes from dative object verbs discussed in Poser (1991). In the causative of a dative object verb, the causee may be accusative: (72)

John ga Mary o Bill ni kataseta. John NOM Mary ACC Bill DAT defeat-CAUs-PAST 'John made Mary defeat Bill.'

According to the theory, the following cases are involved: (73)

a. -ni:

(part of lexical entry of katu 'defeat')

[ARG|2|-HA]

b. -o: [ARG]—HA]

The case in the lexical entry (73a) blocks (73b) since the LINK value of the dative properly includes that of the accusative. Thus Bill gets dative, and the accusative is still available for the causee. The dative with katu 'defeat' and the accusative do not have identical LINK values. The LDC will allow the causee to receive accusative: (74)

kataseta ( x NOM

y

z ) 'caused to defeat'

-ni, -o

-ni: [ARG|2|-HA]

NOM

-o, _V NOM

Note that the dative for the lowest argument comes from the lexical entry. There are no identical Unkings in (74). The LDC is satisfied, and the sentence is correctly generated. 3.3.2.2 Nominative If, as our analysis of Japanese requires, nominative is the default case, then it falls under the Default Linker Non-Dissimilation Prediction. The nominative should not show LDC effects and so should be repeatable in principle. Further, by the corollary in (40), the fact that nominative is repeatable will mean that it cannot be analyzed as anything but the default. The position that nominative is the default case in Japanese is

98

A typology of case systems

not new, although it should be kept in mind that nominative is taken to be the default in some type 1 languages, a position that did not square with the notion of default in the present theory. I will now show in this section and the next that the nominative in Japanese is the default in the sense required by this theory. It passes both predictions for the default case. First, as predicted the nominative is repeatable. Double nominative sentences occur.7 The first class of double nominative sentences are with verbs that optionally do not take a morphological accusative. In (75) the nominative can mark both arguments of desideratives in -tai and potential forms in -re I are (Kuno (1973: 93)): (75)

Dare ga nihongo ga/o hanaseru ka who NOM Japanese NOM/ACC speak-can Q 'Who can speak Japanese?'

Kuno (1973) proposes the following generalization with exceptions: stative verbs may take nominative objects. Poser (1991) demonstrates that this is a fact about surface -o since the statives may take nominative in place of an accusative of path (about which more shortly): (76)

a. Boku wa Asahidake ga I TOP Asahi-peak NOM 'I want to climb Mt. Asahi.' b. Konban hamabe ga this-evening beach NOM This evening I want to take

noboritai. climb-want sanpoositai. take-a-walk-want a walk along the beach.'

In the present approach the generalization can be reconstructed as a systematic exception to any case of the form -o: (77)

[ + stative] -> [- -o]

(optional)

At first this might seem to be evidence against syntactically separate accusatives, but the exceptions to (77) documented by Kuno are stylistic, suggesting this is not a regular linking condition but a condition on form. Alternatively, (77) could be replaced by a spell-out rule that makes -o into -ga. But this involves an extra degree of power that we tried to avoid in Chapter 2. The theory here can do better than having a rewrite rule. The fact that the NP that might have received -o receives -ga rather than some other case follows automatically from the theory. The case -ga is the next most restrictive applicable case. The only unusual thing about the

Four types of case system

99

situation with the statives in Japanese is that a case (actually a group of cases) is prohibited by verbs on a morphological basis. The stative restriction picks out this class on the basis of form as in (77). But the non-application of these phonologically related cases leads to the application of the default nominative, as required by the theory. 3.3.2.3 Adverbial case If, as I have argued, Japanese has accusative as the more restrictive case (and nominative as the default), one should expect few of Jakobson's (1936) "weakly governed" accusatives in Japanese. Rather, one should, by the Elsewhere Pattern Prediction, expect there to be a wide range of nominative adverbials. In fact, both expectations are borne out. Japanese does not use -o to mark time adverbials or cognate objects. Japanese does sometimes mark paths with -o: (78)

a. Kawa o kare ga sono nimotu o watasita. river ACC he NOM that baggage ACC cross-PAST 'He moved that baggage across the river.' b. Sono miti o Isao ga uma o hasiraseta. that road ACC Isao NOM horse ACC run-CAus-PAST 'Isao made the horse run along the road.'

This path accusative is compatible with other accusatives, so it cannot be the same accusative on the present approach. One possibility would be for Japanese to have a default accusative, but we would need some special constraint to block this other accusative from applying in many places. This would be cumbersome. But the accusative of path in Japanese does not seem to be an epiphenomenon of default as in Russian, Greek, Icelandic, and languages like them. Rather the accusative path seems to be a separate and old rule. Its identity of form with the syntactic accusative is more likely to be historical, since historically -o seems to have originated in a focus marker (references in Miyagawa (1989)). Note that, while the present theory does not predict the adverbial accusatives in Japanese, it does not disallow them; they are just no more likely than any other case homophony. The prediction about adverbial case is not absolute. What would be a problem for the theory would be a wide range of adverbial accusatives in a language like Japanese. Also, cognate accusatives should not occur in a language like Japanese. On the other hand, as expected, nominative marks a wide range of nonargument NPs. Saito (1983: 248) argues for a "default" status of nomi-

100 A typology of case systems native in Japanese on the basis of examples like these (bracketings are Saito's on an example from Kuno (1973: 71)): (79)

a. yahari [s natu ga [s biiru ga umai ]] after all summer NOM beer nom tasty 'After all it's during summer that beer tastes good.' b. [s bunmeikoku ga [s dansei ga cilivized-country NOM male NOM [s heikinzyumyoo ga mizikai ]]] average-life-span NOM short 'It is in civilized countries that men are such that their average life-span is short.'

As suggested by Saito's bracketings, it is not at all clear that these NPs are clausemates of any other NPs. However, if the nominative here is a linker then it must be the any-NP case since these NPs are not arguments. Analyzing Japanese as type 2 allows a coherent account of a wide variety of facts about Japanese syntax. It allows the derivation of the Double -o Constraint from the general LDC. It also allows an account of some scrambling facts and the distribution of accusatives and nominatives on adjuncts. Most of all, it provides a unified account of these phenomena from general principles. The class of type 2 languages probably includes many of those dis : cussed by Comrie (1976: 284f.) and so may include Hindi, Punjabi, Persian, Turkish, Finnish, Hungarian, Georgian, Songhai, and Tagalog. Some of these cannot be pure type 2 systems, since they show split ergativity, but many are probably type 2, and others partially so (see §3.4). 3.3.3 Type 3 The next type, type 3, bears a strong resemblance to the last type, type 2. In the present system, the similarities will be captured by the minimal difference between these two types. Type 3 shares with type 2 the yes value for the Limitation Parameter but differs from type 2 in the Preference Parameter, where it has the value low. In this latter respect it is like type 1: (80)

Type 3: Limitation Parameter: Yes Preference Parameter: Low (Top-Down)

Four types of case system

101

As an example of a type 3 language I take Warlpiri. This will lead the Case A of Warlpiri to show an absolutive distribution and the Case B to show an ergative one: (81)

a. ERG: [ARG|-LA]

b.

ABS:

[

]

The absolutive is the default and the ergative is the argument case. The argument case is limited in a way similar but reverse from the accusative in Japanese (type 2). Of the two ergative-absolutive types, type 3 is less unexpected than type 4 of the next section. This is because in type 3 absolutive is the default case. Like nominative case in nominative-accusative systems, the absolutive is the morphologically unmarked, often zero, case. As with nominative, this has led to a general assumption that it is also a default for linking. On the present theory this only holds for type 3, not for type 4. The predictions for type 3 are somewhat less unexpected than for type 4, since type 3 has absolutive as the default case. Accordingly, I will concentrate more on type 4 than on type 3. In this section I will present an analysis based on Simpson (1991) that Warlpiri is type 3 with absolutive default case. In the next section I will present the unexpected ergative type. First of all, the linking of a simple transitive and intransitive is given in (82) and (83) respectively (Hale (1983: 6, 13)): (82)

Ngarrka-ngku ka wawirri panti-rni. man-ERG AUX kangaroo(ABs) spear-NONPAST 'The man is spearing the kangaroo.'

(83)

a. Kurdu child(ABs) 'The child b. Kurdu child(ABS) 'The child

ka wangka-mi. AUX speak-NONPAST is crying.' kapi wanti-mi. AUX fall-NONPAST will fall.'

Hale points out that the absolutive is the regular case of intransitive subjects regardless of theta role. This alone makes the absolutive much more likely to be syntactic than semantic. The linking for these first examples is straightforward and is the reverse of Japanese intransitives and transitives. The transitive will have two

102 A typology of case systems different cases. The ergative, being specified [-LA], cannot apply to the lower of the arguments. It is more restrictive than the default absolutive and so we get ergative on the higher argument, absolutive on the lower argument: (84)

panti- ( x

y ) 'spear'

ERG: [ARG/—LA] ABS:

ABS: f

]

[ ]

Since both cases are non-oblique (about which more below), the ergative NP is the subject, because it is the highest non-obliquely linked argument. In the intransitive, the [—LA] specified ergative is not applicable, because the only argument is the lowest (as well as the highest). This argument being the lowest, no [-LA] case can link it. This leaves the default absolutive for (83a): (85)

wangka ( x ) 'cry' ABS: [

]

The identity of the case which marks the subject of the intransitive and the object of the transitive above is what leads to the label "absolutive." So far, as in the case of type 1 versus type 2, we do not have evidence to distinguish type 3 from type 4. The simple data from transitives and intransitives could reflect type 3 or type 4 case systems. Simple one or two argument clauses are not enough to distinguish type 3 from type 4. Analyzing ergative as the argument case, restricted by [-LA], makes it subject to the Argument Linker Dissimilation Prediction. Double ergative sentences do not seem to occur in Warlpiri. The related Default Linker Non-Dissimilation Prediction says that in principle the absolutive should be repeatable. Warlpiri has a dative of goal and a rich set of semantic cases, with the effect that the default rarely gets a chance to apply more than once per clause. There is, however, one class of double absolutive sentences, the so-called cognate absolutives. These are syntactically and semantically parallel to the cognate objects of Icelandic and Classical Greek. And, in the present scheme, they receive identical analyses. Once again, we analyze the cognate object as a nonargument NP. The case that applies to such a non-argument NP will, in the absence of a more specific case, be the default case. In Warlpiri, the absolutive, as predicted, links cognate objects (Nash (1980: 196)):

Four types of case system (86)

a. Ngaju ka-rna

103

wangka-mi.

I(ABS) PRES-lSgNOM

Speak-NONPAST

'I am speaking.' b. Warlpiri ka-rna ngajulu wangka-mi. Warlpiri(ABs) Pres-I I(ABS) speak-NONPAST 'I am speaking Warlpiri.' This can be analyzed as follows. The verb wangka- 'speak' has the following argument structure in the linking for (86b): (87)

wangka ( x ) ABS:

[

]

( NP ) ABS: [

]

The cognate object has no special case and it is not an argument. The only case that is applicable is the absolutive. The treatment of absolutive as the default captures the cognate object construction. Further, we are forced to analyze the absolutive as the default, since it is repeatable and so does not show LDC effects. The only alternative would be to analyze the absolutive as a special homophonous lexical case for cognate objects. The main problem with this is that lexical cases are rarely if ever zero cases morphologically. The present approach makes interesting predictions about the status of dative NPs in Warlpiri. The verb in (86) and some other verbs in Warlpiri seem to occur in an ABS-DAT case array: (88)

a. Ngaju I(ABS)

ka-rna

wangka-mi.

PRES-1 SgNOM

Speak-NONPAST

'I am speaking.' b. Ngaju ka-rna-rla

(Hale (1983: 18)) ngarrka-ku wangka-mi.

I(ABS) PRES-lsgNOM-3DAT man-DAT

'I am speaking to the man.'

Speak-NONPAST

(Hale (1973: 333))

Although, as mentioned in §2.10 above, optionality of an NP is not a sure sign of adjuncthood, optional complements are very often adjuncts. And Levin (1983: 211) provides arguments that the dative with such verbs should be treated as semantic and the NP as an adjunct. Interestingly, the analysis of Warlpiri as type 3 in the present theory actually forces us to this conclusion. Consider again the argument structure of a verb like wangka- 'speak.' The two possibilities are that the NP for the one spoken to corresponds to an argument or not: (89)

a. wangka- ( x y|GOAL ) 'speak' b. wangka- ( x ) 'speak'

104 A typology of case systems If the argument structure is as in (89b), then the dative must arise as an adjunct.8 Note that, if the argument structure is as in (89a), the ergative will incorrectly apply to the first argument, since it is not the lowest. But this would be wrong. On the other hand, if the argument structure is as in (89b), the ergative is inapplicable, since the argument is the lowest, and so the absolutive is the only applicable case. Thus the present theory forces the conclusion that Levin argues for anyway, a desirable result. Furthermore, this theory makes the prediction that if there are dative arguments in second place in the argument list, then the ergative is not only possible but obligatory. This is confirmed: the ERG-DAT array is also possible: (90)

Ngajulu-rulu ka-rna-rla I-ERG

karli-ki

warri-rni.

PRES-lsgNOM-3DAT boomerang-DAT seek-NONPAST

'I am hunting a boomerang.'

(Hale (1973: 335))

It seems, though, that these datives are different at least in being obligatorily expressed. I take them to be lexically case-marked arguments, which is required by the theory. Otherwise the ergative-marked argument is lowest, which is not possible in the theory, since the ergative cannot apply to lowest arguments in type 3. A complete analysis is beyond the scope of the present section. But one issue that does directly concern the conclusions here is the considerable controversy over whether Warlpiri ergative is a syntactic case or a semantic case. The latter view is necessitated by the analysis proposed by Jelinek (1984). She analyzes the NPs in Warlpiri as adjuncts and the agreement markers/pronouns in the verbal complex as the true arguments. This approach gets the optionality of NPs (since adjuncts are usually optional), but under this account the case-marking facts still need to be explained. The first problem with the NPs being linked by semantic case is that the absolutive is not restricted to any particular semantic class of arguments. The absolutive must at least be a default case as argued here and, in a different framework, by Simpson (1991). I will assume double linking by agreement and case. In that case, Warlpiri allows free compatibility of these two types of linkers. I will now review two possible views of ergative, as a semantic adjunct case or as a grammatical case specified [ARG|-LA]. Evidence that the ergative might be "semantic" comes from intransitives with the sole argument in the ergative (Levin (1983: 149) citing Hale (1982: (232a-d))):

Four types of case system (91)

105

a. Ngarrka-ngku ka ngungkurru-pangi-rni. man-ERG PRES snore-dig-NONPAST 'The man is snoring.' b. Kurdu-ngku ka ngaany-kiji-rni. man-ERG PRES breath-throw-NONPAST 'The child is breathing, expelling breath.' c. Kurdu-ngku ka ngal-ngal-kiji-rni. child-ERG PRES pulsing-throw-NONPAST 'The child is panting (from exhaustion).' d. Karnta-ngku ka kuntul-pi-nyi. woman-ERG PRES cough-hit-NONPAST 'The woman is coughing.'

As indicated in the glosses, these verbs are morphologically complex (nominal preverb plus verb) and denote the performance of a bodily function and so should be regarded as having agents. Most interestingly, however, Hale notes that the verbs in (91b-d) have all been found with the absolutive as well. He suggests a diachronic explanation that the preverb is historically an incorporated nominal and that the verbs are changing towards the regular intransitive pattern. Thus all the stages would fit in with Warlpiri's being a type 3 language. To summarize, two possibilities for the ergative are as follows: (92)

a. ERG: [ARG|-LA] b. ERG: [ARG|AGENT]

While it is true that there is a loose correlation of ergative with agent of causation, if we analyze the ergative as the case for the agent, we would face the following problem: why is the ergative obligatory in transitive sentences but optional in the few intransitive sentences where it can appear? On the semantic ergative account, there is no reason to expect that only ergatives in intransitves will alternate with the absolutive. The correlation of the alternation with intransitivity is left unexplained. Second, if the ergative is a semantic case, it does not seem to pick out a natural semantic class. Nash notices that, in traditional thematic role terminology, the ergative marks agents ((82), (90), (91)) and experiencers of perception: (93)

Ngatjulu-rlu ka-rna-ngku I-ERG

'I see you.'

njuntu

nja-nji.

PRES-lSgNOM-2sgACC yOU(ABS) See-NONPAST

106 A typology of case systems Thus if all ergatives in Warlpiri are semantic, the range is unexpected, or else we need multiple semantic uses of the ergative. The problem still is why we do not get more ergatives on intransitives, if those intransitives qualify for a semantic ergative. Why is there no ergative in (83a) above? Warlpiri is far from being of the "active" type, where the agent gets a special case regardless of transitivity. Two solutions are possible that are compatible with the analysis of Warlpiri as type 3. The first is to follow Nash (1980) and Simpson (1991) in taking the intransitive uses .of the ergative to be quirky case for subjects, e.g. (91b) above: (94)

ngaany-kiji ( x ) 'breathe' (ERG: [ARG|1])

As noted by Levin (1983), this quirky case violates the GB approach to external argument. The obligatoriness of the transitive ergative versus the optionality of the idiosyncratic ergative can be captured by making the idiosyncratic ergative optional as in (94). The homophony of the two ergatives is historically explained. The other possible tack in the present set-up is to adopt a synchronic version of Hale's (1982) solution for this class. The morphological complexity of the verbs in (91) would be taken to reflect their ambiguity between a complex verb with two arguments and a verb with one argument: (95)

a. kiji ( x y[ngaany 'breath'] ) 'throw' b. ngaany-kiji ( x ) 'breathe'

The verb is structurally ambiguous between having two arguments one of which is incorporated into the verb (95a) and the other structure where a complex verb takes one argument. The diachronic trend is from the complex structure (95a) to the simple structure (95b). Note that in the first case the ergative will apply regularly to the first argument, because it is not the lowest, there being a lower argument. And in (95b), the absolutive must apply, since the only argument is the lowest. The optionality of the ergative in these cases is captured by the fact that there are two structures only one of which leads to ergative on the subject. The reason this is probably the less attractive of the two approaches is that this verbal compounding is no longer productive (Nash (1980: 201f.)). I do not have the data to decide between these alternatives.9 It is worth pointing out that both of them are equally compatible with the present

Four types of case system

107

theory and that both seem preferable to the "semantic" case alternative. Both the idiosyncratic case solution of Nash (1980) and Simpson (1991) and the synchronic version of Hale's (1982) solution handle the optionality of the ergative in "intransitive" contexts better than the semantic case view. If ergative is semantic everywhere, there is no reason to expect the different treatment of ergative in the two contexts. Further, the class of verbs that allows ergative in the apparent intransitive context is extremely small. If ergative is semantic, there is no reason why it should be disfavored on intransitives, but the fact is that almost all agentive intransitives take absolutive obligatorily. The two alternatives advocated here actually lead one to expect this pattern. If Nash and Simpson are right, then the rarity of the intransitive ergative follows from the marked character of idiosyncratic case. If the synchronic version of Hale's solution is correct, then the rarity of the ergative "intransitives" follows from the small number of structurally ambiguous verbs. The structural ambiguity is marked and so the apparent intransitive ergative is marked. Type 3 is the expected ergative type with default absolutive. I have argued, albeit tentatively, that Warlpiri is such a language. This is not to deny that there is an active type, and that for these languages we need an ergative of agent.10 However, some languages do seem to be type 3 and bear out the predictions above. In addition to Warlpiri, other likely type 3 languages are surveyed by Austin (1982). This wide range of Australian languages all take absolutive cognate objects and so are probably to be taken as type 3.

3.3.4 Type 4 Type 4 is the ergative-absolutive type not expected on the usual nominativeor-absolutive-as-default view. The prediction that the case with a basic ergative distribution might be the default case is not expected under most theories of syntax. If such a type exists, it would provide a striking confirmation of the typology and the theory that produces it. In this section I will analyze Kabardian, a Northwest Caucasian language, as type 4 and test the resulting predictions. It will turn out that type 4 languages should be the mirror image of type 1. The Limitation Parameter has the value no, and the preference parameter has the value high: (96)

Type 4: Limitation Parameter: No Preference Parameter: High (Bottom-Up)

108 A typology of case systems This leads to an ergative-absolutive system where the ergative behaves as a reverse type 1 accusative and the absolutive as a reverse type 1 nominative: (97)

a. b.

ABS: [ARG] ERG:

[

]

Such behavior follows from the reversal of only one of the parameters in comparison to type 1. Thus the absolutive should only appear once per clause and should mark the lowest available argument. The ergative, on the other hand, should be able to appear multiple times in accord with its default status. Furthermore, in the absence of more specific cases, the ergative should appear. Thus a range of adverbials, barring the availability of specific adverbial case, should appear in the ergative. I will argue that the Northwest Caucasian language Kabardian is of type 4. Kabardian is the name given to Eastern Adyghe (or Circassian, as Adyghe is also known). Eastern and Western Adyghe are extremely closely related and may be considered dialects of one language. Most data will come from Kuipers (1962) who describes a dialect of Kabardian, so I will use this name. The conclusions drawn for Kabardian apply equally to Adyghe generally. For some other descriptions of the language, see Kuipers (1955, 1960), Rogava and Keraseva (1966), and Catford (1976). I will assume for the purposes of the following discussion that Kabardian is syntactically accusative, but nothing depends on this. Rather, in the following, I wish to establish the relation of the ergative as default case and the absolutive as the any-argument case.11 That is, Kabardian is to be shown to be type 4 and all that that entails. Like many Caucasian languages, Kabardian has an ergative-absolutive case-marking system and verbal prefixes. In the basic transitive the subject is ergative and the object is absolutive: (98)

I's-m ss-r j man-the-ERG horse-ABS (ABs3rd)ACT-killed T h e man killed the horse.' (Kuipers (1962: 234, 238))

The linking for this sentence is the following: (99)

J9-W9h'ah§ ( x

y ) 'killed'

*ABS: [ARG] ERG:

[

]

ABS: [ARG] ERG:

[ ]

Four types of case system

109

If Kabardian is type 4, it has the high value of the Preference Parameter. This means that the LDC will question the linking of the higher of pairs of arguments. Applied to (99), this means that if the higher argument were linked by the absolutive, the structure would violate the LDC, because the higher of two arguments would be linked by a case with the same LINK value as the most restrictive case applicable to the lower argument. The use of absolutive on the lowest argument of course cannot violate the LDC. Thus we get absolutive on the lower argument but not the higher one. For the higher one, the only case that passes the AC, the LDC, and the RC is the default ergative. It applies, correctly giving us the linking for the example in (98). The reason that ergative is called "ergative" and absolutive "absolutive" is that the same case links the objects of transitives and the subjects of intransitives, and another case marks the subjects of transitives. In type 4, the reasons for this are different from those for type 3. In type 3, the ergative was limited by the specification [—LA] to appearing on arguments that are not the lowest on the argument list. Otherwise the default absolutive applies. In type 4, on the other hand, the ergative is the default. The ergative-absolutive distribution follows from the high value of the Preference Parameter and the LDC. The [ARG] case could only appear once, by the Argument Linker Dissimilation Prediction (a consequence of the LDC). Because the Preference Parameter is set at high, the [ARG] case could only appear on the lower of two arguments in a transitive. In an intransitive, this same case will be the most restrictive case that passes the AC for the sole argument. The [ARG] case in type 4 is not limited to lowest or highest arguments - type 4 has the value no for the Limitation Parameter. Thus in an intransitive the sole NP is absolutive and the verb agrees with the absolutive NP: (100)

i'a-r k'°-ah§. man-ABS (ABs3rd)went T h e man went.'

(Kuipers (1962: 236))

If the absolutive is 3rd person (singular or plural), as in this example, we get zero agreement. The absolutive agreement for non-third person is overt (see Kuipers (1962: 234)). In the present system, both syntactic cases are applicable to this sole argument, and so the more restrictive absolutive links the NP:

110 A typology of case systems (101)

k'°-ah§ ( x ) 'went' ABS: [ARG] ERG:

[ ]

With only one argument, the LDC is irrelevant. The RC dictates absolutive. Since the same case is used here and on the second argument of the transitive, type 4 and Kabardian in particular are called "ergativeabsolutive." As noted in the last section, such simple examples are consistent with an analysis as type 3 or type 4. Once again, the analysis of a language as of a given type makes a network of predictions. I will now show that these predictions are borne out in the case of Kabardian. Before proceeding, a note on agreement is in order. As with type 1 languages, it is not always clear whether absolutive (or nominative) is the linker and the verb agrees with the absolutive NP or whether agreement is the linker and the absolutive or nominative is the lack of case that complements this type of linking. A third possibility is that both are linkers. Which of these is the right way to analyze agreement does not affect the main point here which has to do with the interaction of this kind of linking with other forms of linking in the language. In the following, the symbol ABS abbreviates absolutive and its corresponding agreement. In Kabardian, the first agreement slot is always with the absolutive. Kabardian, like other Northwest Caucasian languages, has a wide array of other verbal agreement markers which are usually analyzed as agreement slots in the verbal complex. The other agreement morphemes will need more restrictive information, but they will also be compatible with the ergative, my main focus. The actor prefix (see (98)) I will analyze as marking the highest ergative argument (cf. agreement in South Asian languages) and so need not be considered a linker. Whether or not the other agreement markers are linkers is an open question, but, if they are, then they can be handled as follows. First the "indirect object" prefix I will treat as a standard dative of goal: (102)

IOagr: [ARG | GOAL]

I use this terminology to follow tradition without any theoretical significance. The other relevant prefixes are very much like incorporated prepositions, although their synchronic status is not clear from the literature (Kuipers (1955: 202)). They must be more restrictive than the ergative, but they are all compatible with it.

Four types of case system

111

The first prediction, the Argument Linker Dissimilation Prediction, when applied to Kabardian under the type 4 analysis, yields the expectation that no co-arguments are both absolutive. From the sources available, this seems to be correct. Rather, wherever we might get two absolutives, i.e. where there are two co-arguments, the higher should appear ergative. This was true of the simple transitive. I will now show that it is true of many other sentence types as well. As far as the ergative is concerned, we have two predictions. The first is the Default Linker Non-Dissimilation Prediction, which means that multiple ergatives per clause are in principle possible. If two or more co-arguments are not subject to any lexical cases, all but one must be ergative. The first piece of evidence that ergative is the default is that in ditransitives the goal is ergative. Kabardian does not have a case for goals, although it does have goal agreement (see (102) above). In a ditransitive, we will have three arguments, none of which is eligible for a lexical case. In such a clause, the theory leads us to predict ERG-ERG-ABS, and this is precisely what we get: (103)

I'a-m ss-r f6z9-m jarajtahs. man-the-ERG horse-ABS woman-ERG (su3rd)IO-ACT-gave 'The man gave the horse to the woman.' (Kuipers (1962: 238))

The linking for this sentence follows directly. Recall that IOagr and ergative case are compatible: (104)

give ( x

y|GOAL

*ABS

IOagr

ERG

*ABS

z) ABS ERG

ERG

For all three arguments, the most restrictive applicable linker of the type case is absolutive. The LDC in Kabardian will question and reject the absolutive on each of the two higher arguments, because the absolutive is the most restrictive linker of that type (case) that passes the AC for the lowest argument. The absolutive is more restrictive than the ergative, so where the LDC does not rule it out, here on the lowest argument, it will have priority by the RC. For the higher two arguments, the only remaining case is the default ergative, which applies to both arguments.12 Goal agreement is not the only verbal agreement marker which is compatible with the ergative case. The prepositional agreement mor-

112 A typology of case systems phemes are also compatible with the ergative. For instance, a comitative (105a) or benefactive (105b) prefix co-occurs with ergative nominals: (105)

a. I'a-m s9-da-lazah§. man-the-ERG su 1 st-coM(3rd)-worked 'I worked (together) with the man.' (Kuipers (1962: 238)) b. I's-m s3-x°a-lazahs. man-the-ERG su 1 st-BEN(3rd)-worked 'I worked for (in the interest of) the man.' (Kuipers (1962: 238))

One question that immediately arises is whether the benefactives and comitatives are arguments (original arguments or promoted from adjunct) or adjuncts. I will assume that they are adjuncts in view of their free omissability and more importantly their apparent free addability: (106)

lazahs ( x ) 'worked' ABS

da- ( y ) 'together with' ERG

ERG

The comitative adjunct predicate is incorporated into the verb with the comitative morpheme occupying a slot in the verbal complex. Since this comitative NP is not a verbal argument, only the default may apply. In Kabardian the default is the ergative, and so we get ergative comitatives (and similarly ergative benefactives). These last examples also provide evidence for the Elsewhere Pattern Prediction. Adjunct NPs are likely to be marked ergative, and, in the absence of some case to block them, have to be ergative. The comitatives and benefactives are such adjuncts. They have a verbal agreement marker, but this does not preclude case, and so they must get default case, i.e. ergative. The two predictions about the default, the repeatability and Elsewhere predictions, are even more strikingly supported by "semantic" uses of the ergative, where there is no corresponding verbal prefix. In these sentence types, there is no verbal prefix targeting the ergative NP in question, but the ergative appears on the NP. Adjuncts, if they are not linked by a relatively restrictive linker incompatible with case, are eligible for ergative and receive it. The comitatives and benefactives demonstrated possible double linking of adjuncts by prefix and ergative. I now turn to ergative acting as the sole linker of an adjunct.

Four types of case system

113

The ergative applies whenever there is no case at all for the adjunct. I now review adjunct linking where there is no verbal prefix. First, the ergative can link what would be translated by an (in) to PP in English: (107)

a. q'ahla-m mahk'°a. town-ERG (su3rd)goes 'He goes to the town.' b. maza-m jahh. forest-ERG ACT-(ABs3rd)carry They carry it to the forest.'

(Kuipers (1962: 238))

(Kuipers (1962: 238))

Similarly, the ergative can mark an NP denoting the time at which an event occurs: (108)

har za$9-m mabahna. dog-ABS night-ERG (ABs3rd)barks T h e dog barks at night.'

(Kuipers (1962: 238))

The ergative also marks the "object of comparison" with no corresponding verbal prefix: (109)

1'a-r faza-m naxra nax"6z$. man-ABS woman-ERG older (ABs3rd)is T h e man is older than the woman.'

(Kuipers (1962: 238))

The sentence in (108) with the ergative time-at-which adjunct receives the following analysis: (110)

mabahna ( x ) 'bark' ABS

Adverb ( y ) ERG

ERG

In the verb's argument structure there is one argument. The absolutive, as the more restrictive case, wins. The only case that can link the adjunct is the ergative default, since apparently there is no preposition or case meaning 'at.' At least, if there is such a preposition, it is optional. When such a preposition does not apply, the theory predicts the ergative must apply. Either way, the theory correctly predicts the ergative here and similarly for the other adverbial constructions. Note that Kabardian poses a problem for GB Case Theory. If ergative is the surface manifestation of an Abstract Case, this Case is assigned to a class of NPs that do not form a natural class under normal GB assumptions. In GB the thematically highest NP is generated outside VP or more

114 A typology of case systems recently outside V'. At any rate, the external argument is supposed to get the INFL case. If the INFL case is absolutive then it is a mystery why the external argument is ergative in transitives. If the ergative is the INFL case then it is not clear why we can get more than one ergative per clause or why ergatives should appear without agreement. Ergatives on adjuncts would also be a mystery. The present account actually predicts that a language like Kabardian should exist. The absolutive only appears once, but the ergative can appear more than once on co-arguments, as predicted by the LDC interacting with the specifications making absolutive the argument case and ergative the default case. The range of ergative NPs receives a unified treatment. In this section I have demonstrated that Kabardian is an excellent candidate for type 4. Other Northwest Caucasian languages seem to have multiple ergatives (Comrie (1981: 209)) and so should probably be analyzed as type 4. Thus the most unexpected type and the one which rounds out the typology of two-case systems has been found. The theory allows an elegant and unified solution to some very problematic data. 3.3.5 Summary In this section I have shown that the theory of case outlined in §3.1 induces a typology of two-linker systems on the basis of two simple parameters. I have further shown that all the predicted types occur and that they show exactly the Case behavior expected on the present theory. Finally, I would like to consider possible directions for further crosslinguistic work. Some of the predictions made here are based on likelihood. If a given case is the default case for a given language, it is likely to occur more than once per clause, and it is a good candidate for a general purpose adjunct case. Further testing of these predictions is required, but the initial indications above are promising. Note that this is a generalization of an observation suggested by Poser (p.c.) that languages with a double accusative constraint (like Japanese) tend not to have adverbial accusatives. We can now extend this generalization to all the four cases in all the four types. The [ARG] case should show LDC effects, and, since it is limited to arguments, it is not a candidate for general purpose adjunct case. In languages with two grammatical cases, it should be the other, default, grammatical case that shows Elsewhere behavior. By providing cross-linguistic definitions of case based on independently motivated features and by embedding the definitions in a systematic theory of linking,

Split ergativity 115 we get for each language an explanatory account of a wide range of facts. Specific constraints can be seen to follow from universal principles. 3.4

Split ergativity

As touched upon in the previous section, most ergative languages are not uniformly ergative but rather are "split ergative": some classes of NPs show ergative-absolutive marking but other classes show a nominativeaccusative pattern. The "split" may be conditioned by a number of factors.13 Some split ergative languages have a split between an ergative-absolutive system for one set of NP types and a nominativeaccusative system for another set of NP types (those higher on the hierarchy of features). Other split systems are conditioned by properties of the verb or the clause. The system developed so far can handle such systems at least as well as previous theories and may have some special advantages. It is also well known since Silverstein (1976) that some classes of NPs are implicationally more likely to show a nominative-accusative pattern. For example, if pronouns and nouns denoting humans show an ergativeabsolutive pattern then other nouns will. NPs can be ranked on an implicational scale or "hierarchy of ergativity" with pronouns and human-denoting NPs on the "least ergative" end of the hierarchy. The question of course is whether the system outlined in the previous sections can be naturally extended to handle split-NP ergativity. I will outline first a conventional solution which has the same advantages and disadvantages of analogous solutions in other theories. I will then turn to a less conventional solution which draws on the unique strengths of the present approach. First, split systems demand that the two parameters suggested at the outset of this chapter need to be set in a more refined way than I indicated. In general, a language will have to be able to have multiple values of the parameters depending on just those sets of factors that condition case splits (for discussions of case splits, see, for example, Silverstein (1976), Dixon (1979)). So, for instance, Warlpiri is type 3 in nouns, but it is nominative-accusative in its pronouns and agreement markers/verbal clitics. Thus, for a complete account of Warlpiri, we would need to set the values of the parameters differently for different NP classes. The range of such features that can be relevant to case marking suggests another modification to our approach. So far, I have assumed that

116 A typology of case systems linking is done in the lexicon. Argument structures are assigned linkers freely and then subjected to the three constraints. Surface realization must conform to the information contained in the fully specified argument structures, including the information from linking (expressed in the AVM formalism by the value REAL). If an argument is linked by dative at argument structure, it must correspond to a dative NP in PS Structure. The checking between argument structure and PS is maximally simple, because all the work of linking is done at argument structure itself. The problem comes with split systems and other case marking that makes reference to features that are more plausibly characteristic of lexical items than of argument structure. We will see this more in the next chapter (§4.1.2) with Nama linking. To do linking at argument structure in Nama, we will need to specify clause level properties like declarativeness. Verbs may be specified for declarativeness. But the real question is the split systems. If we are to handle a split by noun versus pronoun at argument structure, we would have to freely instantiate features for pronounhood and nounhood on the arguments and generate a case frame for each of the possible combinations of pronouns and nouns. The potential for redundancy is very great. One way out is to have linking occur "after" lexical insertion, as in Kiparsky (1987).14 In the present theory, the only change this would entail from our previous practice would be to treat each of our three main constraints as constraints on lexical insertion. Items would be freely inserted and then checked against the AC, the LDC, and the RC. The AC would refer to applicability not just of the linkers' LINK value in comparison with the argument structure (as above) but also with inherent features of the NP. So a case, say ergative, limited to nouns would be checked both against the argument and the NP it links. The problem, of course, is to limit the number of such features that can be checked. I leave this open for further work. On the lexical insertion approach, the LDC would check pairings of inserted NPs and arguments to make sure there is no higher argument with the linker on that argument-NP pair as its most restrictive possibility of that type. And the RC similarly would check inserted items and their associated arguments for other more restrictive and otherwise legal Unkings. If any are found, the whole structure is out. I leave it open which of the two approaches, linking at argument structure and linking after lexical insertion, is preferable in the long run. What I am developing here is equally compatible with either view.

Split ergativity

117

Because none of the data I examine directly requires the lexical insertion approach, I will for simplicity's sake still speak in terms of linking at argument structure. Before turning to three case systems and a less conventional approach to NP split ergativity, I will first demonstrate for the sake of illustration how the present system can incorporate a conventional feature-based approach to split ergativity and what advantages the present system has. The feature-based approach seems to be most inescapable with splits conditioned on tense and/or aspect. The analysis here is meant as purely illustrative and I do not pretend to make strong claims about this particular analysis. I will use some simple examples from Georgian, a South Caucasian (Kartvelian) language which has an ergative-absolutive system with verbs in the aorist tense, and a nominative-accusative system for present tense verbs (data are from Comrie (1978: 351f.)):15 (Ill)

a. Student-i midis. student-NOM goes T h e student goes.' b. Student-i ceril-s cers. student-NOM letter-Ace writes T h e student writes a letter.' c. Student-i mivida. student-ABS went The student went.' d. Student-ma ceril-i dacera. student-ERG letter-ABS wrote T h e student wrote the letter.'

Note that the same ending, -/, is used both as an absolutive case in the ergative-absolutive system and as a nominative case in the nominativeaccusative system. Although there seems to be no way of avoiding introducing tense into linking on any theory, the present theory does have the advantage of allowing the morphological -/ to correspond to one case regardless of the label:

118 A typology of case systems (112)

a. ERG:

ARG

(-ma)

-LA

b. ACC:

ARG —HA

c. NOM/ABS: [

(-0

Note again that default for linking is not necessarily the same as morphological zero case (or alternatively, lack of case), especially since none of the cases in Georgian is a zero case. This default case need not be specified for tense or aspect. Only the other two cases have to have some additional feature or other provision to limit their occurrence to verbs of a particular tense. This can be done by specifying the cases for tense, but I have not done this in (112). Rather, the arguments they are to unify with must also be specified for tense. Alternatively we could use a feature co-occurrence restriction (FCR), stated as a material conditional, mandating present tense when ergative case is used for linking, i.e. prohibiting any feature structure from containing the path [SUBCATJREAL|ERG] and [TENSE|PRES]:16 (113)

[SUBCAT|REAL|ERG] D [TENSE |AOR]

(114)

[SUBCAT|REAL|ACC] D [TENSE |PRES]

The FCR in (113) will rule out the following AVM and any it subsumes, because [TNS [PRES]] and [TNS [AOR]] are incompatible:

(115)

17

CAT [V] TNS [PRES] SUBCAT [REAL [ERG]]

This gets us the restriction that ergative case will only show up with verbs in the aorist tense. Similarly, the FCR for accusative case will ensure that anywhere we find accusative we find the present tense verb. The main advantage of the present system consists in its ability to capture the identity of cases across the split of a system. For the rest, a feature-based analysis will be required, as in other theories. 3.5

Three-case systems and quasi-split ergativity

The foregoing sections have explained the typology of two-case systems predicted on the present theory and shown that the predicted types actually occur. Two-case systems, however, do not exhaust the possibilities

Three-case systems and quasi-split ergativity

119

of the theory, nor do they represent a limit of the data for there are a number of languages with what must be analyzed as a system of three grammatical cases.18

3.5.1 Three-case systems: Diyari So far the discussion has centered on the question of whether "the" ARG case has the further limitation to [—LA] (type 2 accusative) or [-HA] (type 3 ergative). But the system would allow us to have a language in which there is both an ARG case and an ARG case limited by [—LA] or [—HA]. We have already noted that when dealing with two-case systems, it does not matter whether [—LA] or [-HA] are taken to be values dependent on [ARG] or on [LINK] directly. The reason is that the LDC just compares linkers on the basis of identity versus non-identity of their LINK values: (116)

Linking Dissimilation Constraint: Given two co-arguments, x and y, x higher than y, if y is linked by a linker B that has the same LINK value as the most restrictive linker of that type applicable by the AC to x, call it A, and there is a less restrictive linker C, applicable by the AC to y, then the linking is ill-formed.

Thus the following two cases will not show dissimilation effects with respect to each other; that is, the use of one will not preclude the use of the other: (117)

a. X: [ARG|-LA]

b. Y:

[ARG]

In two-case systems only one of these cases is present, in addition to the default case. But in three-case systems, we do want these two cases. Now it matters whether these cases dissimilate. In the following I will examine only cases where these cases, while separate, do dissimilate. I do not claim that this is the only possibility.19 However, a fully general LDC which would provide for both languages which do have distinct but dissimilating [ARG] cases (see note 18) and those that have distinct but nondissimilating cases would be based not on identity of LINK value but something like following: (118)

Distinctness: Two Linkers A and B are distinct if and only if the LINK value of each contains a path not found in the other. Otherwise they are non-distinct.

120 A typology of case systems Thus, among the following, X is distinct from X' and Y r , but, most importantly, X' is not distinct from Y: (119)

a. X: [ARG|-LA]

b. X':

[ARC

[-LA c. Y:

[ARG]

Since the only path in the LINK value of Y, namely [ARG], is also contained in the LINK value of X', the two linkers are non-distinct. They will therefore not dissimilate under the revised LDC: (120)

Linking Dissimilation Constraint: Given two co-arguments, x and y, y higher or lower than x, if y is linked by a linker B that is not distinct from the most restrictive linker of that type applicable by the AC to x, and there is a less restrictive linker C, applicable by the AC to y, then the linking is disallowed.

The only difference in the new LDC is that dissimilation is based on nondistinctness as defined above, rather than on identity as earlier. Thus the X case above will dissimilate with the Y case, but the X' case will not dissimilate with Y. 20 The system now provides for three-case systems. One case will be the default, one the plain [ARG] case and the other the limited ARG case, i.e. [ARG|— XA] instantiated as [ARG|— HA] or [ARG|-LA] depending on how the Preference Parameter is set. In the following I will concentrate on the latter possibility, i.e.: (121)

a. ERG: TARG

[-LA b. c.

NOM: [ARG] ACC:

[ ]

Note that restrictiveness strictly orders this inventory of three cases. The reason I have labelled (121a-c) "ERG," "NOM," and "ACC," respectively, will become apparent once we consider how they would interact: (122)

( x } intransitive NOM: [ARG] ACC:

[ ]

Three-case systems and quasi-split ergativity (123)

121

y ) transitive

(x ERG:

\ARG

1

*NOM: [ARG]

[-LAJ NOM: [ARG] ACC:

ACC:

[

]

[ ]

To the sole argument of an intransitive verb, only the nominative and the accusative can apply, since the other case, the ergative, cannot appear on lowest arguments. It can apply to the higher of the two arguments of the transitive, and since it is the most restrictive and passes the LDC (the value low is selected for the Preference Parameter). On the remaining argument of the transitive, the nominative and the accusative pass the AC but the accusative must apply since the NOM would violate the LDC. This is because the NOM is non-distinct from the most restrictive case applicable to the higher argument, the ERG. There are languages with three distinct case forms for some class of nominals, and for such languages, a three-case system is called for. One such language is Diyari (Austin (1981), Goddard (1982)) where female personal names, derived dual and plural common nouns, singular first and second person pronouns, and third person pronouns of all numbers have three distinct case forms. These are labelled "A" (mnemonic for transitive subject or - loosely - "agent"), and " O " (for object), " S " (for intransitive subject). Contrast the non-feminine ('nF') singular pronouns in the following (Austin (1981: 241)): (124)

a. tinka-ni nulu nayi-yi parati/ tudu night-LOC 3sg.nFA see-PRES light-ABS fire-ABS y arki-y arki-1 adi-nani REDUP-bum-DUR-RELds

'In the night he saw the light of the fire burning.' b. rjada-ni pulali nina warara-yi patara miri then-LOC 3DIA 3sg.nFO leave-PRES box tree above Then they left him at the top of the box tree.' c. nawu miri-ta tara-irja-na tadi-yi 3sg.nFS above-OI go UP-PROL-PART -PRES

'He went up and up at the top (of the tree).' These forms are also often labelled "ergative" for the case which appears on transitive subject nouns of those classes; "accusative" for the case of object nouns of those classes; and "ergative" for the case which appears

122 A typology of case systems on transitive subject nouns of those classes. The analysis of these sentences would be along the lines of (122) and (123): (122) is the analysis for (124c), and (123) for (124a,b). Other nominals do not show three distinct forms. The morphology of these other nominal classes seems to be on an ergative-absolutive pattern with a zero ending for the S and the O and an ending -// (or -yali), for A. Goddard (1982) argues that the best analysis of this situation is not to say that this second group of NPs has a different case system but rather that it too has a three-case system with syncretism between the S (nominative) and the O (accusative). This results in an "absolutive" pattern that is merely an artifact of the syncretism of forms, not an aspect of the system. Goddard is careful (as many are not) to distinguish between case forms and (morphological) "case systems." These case systems are systems of morphological oppositions and are defined as substitution classes of nominals. There is no reason why a morphological case cannot be defined more abstractly than as the collection of instances of an identical ending. To illustrate how uncontroversial such morphological analyses are, Goddard points out that the traditional analysis of Indo-European languages provides for three cases for all nouns even if one class of them, neuter nouns, never shows a surface distinction between nominative and accusative. Masculine and feminine nouns show a distinction between nominative and accusative (in fact both in the singular and the plural), but neuter nouns never have distinct forms for nominative and accusative: neuter nouns have one form for the singular of both the nominative and accusative and one form for the plural of those cases, e.g. penculum 'danger' (nom/acc sg) and pencula 'dangers' (nom/acc pi). Nonetheless morphological analyses of Latin (and other Indo-European languages with a nominative-accusative distinction on some noun class(es)) have a neuter nominative and a neuter accusative, which happen to sound the same. The test is substitution: substitute (or in the case of a dead language, look for an example of) a noun from one of the classes that does show a distinction and one can "tell" which case the neuter noun bears morphologically, but covertly. According to Goddard's (1982: 169) definition, a case is "a class of nominal forms which are mutually substitutable in certain syntactic or semantic environments given that any two cases, casei and casej, are formally distinguished by at least one subclass of nominals." Interchangeability as a criterion of case facilitates the statement of morphological patterns.

Three-case systems and quasi-split ergativity

123

In fact, there are, according to Goddard, good reasons for preferring a more abstract analysis of the system in languages like Diyari as follows: all nominals have the three-case system but not all have three distinct forms. Some have only two. This analysis follows from the substitution class definition of case. If we substitute a dual nominal which does have three distinct case forms, then we can "tell" which case a given nondistinct nominal has but does not exhibit directly: (126)

f yula 1 \ yunduj

fkintala 1 \ kintala-wula-naj

nanda-na wara-yi

J 2du(ERG) 1 J dog(ACC) 1 hit-PART AUX PRES \ 2sgERG J [2-du-ACC J 'You two/you(sg) hit the dog/two dogs.' Here in each of the NP positions we can substitute for an endingless pronoun (yula) or noun (kintala) one that does show three distinct forms. Thus we can say, based on the substitution test that yula here is ergative even though in the context of an intransitive sentence, the same surface-ambiguous second person singular form yula would be replaceable with the unambiguously nominative second person singular form yini. Thus by the substitution test, the yula in (126) is morphologically ergative and the yula in an intransitive is morphologically nominative despite the lack of a surface opposition. The three-case analysis has several major advantages. Goddard points out that the three-case analysis makes the treatment of certain constructions much more straightforward. First, there is a construction found in many Australian languages including Diyari, in which an NP denoting something possessed as an inalienable part of the person or thing possessing it appears next to the NP denoting the possessor and appears marked for the same syntactic function (Austin (1981: 138)). On Goddard's analysis some NPs have a three-way case distinction and some an ergativeabsolutive distinction. Goddard (1982: 172) points out that the generalization can be stated more simply on a uniform three-case analysis: both NPs appear in the same case (Austin (1981: exs. 286, 287); Goddard (1982: exs. 2-3)):

124 A typology of case systems (127)

a. nulu

feana

mara]

nanda-na wara-yi

3sg.nF-ERG lsg-ACC hand(ACc) hit-PART AUX-PRES

'He hit my hand.' b. [yini milki] fanma-yi-la 2sgNOM eye(NOM) be-open-PRES-Ni

'Your eyes are open now.' In (127a), mara 'hand' is in the accusative just as unambiguous rjana 'me' is. Similarly in (127b), milki 'eye' is in the same case as unambiguous yini 'you'. The three-case analysis is even more advantageous in the pronoun-asdeterminer construction. In Diyari, a third-person pronoun in front of a common noun may form an NP interpreted definitely. Third-person pronouns have three distinct forms but common nouns do not. On the uniform three-case analysis the "definite" NP can be analyzed as consisting of two elements with the same case and can in turn be analyzed as having that single case (Austin (1981: exs. 176, 181); Goddard (1982: exs. 4-5)): (128)

a. [nawu kana] wapa-yi 3sg.nF.NOM man(NOM) go-PRES 'The man is going.' b. nulu pulana [nina putu] yirjki-na wara-yi 3sg.nF.ERG 3duAcc 3sg.nF.ACC thing(ACc) give-PART AUX-PRES 'He gave them that thing.'

On the analysis with three cases for some noun classes and two cases (ergative, absolutive for others - absolutive covering what on the uniform three-case analysis is analyzed as covert nominative and accusative) the "definite N P " in (128a) would be made up of a pronoun in nominative case and a noun in absolutive case. What's worse, what case does the NP as a whole have? On the three-case analysis one can simply say that both parts and the whole have nominative case. Similarly in (128b) the straightforward accusative for both constituents of the bracketed NP and the NP itself would, on the conventional split-system analysis, involve an NP of hard-to-decide case consisting of an accusative pronoun and an absolutive noun. 21 This morphological analysis, if accepted, does have syntactic consequences on the present theory. Recalling that "direct" linking is direct in the sense that difference of morphological case implies difference of

Three-case systems and quasi-split ergativity

125

syntactic case, the three-case linking system is the appropriate analysis for languages like Diyari. Before turning to Goddard's extension of the three-case analysis to some split-ergative systems, it is worth pointing out that the three-case analysis within the present approach accounts very nicely for certain additional facts of Diyari. First, according to the analysis above, the default case in Diyari is accusative, as in type 1 languages. Indeed, Diyari would be of type 1 if it were not for the "additional" ergative case. Since accusative is the default, we have at least a weak expectation that accusative should occur multiple times per clause and that accusative should be the case on adverbials in a heterogeneous "elsewhere" pattern. Both expectations are fulfilled in Diyari. First, one can get multiple accusatives per clause. In fact ditransitives have double accusative. Although this could be the result of accusative dative-accusative homophony, the present system can handle a sentence like that in the previous example: (129)

(x ERG:

IARG f

z ) ditransitive

*NOM: [ARG]

*NOM: [ARG]

—LA

NOM: [ARG] ACC:

y

[ ]

ACC:

[

7

ACC:

f

J

As with the transitive, the use of the ergative is required by the RC on the first argument, precluding (by the LDC) its use on the lower arguments. On each of these the default accusative is the only remaining alternative. A wide and heterogeneous class of adverbs in Diyari arguably gets accusative case. Austin (1981: 116) provides a list of intransitive verbs that take a complement often called a "cognate object":22 (130)

rjani [nina-ya yawada] yata-yi lsgNOM 3sg.nF.Acc-NEAR language-Ace speak-PRES 'I speak this language.'

One can easily tell that this is an intransitive verb with an extra complement, since the subject pronoun which has three distinct case forms is in the form for subjects of intransitives (S rather than A or O on Austin's analysis). On Austin's analysis the other (bracketed) NP here would consist of an accusative pronoun and an absolutive noun. Again, Goddard's (1982) across-the-board three-case analysis allows us to analyze the entire NP and all its constituents as being accusative. Note that

126 A typology of case systems accusative is the default case in this system and is appearing on cognate objects, as in many type 1 languages. The accusative also shows a default pattern in being used on a heterogeneous class of adverbs and in being the case that takes over when a more specific case is optional. First Austin (1981: 110) shows that many adjectives and noun-plus-adjective combinations may function as adverbs without any ending: (131)

kanku nawu miri kari-yi pita-ni boy-NOM 3sg.nF.NOM above climb-PRES tree-LOC 'The boy is climbing up above in the tree.'

The same word can be part of the locative NP with a slight difference in meaning: (132)

kanku nawu kari-yi pita miri-ni boy-NOM 3sg.nF.NOM climb-PRES tree above-LOC 'The boy is climbing in the top of the tree.'

Austin (1981: 110) further notes that in this latter sentence the LOC -ni "can be deleted with apparently no change in meaning." On the present analysis, this is a typical instance of a more restrictive but optional adverbial case giving way to the default case when the more restrictive option is not taken. This is strong additional evidence for the accusative as a default case in Diyari. Tentatively, we can conclude that the present system is equipped to handle three-case systems quite well. 3.5.2 Split ergativity revisited: quasi-split ergativity The three-case system can also form the basis for an analysis of split NP ergativity. Goddard (1982) proposed such an approach for split NP ergativity, and his analysis can be easily adopted in the present scheme, since three-case systems are provided for already. What Goddard proposes is to extend his "uniform" analysis of languages like Diyari with a split between nouns that show three case forms and those that show two to other languages that show an ergative-absolutive form for some noun types and a nominative-accusative distinction for others. No one noun class shows three forms - indeed all show only two - but Goddard argues that all have three cases (nominative, ergative, and accusative) with one set of nouns having syncretism of nominative and accusative, yielding a (misleading) ergative-absolutive pattern and the other class with syncretism of ergative and nominative, thus showing what merely

Three-case systems and quasi-split ergativity

127

appears to be a nominative-accusative pattern. Again, the substitution test is used to tell which case an "ambiguous" NP covertly bears. Goddard demonstrates the advantages of such an approach with data on split NP ergativity from the Yankunytjatjara dialect of the Western Desert Language, another Australian language. The pattern of marking for nominals is the following: (133)

Common nouns

Proper nouns, kin names

Pronouns, anaphors

A -ngku -lu 0 S 0 -nya 0 O 0 -nya -nya Yankunytjatjara case endings (vowel final stems) (Goddard (1982: 179)) On a conventional analysis, all but the anaphors and pronouns would be analyzed as having an ergative and an absolutive (albeit with different endings), and the anaphors and pronouns would have nominative and accusative. Goddard (1982: 179), however, proposes a three-case analysis for Yankunytjatjara despite the fact that no one class of nominals shows three distinct forms. According to Goddard's definition of (morphological) case based on interchangeability, ergative is defined as the case on the class of nominals for which a common noun with the suffix -ngku could be substituted. Similarly, accusative is the case on the class of nominals for which a pronoun suffixed with -nya could be substituted. As for nominative, Goddard notes that nominative case is the class of forms for which either an unmarked pronoun or an unmarked noun can be substituted and notes that this same class of forms appears in the context of citation and in predication. Using his definition of morphological case, Goddard defines nominative in Yankunytjatjara as "a class of forms mutually substitutable in certain syntactic environments (intransitive subject, citation, predicate nominal); nominative and accusative case (casej) are formally distinguished by one subclass of nominals, namely pronouns, and similarly nominative and ergative are formally distinguished by one subclass of nominals, namely nouns" (Goddard (1982: 180)).23 The advantages to a three-case analysis for seemingly split ergative Yankunytjatjara are similar to those for seemingly split 3case/2case Diyari. Analogous constructions receive similarly simple statement under the three-case analysis. First, for "inalienable possession," the

128

A typology of case systems

statement of the similarity between the NP denoting the possessum and the NP denoting the possessor can be directly in terms of identity of case whether accusative (134a), locative (LOC) (134b) or purposive (PURP) (134c): (134)

a. Katja-lu wangka ngayi-nya kuli-ntja wiya. son-ERG talk(Acc) lsg.Acc listen-NOML NEG '(My) son doesn't heed my words.' b. Punpun tjuta wati-ngka mara-ngka nyina-ni. fly many(NOM) man-LOC hand-LOC sit-PRES 'There are a lot of flies on the man's hand.' c. Ngayulu Yami-ku yunpa-ku ngurpa-ri-ngu. lsgNOM Yami-PURP face-PURP ignorant-iNCHO-PAST 'I forgot Yami's face.'

Similarly, in the "phrasal conjunction" of NPs in Yankunytjatjara, a three-case analysis allows one to state the rule more simply. Phrasal conjunction is the placement after one or more nominals of a pronoun of a number subsuming that of the preceding elements. The rule on Goddard's three-case analysis is that the initial words and the final pronoun must be of the same case: (135)

a. [Nyuntu 2sg(NOM)

ngali]

ya-nu.

ldu(NOM) gO-PAST

'You and I went.' b. [Kanytji-nya ngali]

ya-nu.

Kanytji(NOM) ldu(NOM) go-PAST

'Kanytji and I went.' c. Ngayulu [Kanytji-la lsg(NOM) Kanytji-LOC

tjana-la] nyina-ngi. 3pl-LOC

Sit-PAST.IMPERF

'I stayed with Kanytji and the others.' d. Wati-ngku [nyuntu-nya ngali-nya] nya-ngu. man-ERG 2sg.Acc ldu-Acc see-PAST 'The man saw you and me.' On a conventional two-case analysis, for instance, (135b) would have to be analyzed as having an absolutive NP (Kanytji-nya) and a nominative pronoun (ngali), whereas on the three-case analysis they match - both are nominative (see the table in (133)).24 If the three-case analysis of at least some split ergative languages is on the right track, the scheme developed in the present chapter can incorpo-

Three-case systems and quasi-split ergativity

129

rate these insights very easily. We have seen how the system allows for three-case systems. These three case-linkers specify the morphological form for an NP based on linking properties like argumenthood and not-being-the-lowest-argument. There is no barrier in this system to syncretism of morphological cases of the sort argued for by Goddard. We just use a three-linker analysis - one linker for each of the morphological cases proposed by Goddard. These cases show some syncretism. Note that this does not weaken the system here in the slightest. Most of the time, the morphological analysis is not controversial as far as number of cases goes. But even where the morphological analysis is not so straightforward, each of the approaches has consequences for linking. If two NPs are (morphologically) analyzed as having the same case they may involve different linkers, but if two NPs are analyzed as having different cases we cannot say they are linked by the same case. In fact if we adopt the Goddard analysis we have to use a three-linker analysis. Morphology constrains the theory of linking; morphological analysis will make a difference to the linking analysis.

4

Linker interactions

This chapter will be devoted to the interaction of various aspects of morphosyntax. First I will continue with some implications of the Elsewhere Pattern Prediction for linker interaction. The default is supposed to apply whenever nothing else can. The strongest evidence for the default status of any case is when its opportunities for application form as unnatural a class as possible. The choice is between, on the one hand, positing a very unlikely set of homophonous cases (or other linker type) which just happen to fill in all the gaps of the other linkers and, on the other, simply positing a default linker. The second type of interaction examined in this chapter will be between the valence-changing operations known as passive and antipassive and the linkers. The question is whether the system as it stands will predict the correct case even under circumstances of changed valence. Section 4.2 will show that a suppression view of the passive and the present theory of case interact to give exactly the right derived case frames. I will then extend the discussion, which has so far concentrated mostly on case (and to a limited degree on agreement), to linking by word-order position. I will present an analysis of a language that relies primarily on word-order position, English. Finally, it will be shown that the present system is equipped to handle the interaction of case and configuration in oblique and non-oblique contexts.

4.1

Other elsewhere patterns

The Elsewhere Pattern Prediction has further implications which I would like to turn to now. The nature of a default is that it applies in the complement of the places where anything else applies. So if we have a default, another syntactic case, lexical cases, and a range of adjunct cases, then the default case applies in the complement of the conjunction of the domains of all the other cases. In other words, the default's actual 130

Other elsewhere patterns

131

domain of application is determined by the nature and inventory of all the other linkers. I noted above that it would be possible though uneconomical to have separate cases for all the uses of the default. So in Icelandic, which is type 1, we could of course have a range of adverbial accusatives. If we assume that the accusative is the default, we can explain the fact that accusative marks adverbials but nominative does not, an otherwise unexplained fact. Further, we can explain why accusative but not nominative is repeatable and why accusative adverbials correlate cross-linguistically with accusatives that can be used with more than one co-argument. And these predictions generalize to the defaults of languages of the other types. But there is another compelling reason why the accusative of type 1, the nominative of type 2, the absolutive of type 3, and the ergative of type 4 should be regarded as the default. This is because in languages with a very unusual domain for their other cases, the default (as I am analyzing it) applies in exactly the complement of this unconventional distribution. The first of these is optionality. If a wide range of adjunct cases are optional, the accusative is predicted to apply just in case the more restrictive optional case does not. The second area where the default analysis is desirable is where the other, i.e. [ARG], case has an unusual restriction causing it to fail to apply in certain contexts. Here too, we predict that the accusative or other default will apply. The fact that it does is striking confirmation for the restrictiveness-based approach. 4.1.1

Optional adjunct cases

Optionality of an adjunct case is the perfect testing ground for the restrictiveness-based approach. The theory predicts that when any linker X is optional and the no-X option is taken, then the next most restrictive linker must apply. Applied to adjuncts the predictions are particularly clear. For adjuncts the only linkers that could apply are the special very restrictive adjunct case and the default, e.g. accusative in type 1. Recall that the [ARG] linker cannot apply, because by definition it is restricted to arguments and so cannot even pass the AC for adjuncts. Other cases that are restricted to arguments (argument category cases, idiosyncratic cases with verbs) similarly are not applicable to adjuncts. This means when an adjunct case is optional the default is predicted to apply.1 I now present examples of this. The alternation of accusative with prepositions on adjuncts is quite common in Greek. Thus for the accusative of respect, repeated here in

132 Linker interactions (1), one can have the preposition eis 'to,' kata 'according to,' pros 'for' and dative alone: (1)

(2)

(3)

deinoi de makhen terrible(NOM.pl) PTCL battle(Acc) 'terrible in battle'

{KQsch.Pers.21)

pr6tos eis eupsuchian first regarding good-courage(ACc) 'first in good courage'

(Aesch.Pers.326)

en mekei kai platei kai bathei in height(DAT) and breadth(DAT) and depth(DAT) 'in (with respect to) height and breadth and width'

The accusative is simply used when none of the more specific prepositions or cases is not. As noted above, the accusative in Greek can also be used for extent of time adverbials. Many other prepositions can express extent of time (Smyth (1956: §1587)). Two prepositions especially vary with the accusative: para 'for' with accusative (Smyth (1956: §1692.3.b)) and did 'through' with genitive (Smyth (1956: §1685.1.b)): (4)

(5)

para panta ton chronon throughout all(ACc) the(ACc) time(Acc) 'throughout the whole time' dia nuktos through night(GEN) 'through the night'

(D.5.2)

(X.^.4.6.22)

Both these prepositions emphasize the uninterrupted nature of the action or state denoted by the verb. However, truth conditionally, they probably contribute in the same way as the accusative; accusative NPs of duration also imply uninterrupted action or state. Not surprisingly, the accusative is the unmarked case. The default accusative should not contribute to the meaning, other than allowing the NP it attaches to to be linked. The accusative does not constrain the interpretation itself. Note too that the prepositions that alternate with the accusative in the adverbial constructions do not all take accusative, e.g. did 'through' in the last

Other elsewhere patterns

133

example. Therefore we cannot be dealing with some kind of surface preposition deletion. Genitive of duration adverbials is not found. Another place where optionality reveals the default nature of accusative is with time-at-which adverbials. In Greek the dative is normally used for time-at-which adverbials. But in the few cases where the dative does not apply, we get accusative (Smyth (1956: §1583)). The dative in (6) alternates with the accusative in (7) (Note that (6) contains both an accusative of duration and a dative of time-at-which): (6)

tauten men ten hemeran autou emeinan that(Acc) PTCL the(ACc) day(Acc) there waited(3pl) tei de husteraiai . . . the(DAT) PTCL next(DAT) 'throughout that day they waited there, but on the following (day) . . . ' (X.H. 1.1.14)

(7)

tende ten hemeran that(ACc) the(Acc) day(Acc) 'but on that day'

(Aeschines.3.7)

Once again, a more specific adjunct case is optional, and exactly when it does not apply, the accusative appears, providing more evidence for accusative's default status in Greek. In Icelandic, the evidence from optionality is more limited. It was noted that there are a range of non-argument accusatives. In some cases, a preposition is optional with an adjunct. In this case the accusative is the other option: (8)

Eg

var J)ar (um) tvo daga. I(NOM) was there for two(Acc) days(ACc) 'I was there (for) two days.'

Similar alternations occur in German and Ukrainian (Shevelov (1963: 172)). As indicated in the gloss for the last example, a similar phenomenon occurs in English. In the next chapter I will argue that English is type 1 and that its default linker is position within the VP. The morphosyntactic feature [VP] will be assigned to NPs in the same way as accusative is in Icelandic: (9)

VP: [

]

134 Linker interactions This linker can be thought of as the equivalent of the type 1 accusative but with a different realization. Clearly adverbials can be linked by location in the VP, since they do not receive a morphological case nor do they occur with a preposition or agreement marker. If they are linked it must be by position, and by hypothesis the within-VP position: (10)

John stayed there three hours.

(11)

John drove this path.

Note that, prepositions leading to at least roughly the same meaning can be used as well: (12)

John stayed there for three hours.

(13)

John drove along this path.

As with the case-marking languages, this can be captured by the optionality of the adverbial case: (14)

Adverb ( NP ) (for: [ADJUNCT])

VP:[

]

If the /br-option is taken, the preposition applies. If the no-preposition option is taken, then the default VP-linker applies alone leading to the bare NPs in (10-11). One main advantage of the default approach over multiplying homophonous adjunct cases is that the distribution of accusative is no longer accidental. The fact that the accusative applies just when optional cases, especially adjunct cases, do not can hardly be coincidence. The recurring alternation of these cases with the default is evidence for the restrictiveness-based approach. There is also a similarity between this type of optionality and the case alternations of Chapter 2. There it was argued that relating two cases directly, e.g. lexical accusative and lexical dative, was too powerful, because it leads one to expect all sorts of non-occurring alternations, e.g. anti-DS. So too with the adjunct cases, if the adjunct cases alternate with a raft of separate adjunct accusatives rather than with one default, we should equally expect alternation of adjunct cases with other cases, like dative. These alternations do not occur systematically. The fact that the default is the case alternating on adjuncts is further confirmation of the default approach.

Other elsewhere patterns

135

4.1.2 Nama Hottentot nominal suffixation Optionality is not the only way for the conjunction of the distributions of the non-default cases to be unusual. The argument cases can have restrictions which force the default to appear in otherwise unexpected places. Nama Hottentot (Hagman (1977), Sadock (1991: 147ff.)) is one such language. Its [ARG] case has very unusual additional restrictions on its appearance. On the present theory, the default linker is predicted to apply in just those places where the [ARG] linker would apply but for its unusual restriction. Nama presents a striking confirming instance. The inventory of linkers in Nama consists of agreement, position before a "declarative particle" ke, a "subordinative suffix" -a, and various postpositions. Most interestingly, the declarative particle seemingly shows a "nominative" distribution but only for preposed NPs in declarative clauses. Postposed declarative subjects, subjects of interrogative and imperative/hortative clauses, objects, time adverbials and some other NPs all get the subordinative suffix -a. After reviewing linking in Nama, I will show that this pattern is exactly what we expect if Nama is type 1 and the declarative suffix is limited to declarative clauses. In declarative clauses, we can say at first approximation that the highest argument is followed by the particle ke and other argument(s) get the suffix -a. Additionally, Nama shows agreement for person, gender, and number with the highest argument. So, in an intransitive, we get just one NP preceding ke and person/gender/number (PGN) Agreement: (15)

'ao-p

ke

//naapa ke

man-iNFL(3Msg) DECL

there

maa

PAST stand

T h e man stood there.' The morpheme glossed INFL is a clitic and agrees with the highest nonobliquely linked argument, i.e. with the subject. In a transitive, the subject is still the source of sentential agreement, and the other NP appears with the suffix -a: (16)

'ao-p

ke

man-iNFL(3Msg)

DECL dog-3Msg-suB PAST see

'ari-p-a

ke

muu

'The man saw the dog.' As Sadock notes, the -p is both a person/gender/number suffix, which all nouns require, and the sentence-level inflection, which every sentence requires. Both the PGN suffix and the sentential clitic have the same complex allomorphy. The -p on the subject NP seems to satisfy both

136 Linker interactions the requirement for the subject NP and the INFL requirement for S just if the subject NP is before the particle he, as in (15-16). If the subject is not located here, we get separate -ps. For example, some other constituent can be topicalized, e.g. a PP (17) or the verb (18): (17)

!'aa-s !'oa-p ke nee 'ao-p-a ra !uu town-3Fsg to-iNFL(3Msg) DECL this man-3Msg-suB PROG go 'This man is going to the town.'

(18)

!uu-p ke ra nee 'ao-p-a !'aa-s !'oa go-iNFL(3Msg) DECL PROG this man-3Msg-suB town-3Fsg to This man is going to the town.'

In each of these the sentential clitic still agrees with the subject (man), but is attached to the fronted constituent. The subject NP comes later in the sentence and receives the subordinating suffix -a, in addition to the PGN suffix. In the following, the subject seems to be postposed, and there is a conjunction at the front of the sentence. In such sentences, the agreement clitic attaches to the conjunction: (19)

(20)

tsii-p ke nee 'ao-p-a ke !uu and-iNFL(3Msg) DECL this man-3Msg-suB PAST go 'and this man went' 'oo-s

ke

//'u-s-a

ke

til

then-iNFL(3Fsg) DECL this-3Fsg-suB PAST ask

'then she asked Note in (20) that the subject is feminine (s), and the sentential clitic is also feminine. One might be tempted to see this as the operation of some surface filter which disallows the subordinative suffix on the NP preceding the declarative particle he. This cannot be so, however, because if the object is fronted, then the object must have -a even though it precedes the ke: (21)

hai-p-a-p ke 'ao-p-a stick-3Msg-suB-iNFL(3Msg) DECL man-3Msg-suB //'u-p-a ke maa this-3Msg-suB PAST give 'The man gave him a stick.'

Other elsewhere patterns

137

The first NP is the object, and it has its own PGN suffix followed by the suffix -a, followed in turn by the sentential clitic. Contrast this with the situation in (16), where the subject is in pre-fe position, and has only one PGN suffix, which is of course identical to what the sentential clitic would be. This last example (21) also illustrates the uses of -a, to which I now turn. Nama does not have a special case for goals. Rather they are marked with the suffix -a and, according to Hagman (1977), precede the direct object in non-topicalized sentences: (22)

'ao-p ke //'ii-p-a hai-p-a ke maa man-iNFL(3Msg) DECL this-3Msg-suB stick-3Msg-suB PAST give 'The man gave him a stick.'

(23)

'ao-p ke tara-s-a pere-p-a man-iNFL(3Msg) DECL woman-3Fsg-suB bread-3Msg-suB ke maa PAST give 'The man gave the woman bread.'

In the present theory we need linkers like the following, among others: (24)

_V': [ARG|GOAL]

The word-order linker will refer to the syntactic argument category GOAL, just as the dative does. For now, it is just necessary to assume that linking by a word-order position (like (24)) is compatible with linking by -a. I will analyze the suffix -a as the default linker in Nama. Before turning back to linking on declarative subjects, it is very striking that - a fulfills all the expectations of a default case. It is repeatable, as we already saw in the ditransitives in (21-23). It also links the subject under certain circumstances, about which more shortly. But note that it fulfills the Elsewhere Pattern Prediction in marking a wide range of NPs which do not have special cases. Hagman (1977) documents a wide range of uses of -a: equational predicate NP-a (p. 56), direct object (p. 76), indirect object (p. 76), temporal adverb NP-a (p. 105), with certain postpositions (p. 101), postposed subject (p. 109), interrogative subject (p. 139f.), imperative-hortative subject (p. 145). We have already seen the direct and indirect object and postposed subject uses. Time adverbials get the suffix -a:

138 Linker interactions (25)

sii-kxrh ke am tsee-ra PRO-iNFL(lMdu) DECL two day-3F-Cdu(suB) //naapa ke haa there PAST stay 'We stayed there two days.'

The adverbial tseerd 'for two days' has the subordinative suffix.2 The subject of non-declarative clauses also gets the subordinative suffix. Thus in interrogative sentences there is of course no declarative particle ke, and the subject gets the suffix -a:3 (26)

//'n-p-a //an'-e ke PRO-iNFL(3Msg)-suB meat-3Msg(suB) PAST eat 'Did he eat meat?'

Sadock (1991) argues that the PGN suffix on the subject is also the sentential clitic. The use with postpositions is very interesting, because it illustrates the optionality and Elsewhere behavior of the last section. There it was maintained that the fact that what we are analyzing as the default case applies just in the event some optional more restrictive case does not apply is strong evidence for its status as default. This phenomenon occurs with Nama -a as well. According to Hagman (1977), most postpositions require a bare NP (or a locative demonstrative adverb, which need not concern us). In the present theory, the prepositions themselves link their NP. There are two postpositions of motion, however, which combine with an NP which must be followed by -a (NP-a + Postposition): xuu 'from, away from,' and 'MM 'along, following.' These postpositions may be said not to link their NP. But, most interestingly, there is a third postposition of motion,'^ 'to, toward,' which can take an NP with or without -a: (27) v '

., , , sn-kxm

. ke

j/'ae//am-s-a I „ ,,^ , > , < ' //, > ! oa ke !nan [/ae//am-s J f Windhoek-3Fsg-suB I , /1A,, x PRO-iNFL(lMdu) . „ , > to PAST travel v } DECIJ |Windhoek-3Fsg J 'We travelled to Windhoek.'

Note too that (17) above is an example of this postposition with an -a-less NP. This is exactly the same situation as with the optional adjunct linkers of the last section. The analysis would be identical here: whenever the

Other elsewhere patterns

139

option of not linking by the more restrictive linker is taken, the default must apply. Otherwise we would have an unlinked NP. The following five facts need to be explained: (28)

a. The clitic always occurs suffixed to the word before ke. b. If the subject occurs before ke, it has one affix that seems to do double duty as PGN suffix and sentential clitic. (This is the analysis of Sadock (1991).) c. If the subject occurs before ke, it cannot bear the suffix -a. d. The clitic always agrees with the subject (overt or covert). e. If there is no ke, the subject is marked with -a.

Furthermore, the unusual distribution of the subordinative suffix -a needs to be explained. As I indicated above, the analysis of Nama on the present theory should involve type 1 settings for the Limitation Parameter and the Preference Parameter. This will make -a the default and like the accusative in Icelandic and Greek. The one unusual feature of Nama linking is the position before ke, which is a nominative-like linker but only in declarative clauses: (29)

a. _ke:

[ARG] + DECL

b.-a:[

]

The pre-fce positional linker is limited to verbs that are declarative. I assume that verbs are assigned [ + /-decl], [ + /-inter(rogative)], etc. freely, and that positive values of these features are incompatible, because a clause cannot be simultaneously declarative and interrogative or imperative. Feature geometry will not matter here. The linkings for simple intransitive and transitive declarative sentences are straightforward and are just like those for the corresponding sentences in type 1 languages. An intransitive like (15) will get the [ARG] linker, which in Nama declarative clauses is the pre-£e position: (30)

m a a [ + decl] ( x ) 'stand' _ke: [ARG] + DECL

-a:[ ] VP:[ ] There is only one argument, so the LDC is not relevant. The RC dictates use of the more restrictive linker, pre-A;e position. Since it is incompatible

140 Linker interactions with linking by -a or the VP-internal position, the NP shows up before he with no subordinative suffix. The transitive (16) is likewise familiar: (31)

muu [ + decl]

X

_ke:[ARG] + DECL

VP:[ ]

y ) 'see' *_ke: [ARG] + DECL

VP: [ ]

The LDC rules out the use of the pre-fce linker on the lower argument, because that linker is the most restrictive linker of its type applicable to the higher argument. The pre-A;e linker does link the higher argument by the RC. The second argument must get the default. The second argument is also linked by the default word-order linker, but the first one cannot be, since the more restrictive pre-A;e position linker has priority by the RC and the NP cannot be in two places. The only thing on the present account that makes Nama unusual as a type 1 language is the restriction of its [ARG] linker to declarative clauses. The restriction follows from the fact that the [ARG] linker is a declarative particle. On this analysis of Nama, we predict that in all instances where this [ARG] linker cannot be used, the default subordinative suffix will be. This is exactly correct. The linking for non-declarative clauses can be exemplified by the structure for the interrogative in (26) above: (32)

^'uu[ + int](x -a: [ ] VP: [ ]

y)'eat' -a:[ ] VP: [ ]

The [ARG] linker is not available, because this verb is not insertable in a declarative context, captured here by specifying the verb [ + inter(rogative)]. The only linkers applicable to the subject and the object are the -a marker and positional linkers, which apply. A note is in order on topicalization and other "Wh-Movement" processes in relation to linking. Kiparsky (1987) takes linking to happen at a level after the lexicon and before Wh-Movement. Since the present theory of case is embedded in a unification-based approach close to Generalized Phrase Structure Grammar (GPSG) and Head-Driven Phrase Structure Grammar (HPSG), I assume too that linking happens "before" topicalization, but the meaning of "before" is different. I will assume that topicalization happens "after" linking in the sense that a VP with an NP missing (VP/NP) satisfies the requirement that the NP be linked by

Other elsewhere patterns

141

VP-internal position. In other words, the feature [VP] added to an argument that corresponds to a gap is well-formed, whereas a [VP] feature added to an argument that corresponds neither to an NP within the VP nor to a gap is ill-formed. So in the last example, the subject is linked by the VP-internal position linker but the realization of this linking is a VP/ NP (which is satisfied in turn by the presence of the topicalized NP) rather than an NP dominated by the VP as is true of the object. Topicalization of objects works similarly. Another notable fact about Nama linking was that postposed subjects were linked differently even in declarative clauses. This might seem to contradict what I, following Kiparsky (1987), just assumed, that linking is not affected by "Wh-Movement" type processes. This contradiction is only apparent. The difficulty here relates to the unusual fact noted above that the sentential subject clitic is only allowed if the word it attaches to is not the subject itself. Recall from (16) above that the PGN marker in non-postposed subject sentences seems to be doing "double duty," and this is the analysis of Sadock (1991). Another way to capture these facts is to assume that the clitic, when it appears, is the subject, and so is in complementary distribution with a subject NP. The subject NP has only one PGN suffix which is not a clitic: (33) COMP

I have left off optional positions for adjunct NPs. I assume that adjunct NPs may be generated below any node from S on down. Either an NP or a clitic may be generated in the NP of S position. If it is a clitic, then it must attach leftwards, as Sadock notes. Thus in declarative sentences the clitic is the subject and is linked. The "postposed subject" is then not the subject but rather an adjunct coindexed with the true subject. Evidence for this comes from the fact that the postposed subject is optional:

142 Linker interactions (34)

'/lii tsee-ts ke nil ^'oa !aro-p one day-iNFL(2Msg) DECL FUT go.out forest-3Msg 'One day you will go out into the forest.'

!naa. into

Other NPs in the sentence, e.g. the object, seem not to be optional. Thus the special status of the postposed "subject" needs an account, which it receives from its treatment as an adjunct. The linking for this sentence is the following: (35)

! ^ ' o a [ + decl] ( x ) ' g o ' _ke:

[ARGJ

+ decl -a:[ ] VP:[ ] The sole argument is linked by the pre-ke position as in the normal declarative intransitive in (15), the only difference being that what the NP the argument corresponds to is a clitic in (34) rather than a full NP. Now consider the N P I am analyzing as an adjunct coreferential with the subject. The linking for a sentence containing such an N P is as follows, which is the linking for (21): (36)

maa [ + decl] ( x _ke:

y|GOAL *_ke: [ARG] + decl + decl -a:[ ] *a:[ ] VP:[ ] -V':[ ] VP:f ] [ARG]

z ) 'give' ( NP ) *_ke: [ARG] -a: [ ] +decl VP: [ ] -a: f ] VP:[ ]

This is a ditransitive declarative with a "postposed subject." By the LDC only the highest argument gets linked by the pre-fce position, which it must by the RC. The subject in this case is a clitic. The linking for the coreferential adjunct is separate and is given at the right. On the present theory, the only case that can apply to the adjunct is the subordinative suffix -0, which does apply. The theory correctly predicts that the adjunct must get -a, since, being an adjunct with no special case, that is the only case it is eligible for. This sentence is also interesting because it illustrates the linking of a ditransitive. All arguments not linked by pre-£e position get -#, the default case. Additionally, the goal gets linked by the position adjacent to V' (at the front of the VP). The other object must appear in

Other elsewhere patterns 143 the VP as in (22) or be topicalized, which it is in (21). The VP-internal position also applies vacuously to the goal. To summarize, the present analysis accounts for all the facts in (28) about the subject. First, the clitic must occur suffixed to the word before ke because it is linked by left adjacency to ke. It is a clitic, however, and so phonologically it must attach to the preceding word. The "double duty" phenomenon follows from the choice of generating a clitic pronoun or a full NP as the NP linked to the highest argument. If there is a full NP before ke, the PGN suffix is just a PGN suffix. There is no additional clitic, because the NP is standing where the clitic would be. The subject, if it stands before ke, is linked by ke. This linker is more restrictive than the default linker, the suffix -a, and the two are incompatible by hypothesis. By the RC, linking by means of the pre-£e position will block linking by means of -a, and so a subject NP to the left of ke cannot have -a. The clitic of course appears to agree with the subject, because on this analysis it is the subject. Any adjunct NP coindexed with the subject clitic must agree with it. Finally, in clauses without ke, i.e. all non-declarative clauses, the subject of course gets -a. This is because in such clauses there can be no linking by pre-ke position, and so the next most restrictive linker must apply, i.e. the default -a. The default status of -a is very dramatic in Nama, and the fact that the theory forces us to analyze this subordinating suffix as the default case provides an elegant way to capture the wide and disparate class of NPs that receive -a. Without default -a, the distribution of -a would be accidental, which, in the case of Nama, is very unlikely. 4.1.3 Defaults and grammatical relations Another result of of the present theory's use of a default is that it reduces the number of notions needed to capture morphosyntactic regularities like case marking. Some work in RG has been devoted to the question of what range of grammatical relations need to be referred to by rules of case marking (case marking is effected by rule in RG). The use of final grammatical relations for case marking is the least controversial and has its antecedents in traditional grammar and analogs in other theories. For example, the RG rule assigning nominative to Is is a recasting of the traditional observation that nominative is the subject case. In other contemporary theories, nominative for Is is paralleled by case-marking rules such as "assign Nominative to SPEC of IP" in GB and the rule assigning nominative to SUBJ in much of LFG (but see Zaenen, Maling, and

144 Linker interactions Thrainsson (1985) for a different approach). Direct Linking seeks to minimize even this type of statement, as I have noted above. But RG posits case-marking rules that refer to grammatical relations that are unique to that theory. Two such notions are Initial 1, 2, etc., and Acting 1, 2, etc. An Initial 1 in that framework is a nominal heading an arc which bears the relation 1 at the initial stratum. Thus, in a passive the final subject is a 2 which advances to 1. Acting 2 is more inclusive. Nominals that head arcs that bear the 2 relation at any stratum, but do not subsequently bear a term relation, are Acting 2s. For arguments that linking does not refer to initial grammatical relations, see Kiparsky (1987). I would now like to point out how the facts adduced to support Acting 2 are also fully covered by the present theory which contains a minimum of primitive functional information. As an example of a language requiring the notion Acting 2, Perlmutter (1982: 312f.) cites Latin as typical of languages requiring a rule assigning accusative (or some other morphosyntactic device) to Acting 2s. The construction cited is very similar to the double accusative ditransitives in Greek (German also has a similar construction with three or four verbs): (37)

Magister pueros grammaticam docet. teacher(NOM) boys(ACc) grammar(ACc) teaches 'The teacher teaches the boys grammar.'

The RG analysis proposed for (37) is that the semantic goal starts out as a 3 and advances to 2, forcing the 2 on the previous stratum into chomage: (38)

docere

| | pueros magister grammaticam

Case marking for accusative could be stated just in terms of nominal heading a 2 arc, since both accusative NPs head a 2 arc at some stratum.

Other elsewhere patterns

145

The crucial case for deciding between a nominal that heads a 2 arc and Acting 2 is where a nominal heads a 2 arc (2 at some stratum) and there is a later stratum at which it bears another term besides 2. Such a nominal of course heads a 2 arc but it is not an Acting 2, because an Acting 2 cannot bear another term relation on a stratum later than the last one where it bears the 2 relation. Perlmutter takes the passive of (37) as evidence for the need for Acting 2: (39)

Puerl grammaticam a magistro docentur. boys(NOM) grammar(ACc) by teacher(ABL) are-taught T h e boys are taught grammar by the teacher.'

The relational diagram is the same with the addition of another stratum for the passivization. The 2 created from the previous 3 now advances to 1. Thus puerl 'boys' heads a 2 arc but is not an Acting 2. And it is not accusative. The other nominal is an Acting 2 and is accusative: (38)

docere

| | puerl magistro grammaticam

As noted before, such sentences are handled straightforwardly in the present framework. Just as in Greek, a small number of ditransitives take three direct arguments rather than the more usual two direct arguments and an argument belonging to the syntactic argument category GOAL. With direct arguments, the dative is not applicable: (41)

docere ( x

y

NOM: [ARG] ACC:

[ ]

z ) 'teach'

*NOM: [ARG] ACC:

f

]

*NOM: [ARG] ACC:

[

]

I assume that the nominative and accusative are specified as in Greek, and Latin indeed shows type 1 characteristics (see Kiihner (1912-14) and Gildersleeve (1894) for details on accusative adverbials and cognate accusatives). If so, the LDC prohibits nominative on all but the first argument, where it must apply by the RC. The lower two arguments must get

146 Linker interactions the next most restrictive case, which is accusative. In the passive, everything is the same, except for the removal of the highest argument from the argument list: (42)

doceri (xA

y

z ) 'teach'

NOM: [ARG] ACC:

[

]

*NOM: [ARG] -ACC:

f

]

The second highest argument in the active corresponds to the highest argument in the passive. Since there is no higher argument, the second argument (the semantic goal) gets the nominative. The other remaining argument has no choice left but accusative. The accusative as default is in a position to capture the facts that otherwise would require an additional notion of Acting 2 and 2 itself. The purpose of this section has been twofold. First, we illustrated that an unusual distribution of non-default cases leads to an unusual distribution of default marking. The more unusual this distribution is, the less likely that the case I am analyzing to be the default just happens to have this unusual distribution as a result of the conjunction of specific rules. The more unusual the boundary between non-default and default, the more unlikely that they mesh out of coincidence. I illustrated this prediction from adopting a default case with optional adjunct cases, the declarative particle linking in Nama, and the data for Acting 2, which show the default applying in what is far from being a natural class of environments. The default approach allows a unified account and gains further support. 4.2

Passive and antipassive

The above proposals make interesting predictions about the interaction of case and the passive (and antipassive). This section will start with a review of the interaction of passivization, which, following recent practice, I will take to be the suppression of the highest argument. I will then show how the two parameters proposed in the last chapter provide an account of passive in each of the types in the typology of two-case systems. This approach automatically leads to an account of antipassivization in types 3 and 4. Passive is usually viewed as consisting of two main aspects (Chomsky (1981)): it absorbs the highest theta role and absorbs the structural Case assigned by the verb. The correlation between lack of ability to assign

Passive and ant (passive 147

Case and the absence of an external argument is known as Burzio's Generalization. The correlation of non-Case assignment and lack of an external argument has remained a stipulative generalization, and attempts to derive it from other principles or to derive lack of Case assignment from lack of external argument or vice versa have been largely unsuccessful. The initially most promising route is to derive Case absorption from external argument suppression. One of the main obstacles to doing this is the existence of a class of languages like Ukrainian with non-promotional passives, where the external argument is suppressed but the internal argument still gets accusative rather than the expected nominative from INFL. I will show that, assuming the case theory above, we can derive Burzio's Generalization and begin to explain the distribution of Ukrainian-type passives. In this section I will focus on the interaction of passive and the argument linkers. I will assume a suppression view of the passive, meaning that the one aspect of the passive operation is the removal of the highest argument of an active verb from the argument list. The next highest argument in the active automatically becomes the highest in the passive. The suppressed argument can in some sense be expressed by a by-phrase which will be interpreted in the same way as the highest argument of the active. This by-phrase is sometimes considered an optional argument and sometimes an adjunct coindexed with the suppressed top argument. It is not clear that either of these ways of dealing with the by-phrase is universally valid. For present purposes, however, I will assume it is some sort of adjunct. Some languages do not allow the by-phrase to appear at all, some languages have a range of prepositions for the by-phrase, and some, like English, have only one preposition. For English I will tentatively treat the by-phrase as part of the passive operation since by does not have a consistent semantics, unlike other adjunct prepositions which do. The passive sentence The tree was hit by the car involves at most an actor phrase, not an agent phrase. The essential assumption for the interaction of passive and antipassive with the four main types of two-case system is that the highest or lowest argument of the active no longer be the highest or lowest argument of the passive or antipassive, respectively. 4.2.1

Passive in type 1 and type 2 languages

Type 1 and type 2 languages differ in the status of what is labeled accusative case. In type 1 languages it is the default case, but in type 2 languages it is a more restrictive case with nominative as the default:

148 (43)

Linker interactions Type 1: NOM: [ARG] ACC: [

Type 2: NOM: [ ]

]

ACC: [ARG|-HA]

Since the two types of accusative differ, they might be expected to interact with the passive differently. I now show that for each type, the case absorption property of the passive follows straightforwardly. As touched on briefly in Chapter 2, the accusative in type 1 languages is not actually absorbed but is prevented from applying by the nominative. In a passive sentence like the following, the accusative does not appear: (44)

Eg

var

I(NOM) was

laminn (afhonum). hit

by him(DAT)

'I was hit (by him).' The top argument in the lexical entry of the active is suppressed: (45)

lemja-PAss ( xA y )

The cases apply to this derived structure: (46)

lemja-PAss ( xA y ) NOM ACC

Since this is a type 1 language the NOM ([ARG]) is more restrictive than the ACC ([ ]) and so takes priority. The passive subject is nominative not accusative. In a type 2 language such as Japanese, the accusative should take priority over the nominative. However, the accusative, specified [-HA], will not be applicable to the highest argument of the passivized verb. Consider the active-passive pair in (47) from Kuno (1973): (47)

a. John ga Mary o korosi-ta. John NOM Mary ACC kill-PAST 'John killed Mary.' b. Mary ga John ni koros-(s)are-ta. Mary NOM John by kill-PASs-PAST 'Mary was killed by John.'

The linking for the passive (47b) will start with the lexical entry for koros'kill': (48)

koros- ( x y )

Passive and antipassive

149

In the passive the highest argument is suppressed: (49)

koros-(s)are ( xA y )

When the linkers apply, only nominative not accusative can apply, as in (50): (50)

koros-(s)are ( xA y ) NOM

The accusative, bearing [-HA], is not even applicable. The subject will be nominative. I will continue with the account of passive in the next chapter, where ditransitive passives will be considered. Thus Burzio's Generalization follows from general principles and the parameters of the last chapter in the two types of languages. In type 1 the accusative does not show up in the passive, because it is preempted by the more restrictive nominative. In type 2 languages the accusative (and the dative if it is [—HA]) does not show up in the passive, since it may not apply to the highest argument.

4.2.2 Antipassive in type 3 and type 4 languages The antipassive works in a way very similar to the passive, only it is affixmediated suppression of the lowest rather than the highest argument. If the antipassive is indeed like a reverse passive, the antipassive should work the same in types 3 and 4 as the passive in types 1 and 2. This is because, by the Preference Parameter, these languages bear the same relationship to type 1 and 2 as the antipassive does to the passive. The antipassive should work in type 4 the same as the passive in type 1, but with the ergative left over. The antipassive in type 3 should work the same as the passive in type 2 with the ergative not applicable to the passive's lowest argument since the ergative is [-LA]. Kala Lagaw Langgus is one of the first languages to have been analyzed as having antipassivization (Postal (1977)). Kala Lagaw Langgus is a split-ergative language, where the ergative is limited to common nouns in the singular and dual, and accusative is limited to proper nouns and non-singular pronouns (plural common nouns are never inflected). Klokeid (1978) presents an analysis where absolutive is the default case. In our terms, Kala Lagaw Langgus must be partially type 3 with respect to the ergative (on split ergativity, see §3.5). In other words Kala Lagaw Langgus contains the following cases applicable to a transitive:

150 Linker interactions (51)

a. b.

ERG: [ARG|— LA] ABS:

(limited to certain noun classes)

[ ]

Evidence that the absolutive is the default comes from its Elsewhere distribution. As Klokeid notes, it applies whenever the other cases do not. The fact that a clause can contain multiple absolutives is a good indication in the present theory that it is indeed the default: (52)

Moegikazin ngoena gasamdhin a Zon child(ERG) me(Acc) caught and John(ABs) kosar garaka gasamawmadhin. two youth(ABs) caught The child caught me, and John caught two youths.' (Klokeid (1978: 594))

In the second clause of (52), both nominals are absolutive (the zero case). This follows on Klokeid's account and on ours, because the first nominal is not in the ergative noun class, so it cannot get ergative. It is the highest argument, so if Kala Lagaw Langgus accusative is like the ergative but [-HA], the accusative cannot apply here. The other nominal is in the ergative class, but as the lowest argument here it cannot get ergative. For both nominals, this leaves only the default absolutive. The fact that we can get more than one absolutive per clause means that, in the present analysis, the absolutive is to be analyzed as the default. The antipassive suppresses the lowest argument. In Kala Lagaw Langgus, this will mean that the argument that would be ergative in the active cannot be ergative in the passive because that argument is now the lowest: (53)

a. Ngath (urapun) puuyi pathan. I(ERG) one tree(ABs) chopped-down 'I chopped down (one) tree.' b. Ngay (matha mura) puyin pathi. I(ABS) all tree(ERG) chopped-down 'I chopped down (all) the tree(s).' (Klokeid (1978: 590))

As in many other languages, the case for the by-phrase is homophonous with the ergative. The linking for the active and the antipassive are as follows:

Passive and an tipassive

(54)

a. pathan ( x

y ) 'chopped down'

ERG: [ARG/—LA] ABS:

151

ABS: [

]

[ ] yA ) 'chopped down'

b. pathi ( x ABS: [

]

If the second nominal in (53a) belonged to a different noun class, the accusative would be applicable. As I indicated at the end of the last chapter, the two ways of building in such restrictions are to load argument structure with noun class features and any other features relevant to case marking or to have the constraints (AC, LDC, RC) be constraints on lexical insertion. Probably the latter is better, but exactly how the noun class restrictions are implemented does not affect the main point here which is the effect of the antipassive on type 3 linking. In the active (53a) (the structure for which is given in (54a)), the top argument, provided it belongs to a given set of nominal classes, is eligible by the AC for ergative or absolutive. The argument is the highest, so there can be no problem with the LDC (no argument is higher than it), and ergative applies by the RC. The other argument will get accusative or default absolutive depending on noun class. In the antipassive (53b), represented in (54b), the highest argument of the active is also the lowest argument, making ergative inapplicable, since ergative cannot appear on lowest arguments. This leaves the default absolutive, which is correctly predicted to apply. Note that we get a sort of reverse of Burzio's Generalization in type 3. In type 2 (Japanese) the accusative cannot apply to the logical object in the passive because the logical object is now the highest argument, which the type 2 accusative cannot link (being [-HA]). In type 3, the ergative is precluded in the antipassive for a similar reason. When the antipassive suppresses the lowest argument, the argument that is the highest in the active is now also the lowest and so cannot get ergative. Default absolutive must apply instead. Thus the correlation of antipassive and the inability of ergative to appear follows directly. For a type 4 antipassive, a possible instance comes from Kabardian, the Northwest Caucasian language discussed in §3.3.4 above. Davies (1984: 333) in his discussion of antipassivization cites a construction in Kabardian, giving an example from Catford (1976) originally due to Jakovlev:

152 Linker interactions (55)

a. rie-m dog-ERG 'The dog b. fie-r dog-ABS T h e dog

qwipsrie-r je-dzaq'e bone-ABS bite is biting the bone.' qwipsfre-m je-w-dzaq'e bone-ERG bite is biting the bone.'

(Catford (1976: 45))

Catford cites Jakovlev as claiming that the first construction, the ergative one, implies that the dog bit right through the bone, but the "nominative construction" in (55b) implies that the dog is only gnawing at the bone. It is hard to tell whether this is an antipassive. Davies does take this to be an antipassive and glosses the -w- in \hs je-w-dzaq'e 'bite' as a morphological indication of the antipassive. There is quite a large literature on the "nominative construction" and what its status is. I will assume for the sake of illustration that this is an antipassive but without any certainty that this is the best way of looking at Kabardian sentences like (55b). The antipassive in Kabardian is predicted to work like the passive in type 1 only in reverse. Compare the Unkings for the active and the passive: (56)

a. je-dzaq'e ( x

y ) 'bite'

ABS: [ARG] ERG:

[

]

ABS: [ARG]

[ ] y ) 'bite' ERG: A

b. je-w-dzaq'e ( x ABS: [ARG] ERG:

[ ]

Type 4 languages have the high value for the Preference Parameter, meaning that the LDC will question the linking of higher arguments relative to the possibilities for the lower arguments. This means that the [ARG] case has an absolutive rather than a nominative distribution. If the antipassive suppresses the lowest argument, then the higher argument is no longer prevented from getting the absolutive by the LDC, since there is no lower argument to compare its linking to. This means that where the ergative had to apply in the active, the absolutive now becomes obligatory by the RC. The ergative absorption effect follows from the Preference Parameter and the LDC, just as the accusative absorption effect followed in type 1.

Passive and antipassive

153

4.2.3 Non-promotional passives At the beginning of this section, I mentioned the possibility in some languages of "non-promotional" passives. A well-known instance is Ukrainian from Sobin (1985), exemplified by the Ukrainian sentence in (57): (57)

Cerkvu bulo zbudovano church + fern, ACC be + past + neut build + part + neut v 1640 roc'i. in 1640 T h e church was built in 1640.'

In such passives, the accusative NP of the active remains accusative in the passive. If it is type 1, the accusative is not actually unavailable in the passive but only preempted by the nominative. Thus if there is a way of using the nominative up in the passive then the accusative will automatically apply, and we will get a non-promotional passive. Dummies inserted at the top of argument lists have often been proposed. In fact, Sobin proposes a dummy for Ukrainian. In GB the consequence is that Case absorption cannot be the basic property of the passive. On the present account, this is expected. I would like to show that positing an optional null dummy for Ukrainian gets the right results. First of all, before we can give an account of the interaction between passivization, dummies, and case in Ukrainian, it is necessary to establish what type Ukrainian should be analyzed as belonging to. I take Ukrainian to be type 1, primarily on account of its wide range of adjunct accusatives, including accusative of price, accusative of measure, accusative of quantity, and accusative of time (see Shevelov (1963: 17If.) for details about both literary and colloquial Ukrainian). The non-promotional passive in (57) is a passive with the insertion of a dummy argument. The linking for (57) is given in (58): (58)

xA

build ( DUM NOM: [ARG] ACC:

[ ]

y) *NOM: [ARG] ACC:

[

]

The dummy gets the most restrictive case applicable to it. Since it is an NP and a dummy argument it gets nominative. Thus nominative is unallowable on the second argument by the LDC. This leaves accusative, and it does appear in (57).

154 Linker interactions This dummy must be optional, since sentences like the following are also grammatical: (59)

Cerkva bulo zbudovana church + fem.NOM be + past + fern build + part + fem.sg v 1640 roc'i. in 1640 'The church was built in 1640.'

Thus, the dummy is optional. If not inserted, the passive linking is unexceptional, and the passivized subject takes nominative and agreement. Further, if we assume that agreement is the NOM linker and the nominative case is the flip-side of agreement, a sentence in which we get a nominative subject in the passive without agreement should be out, which is correct: (60)

*Cerkva bulo zbudovano church + fem.NOM be + past + fern build + part + neut v 1640 roc'i. in 1640. 'The church was built in 1640.'

Even if we assume that agreement and nominative are syntactically equivalent (their LINK value is [ARG]) but separate linkers, the LDC would still rule out (60). The dummy either absorbs the agreement (NOM) linker as in (57) forcing the accusative to apply or it does not, allowing the agreement and/or nominative to link the passive subject as in (59). There is no way for (60) to be generated. The passive receives a natural account in all four types of two-case language. In the next chapter, I will show how the present account of case interacts with raising. 4.3

Word order

Linking is not accomplished by case alone. In addition to case and adpositions, agreement and word order are the other two major types of linkers. We have already seen two examples of positional linking, the preverbal position in Icelandic (about which more in the next chapter) and the positional linker adjacency-to-V in Japanese. The Icelandic position was like a type 1 nominative, and the Japanese _V linker had the same LINK value as the accusative. In both languages, the positional linker

Word order

155

was in addition to the case linkers, and furthermore the position had the same LINK value as one of the cases in that language. There are, however, many languages that rely primarily on position for linking, a notable example being English. The present approach to direct linking is meant to generalize to all forms of linking, and so such languages should behave similarly to the ones we have examined so far except for the realization of the linkers. The features I have been using (ARG, EXP, GOAL, etc.) as parts of the value of LINK define the cases with respect to the theory, because the theory consists of three constraints that refer only to LINK values. Despite their radically different surface properties, linkers with the same basic distributional properties, captured here as equivalent LINK values, should behave similarly with respect to the AC, the RC, and the LDC. This section will be an analysis of English as a type 1 language, and the next will consider the interaction of case, configuration, and obliqueness. Recall first Kiparsky's (1987, 1989) analysis of English. "Grammatical relations," whether primitive or derived, correlate with word-order position in English. This effect can be achieved in direct linking by means of the rules he proposes for English: (61)

a. Agreement and _V": unrestricted b. V_ : unrestricted c. V'_ : Theme

These rules differ from the linkers on the present approach in the following respects. First, they operate over a list of thematic roles. Each argument is classified into categories such as agent, goal, theme, etc., hierarchically arranged on a universal Thematic Hierarchy. "Unrestricted" in (61) means not restricted to any particular thematic role on the hierarchy. Furthermore, Kiparsky takes agreement to link the highest (available) role because of its special anaphoric nature. On the present approach, however, agreement will be treated as an [ARG] linker and the "highest available" behavior will follow from the type 1 low setting of the Preference Parameter. Then the LDC will allow only one occurrence of agreement per clause, on the highest argument which can get agreement. The LDC, the RC, and the Preference Parameter make the "highest available" nature of English agreement fall out. This makes the correct prediction that if the value of the Preference Parameter is reversed, as in type 4, the analog of English agreement ([ARG]) should

156 Linker interactions show an absolutive pattern, "lowest available" argument. Kabardian was a case of this. Kiparsky's rules in (61) are to be taken with the phrase structure rules in (62): (62)

a. S -> {V", N", P"} b. V" -> {V', N", P"} c. V' -> {V, N", Prt}

For a discussion of how Kiparsky's analysis elegantly captures the possibilities for ditransitives, see Chapter 1: (63)

N' a. b. c. d. e. f. g. h.

sent sent sent *sent *sent *sent *sent *sent

Prt

a drink up up John up John a drink up up up up

to John a drink to John a drink a drink up John John a drink to John a drink a drink John

It will be shown shortly that the present approach also affords a simple analysis of the facts in (63a-h). I will also assume the phrase structure in (62) and (63). For linkers I will adopt the following: (64)

a. _V": [ARG] b. agr: [ARG] c. V":[ ] d. V_: [ARG|GOAL|-HA]

The position _V" ( = VP) and agreement are identical syntactically but distinct with respect to realization. In the unmarked situation they can link the same NP (the subject). The linkers in (64c,d) will, because of the three constraints, normally be used for objects, but they need not be directly stipulated to apply to

Word order 157

objects. One of these linkers needs to be more restrictive than the other. The theory traces restrictiveness back to basic distribution. There are two possibilities. Either we can say that a position to the right of V' is restricted to themes as in Kiparsky's (61c). Goals are then forced into the V_ position in ditransitives. This assumes that linkers make reference to theme, a thematic role which is difficult to define. So far we have avoided making an analog of theme in the system of syntactic argument categories. The other possibility is to reconstruct the notion of an indirect object position after the verb and so have the position V_ limited to goals, as in Wechsler (1995). Translating this into the argument category approach, we would just be using the argument category goal which we have already. This has the virtue of making the position parallel to the goal dative from many languages we have seen already. In the next section we will see that this latter choice, a goal position, is in fact the only one possible in this theory. The default, then, is the within-VP position (V") position, and therefore, like the default in other languages, should link the otherwise unmarked adverbs. If English is indeed a type 1 language, the withinVP position should behave exactly like the accusative in Icelandic, German, and Classical Greek. This position is roughly analogous to structural case in GB: any NP in the VP gets this linker. To show how the present proposal works, I return to the example in (63). For both (63a) and (63b), the linking is as follows: (65)

y|GOAL

agr: [ARG]

to: [ARG/GOAL/-HA] + OBL

z) *_V": [ARG] *agr: [ARG]

V": [ ]

V_: [ARG GOAL - H A ]

V":[

send up ( x

_ F ' : [ARG]

]

*_V": [ARG] *agr: [ARG]

For the first argument the VP-external position and agreement are the most restrictive applicable linkers. There is no higher argument, and so the LDC cannot be violated. VP-external position and agreement are compatible, and so both apply. On the lower two arguments, these linkers are unavailable, because they are the most restrictive linkers that pass the AC for a higher argument. For the goal argument, the preposition competes with the verb-adjacent-position-for-goals and the default VP-internal position. The default must give way to either goal linker since they are more restrictive. Prepositions and word-order linkers are normally not

158 Linker interactions compatible, a subject I return to in the next section. If the preposition option is taken, we get (63a) and (63b). For the lowest argument, the only linker that passes the AC and the LDC is the VP-internal position, which just means that the object must appear somewhere in the VP, either as sister to V (63a) or sister to V' (63b). The VP-internal linker underdetermines the position of the NP, which leads to two options (63a,b). The sentence in (63c) is the result of taking the V_ option over the equally restrictive preposition for the goal. Note that, if this is done, then the last argument cannot appear as sister to the verb. It will still be linked by VP-internal position but will be forced to appear as sister of V', since the sister of V position is taken by the goal. As for the ungrammatical sentence types in (63), in each one some general constraint is violated. First, as in Kiparsky (1989), (63d) is ruled out, because only one NP can be sister to V in the phrase structure. To get (63f,g), at least one constraint must be violated. If the goal is linked by VP: [ ] over to: [ARG|GOAL] and V: [ARG|GOAL], then the RC is violated. Sentence (63e) is well-formed, but only if the NP a drink is linked to the goal argument. This will lead to the bizarre interpretation where a drink is the entity that John is being sent to. If, however, the NP John is linked to the goal argument, then the RC has been violated and the sentence is ungrammatical under the normal interpretation. As for (63g), the adoption of Kiparsky's phrase structure will put a single NP before the PP if both are sisters of V'. The structure necessary for (63g) is not generated. In general, English behaves as a type 1 language with positional linking rather than cases. This should not be too surprising, since most analyses of English identify some position, e.g. SPEC of IP in GB, with nominative case in languages like Greek and Latin, and identify position within the VP with accusative in those languages. As I showed briefly in §4.1.1, the VP-internal position does seem to link a range of adjuncts in the default manner expected. In Chapter 6, Old English will be analyzed as being a case-marking type 1 language. It will be clear that despite radical changes on the surface, English has remained a type 1 language. 4.4

Obliqueness and compatibility of case and configuration

We have already seen several examples of the interaction of case and configuration. One was the preverbal position in Icelandic and the cases. Another was the accusative and V-adjacent position linkers in Japanese. Earlier in this chapter we saw the interaction of the subordi-

Obliqueness and compatibility of case and configuration 159 native suffix -a and the pre-A:e position in Nama. In each of these languages, the cases and the configuration that are interacting are all non-oblique. In the last section, on English, the preposition to and the V-adjacent position conflict in obliqueness, and they are not compatible on the same NP. This will turn out to be true generally. This section will consider the question of the interaction of case and configuration when one or more of those linkers is oblique. It is a commonplace that obliques show a greater freedom within the clause than non-obliques. For example, in English the non-oblique goal phrase has less freedom than the corresponding oblique to phrase (66,67) and the corresponding oblique dative phrase in German (68): (66)

a. It was to Bill that John gave the book, b. ??It was Bill that John gave the book.

(67)

a. To Bill John gave a book, b. ??Bill John gave a book.

(68)

a. Der Schiiler gab dem Lehrer den Apfel. the(NOM) pupil gave the(DAT) teacher the(ACc) apple The pupil gave the teacher the apple.' b. Der Schiller gab den Apfel dem Lehrer. the(NOM) pupil gave the(ACc) apple the(DAT) teacher 'The pupil gave the teacher the apple.' c. The pupil gave the teacher the apple. d. T h e pupil gave the apple the teacher.

At the level of whole languages it has often been remarked that case and configurationality or strict word order are inversely correlated. The question is whether obliqueness or case is what is incompatible with configurationality and how exactly this relationship works. One initially plausible way to capture this is to have oblique case be incompatible with linking by word order, leading to greater freedom of word order, i.e. relative non-configurationality. Typically behavioral obliqueness is associated with lexical case. But Icelandic shows lexical case that marks NPs that behave otherwise much as they do in configurational languages such as English and Norwegian (see Chapter 6). Two types of approach are possible. One is to say that case and word order are inherently incompatible and claim that special cases like Icelandic involve a special type of case. The problem with this is that it creates an otherwise unnecessary division among cases. Another approach is to say that case and word order are always compatible but

160 Linker interactions not when the case is oblique. I will pursue this possibility and attempt to show why this should be the expected situation. The first generalization that needs to be captured is that the only linkers that are oblique are lexical, i.e. they refer to some information more specific than argumenthood (and [—XA]). Under this approach, a case or a positional linker may be oblique or non-oblique, but only if it is lexical, that is, only if it fits the following: (69)

Obliqueness Constraint Linker: +OBL D Linker: [ARG|X] where X is non-null4 + OBL

Any linker may be non-oblique, but only a lexical linker may be oblique. This leads to the following types of linkers: (70) lexical non-lexical

oblique Y N

non-oblique Y Y

This means that type 1 nominative or accusative (or their equivalents in other types) may not be oblique. A lexical case may be oblique, as in German (see Chapter 2), or it may be non-oblique, as in Icelandic. Furthermore, I will take oblique to be the marked value of the pair [±OBL]. This makes sense, since the less restrictive, i.e. the more general, linkers cannot be oblique. More importantly, obliqueness is a property of only those cases that are language-specific. By hypothesis, all languages have a default linker, and most if not all have an [ARG] linker. Thus there is nothing stipulated about these linkers except their realization. These linkers seem never to be oblique. Therefore it is reasonable to assume that the linkers which are least marked with respect to their basic distribution (LINK) are also the least marked with respect to obliqueness. When obliqueness occurs, it occurs on lexical cases, i.e. those cases involving more language-specific stipulation. It is not surprising that the value for obliqueness of these cases is the marked one. The unmarked status of [—OBL] will be captured by the following default: (71)

[

]->[-OBL]

All linkers that are [ + OBL] cannot get [-OBL] by the default since the values are contradictory. But any linker not specified [ + OBL] is automatically [—OBL]. And, as emphasized already, only the lexical cases can be

Obliqueness and compatibility of case and configuration

161

specified [ + OBL], leading all non-lexical linkers to be [-OBL], the desired result.5 Positional linkers are typically non-lexical. Since non-lexical linkers are non-oblique (by the constraint in (69)), positional linkers are typically non-oblique, which is correct. However, the non-obliqueness of positional linkers is not inherent but rather contingent on their being non-lexical. This dependency leads to the following predictions: (72)

A lexical positional linker can be oblique.

(73)

If a positional linker is oblique, it must be lexical.

In the last section, we saw one lexical word-order position, for goals in English. What (72) says is that such a linker may be oblique in some language. Later I will present Dutch as such a language. One last prediction is that oblique case is not in principle incompatible with linking by position. The only constraint on co-occurrence of case and position is that they should not have contradictory values for obliqueness, which follows if the AVMs of the linkers must unify. The main constraint we need is one to prevent an NP from being linked both by an oblique linker and a non-oblique linker at the same time. This could of course be stipulated, but it should follow from something more basic. Intuitively, what we want to express is a generalization that the properties of a linker are inherited by the NP which is linked by that linker. If two incompatible linkers link the same NP, then in some sense, the NP is a contradiction. It is not clear whether it has the property of the one linker or the other. When it comes to obliqueness, the situation is clear. If an NP is linked by an oblique linker, the NP is not a term but an oblique. If, on the other hand, an NP is linked by a non-oblique linker, then it will be a term, by definition. But if an NP were linked both by an oblique linker and by an non-oblique linker, the NP would have contradictory properties associated with it. What we want is for the information about obliqueness versus non-obliqueness (and presumably other mutually exclusive properties) to transfer (e.g. by "percolation") to the NP where consistency is checked: (74)

Obliqueness Inheritance fLINK

LINK

REAL [LINKERX [OCOBL]]

D

NP

REAL [LINKERX aOBL

NP

162 Linker interactions This constraint may be derivable from more general principles of inheritance, but I will leave (74) as is for the discussion to follow. Turning back to case and configuration, we are now in a position to consider how an oblique case and a non-oblique positional linker will interact. If the oblique case is more restrictive than the position, we have on some argument a competition between the following: (75)

a. Case: [ARG|X]

where X is non-null

+ OBL

b. Position: [Y]

where Y = ARG or is null

—OBL

If both are applicable to some NP (pass the AC and the LDC in isolation), then we have to consider Obliqueness Inheritance and the RC. If both linkers applied, then the linking would violate Obliqueness Inheritance. If just the position applied, then the RC would be violated, because there is a more restrictive unused linker that could apply to that argument in isolation, namely the other case. The case alone is fine and must be used. The oblique case-marked NP is ineligible for positional linking. One interesting prediction is that if a word-order position linker refers to information more specific than argumenthood, i.e. it is lexical, it could be associated with [ + OBL]. Thus we predict the existence of a language in which the English-type goal position links an NP that behaves like an oblique (e.g. does not passivize). A possible example of such an oblique lexical positional linker is the verb-adjacent position in Dutch. In that language the indirect object must appear next to the verb (Verhagen (1986: 202f.) and references there): (76)

a. Toen heeft Jan het kind zijn laatste boterham gegeven. then has Jan the child his last sandwich given Then Jan gave the child his last sandwich.' b. ??Toen heeft Jan zijn laatste boterham het kind gegeven. then has Jan his last sandwich the child given 'Then Jan gave the child his last sandwich.'

From the discussion of these examples it is clear that Verhagen gives (76b) two question marks because the string is well-formed but only under the bizarre interpretation where a child is being given to a sandwich, whereas the pragmatically expected reading goes with (76a) not (76b). So far Dutch is parallel to English. Where they differ is in the

Obliqueness and compatibility of case and configuration

163

fact that the goal may not be the subject of the passive in Dutch (Hoekstra (1984: 159)): (77)

a. dat that 'that b. *dat that 'that c. dat that 'that

ik Marie bloemen gaf I Mary flowers gave I gave Mary flowers' Marie bloemen werd gegeven Mary flowers was given Mary was given flowers' Marie bloemen werden gegeven Mary flowers were given flowers were given (to) Mary'

The ditransitive passivizes only if the lower, "theme" argument is subject of the passive, as indicated by the singular versus plural agreement in (77b) and (77c). Most theories of the passive predict that if neither object is oblique, then either both objects passivize (symmetric) or only the higher object passivizes (asymmetric). By "higher" is meant higher in argument structure or "thematically" higher. To preserve this generalization, the goal must be oblique in Dutch and so not an object. However, there is no oblique case (no dative) on that NP, but its position is fixed, so the oblique linker is likely to be the position. Dutch thus illustrates our prediction in (73) that if a word order is oblique, it has some restriction beyond [ARG]. This can be handled by having a word-order position for goals which is oblique. Consider again the choice for ditransitives we faced in English between a position restricted to themes (as in Kiparsky (1987, 1989)) or one for goals (as in Wechsler (1995). Using argument categories, the latter analysis can be adopted by using the argument category goal which we have made use of already. The other analysis would involve translating the thematic role theme into an argument category. The preference for the goal analysis coincides with the necessity for it in Dutch. The non-passivizability of the goal in Dutch indicates that we should analyze it as an oblique. But, to be oblique, that NP must be linked by a lexical linker. The only way for the goal to be linked by a lexical linker is to posit a positional linker for goals. If the goal were linked by the default position, it could not be oblique on the present theory. The word-order position for goals in Dutch is roughly "before the other object." Very tentatively I suggest _V' as the positional linker for (syntactic) goals in Dutch:

164 Linker interactions (78)

_V':

[ARG|GOAL] + OBL

Just like the German dative, this oblique word-order position will prevent the goal from becoming subject (highest non-oblique argument) in the passive (see Section 5.3 for the German passive). We predicted (in (72)) that an oblique positional linker should be found, and we predicted (in (73)) that if such a linker were found, it would be lexical. Two predictions converge on Dutch and are borne out. This approach further predicts that if a word-order position like the pre-V' position in Dutch can bear [ + OBL], then that word-order linker would be compatible with a lexical oblique case. This would appear as a dative that is fixed in its position. This prediction is particularly interesting, because it contradicts the traditional assumption that case and fixed position are complementary. What we are predicting is that even an NP that is both a behavioral and a coding oblique can be fixed in position. This seems to be borne out by the facts in certain dialects of Swiss German, especially Grisons German (Andreas Ludwig, p.c). In Grisons German, there is a distinct morphological dative, and it appears on goals: (79)

und den het dr dogdor Suarer em Bbuur and then has the doctor Saurer the(DAT) farmer de KB erklart the KB explained 'and then Dr. Saurer explained the KB (artificial fertilization) to the farmer'

(80)

de Bbuur git mier noch duusigg Franka the fanner gives me(DAT) still thousand Francs 'The farmer will give me a thousand Francs.'

The goal is fixed in its position. If it is placed after the other object, the result is ungrammatical. So corresponding to (79) and (80), we get: (81)

*und den het dr dogdor Suarer de KB em and then has the doctor Saurer the KB the(DAT) Bbuur erklart farmer explained 'and then Dr. Saurer explained the KB (artificial fertilization) to the farmer'

Obliqueness and compatibility of case and configuration 165 (82)

*de Bbuur git noch duusigg Franka mier the farmer gives still thousand Francs me(DAT) 'The farmer will give me a thousand Francs.'

The other order of the dative and the accusative NP is ungrammatical. In (82), this could be ascribed to the pronoun following the noun, but in (81) both the dative and the accusative NPs are definite NPs. The lack of the accusative-dative order cannot be reduced to ordering statements based on pronounhood versus nounhood or definiteness versus indefiniteness, familiar from Standard German. Rather Swiss German shows truly fixed order of dative goal before accusative object. The goal, despite its being positionally fixed, is oblique. In ditransitive passives, the goal does not become the subject. This can be tested by trying to make the dative goal NP the target of control: (83)

*ich

hoffe duusig Franke ggee ze werde hope thousand Francs given to become 'I hope to be given a thousand Francs.'

I(NOM)

In general, dative goals, like dative goals in Standard German, do not pass the tests for subjecthood, showing that they are behavioral obliques. The goal in Grisons German seems to be doubly linked by dative and verb-adjacent position. Furthermore, NPs linked by these linkers are oblique. The two goal linkers for Grisons German would be the following: (84)

a. DAT: [ARG|GOAL] + OBL b. _V': [ARG|GOAL] + OBL

Note that the position is oblique and lexical, as predicted: (85)

geen ( x

y|GOAL

NOM: [ARG]

DAT: [ARG/GOALJ

+ OBL ACC: [

]

_ V1:

[ARG/GOAL]

+ OBL NOM: [ARG] ACC:

[ ]

z ) 'give' *NOM: [ARG] ACC:

[

]

166 Linker interactions In the passive, the goal is not the subject (highest non-oblique argument), because it is oblique: (86)

geen ( xA

y|GOAL

z ) 'give'

DAT: [ARG 1 GOAL]

NOM: [ARG]

+ OBL

ACC:

[ ]

_ V': [ARG1 GOAL] + OBL NOM: [ARG] ACC:

[ ]

The goal is doubly linked by dative and verb-adjacent position, but Obliqueness Inheritance will not cause a clash: (87)

LINK [ARG [CAT [GOAL]]]

[ V_

DAT [ + OBL] 1

[+

OBL]j

+ OBL

NP

The NP is oblique, and this comes from both linkers, so the structure is fine. Grisons German is a confirming instance of the prediction that an oblique word-order linker is possible, and that such a linker must be lexical. Further, such a positional linker is compatible with oblique case in Grisons German, also as predicted. 4.5

Compatibility of linkers

Having opposite values for obliqueness is not the only way for two linkers to be incompatible. I have assumed that different types of linkers are in general compatible, and different linkers of the same type are incompatible. Such a view is too simple, and I will end this chapter with a short discussion of the compatibility of linkers. This discussion will be far from conclusive. Most versions of GB assume some version of "case clash" which prevents more than one Case per theta role. Direct Linking also originally (Kiparsky (1987)) incorporated a restriction that each NP must be linked exactly once. I have formulated the above conditions on linking to keep open the possibility of multiple linking, and have analyzed some languages as involving multiple linking, typically by a case and a wordorder position. The Restrictiveness Constraint disallows a less restrictive

Compatibility of linkers 167 linker to link an NP to the exclusion of a more restrictive linker that is available. The RC, however, says nothing directly about blocking. The RC only enforces disjunctive ordering between linkers if the linkers are incompatible. I now consider the general question of compatibility. Linkers can be divided into three main types: (i) morphological case and adpositions, (ii) agreement and (iii) word order. As discussed already, case and adpositions work similarly, although adpositions have a tendency to be independent predicators, and cases tend to fall on the more grammaticalized end of the spectrum from independent predicators to syntactic linkers. For example, the preposition to in English is both a linker for goal arguments and an independent predicator. Both case and prepositions sometimes co-occur with word-order linking, but generally only if they are highly grammaticalized. I proposed in the last section that, in principle, cases (and adpositions) are always compatible with word-order. However, one factor prevents their co-occurrence in most cases. Word order rarely refers to lexical information and so can rarely be oblique. Lexical case is often oblique. By Obliqueness Inheritance, then, such cases and word-order positions do not mix in general. But if the case and the word-order position are both oblique (Grisons German) or both non-oblique, then they should be able to co-occur. We correctly predict, then, that there are two main ways for case and word order to be compatible. One is if the case is not oblique, as with the syntactic cases (nominative, accusative, ergative, absolutive in all types), or with nonoblique lexical cases (as in Icelandic, see Chapter 2 and the next chapter). The other way to be compatible is if the word-order position is oblique, as we have already seen for Grisons German. Thus the theory seems to handle the relation of case and configuration in an explanatory way. Likewise, agreement and word-order position seem to be in principle compatible. If agreement is much like case but just differs in appearing on the verb rather than the argument, then this is quite expected. Agreement is perhaps for diachronic reasons rarely lexical and so rarely oblique. Thus a combination of agreement and word order should usually be possible. Agreement and case is trickier. Many languages (e.g. Kabardian) have agreement with a wide range of cased NPs. One could say that this agreement linked the NP in addition to the case, instead of the case (but then nominal case is a mystery), or that it does not link. Also, it is not clear whether in these cases agreement picks out NPs on the same basis as a case (e.g. by argument category) or whether it picks out NPs on

168 Linker interactions the basis of their case. This problem is particularly acute for nominative. As I have argued, nominative in Classical Greek, German, Icelandic, etc. is not the default. But it does seem to be the reverse side of agreement in at least some of these languages. Kiparsky (1987) analyzes nominative as a no-case and agreement as the linker. On the other hand, as we shall see, Icelandic allows nominatives which do not agree. Many South Asian languages have an agreement rule that could be described as: agree with the functionally highest nominative (Joshi (1987) on Marathi, Mohanan (1994) for Hindi). Next comes the question of the compatibility of two linkers of the same type. First of all, Obliqueness Inheritance should rule out an oblique and a non-oblique linker of the same type from linking the same NP (e.g. an oblique lexical case and an accusative). I know of no cases of double agreement. Double word order is hard to imagine. Certainly two NPs cannot occupy more than one position at the same level. On the other hand, if an NP were linked by adjacency to the verb and internality to the VP, it could be linked twice, by two word-order positions. I know of no analysis that seems to hinge on this issue. Multiple morphological case may be possible, but two types must be immediately distinguished. Dench and Evans (1988) do show that a given NP may bear more than one case, but their examples are of cases used on different "levels" marking the same NP. An example is the NP warrapala-ku 'grass-LOC-ACc' in (88) from Panyjima (Dench and Evans (1988: 8)): (88)

ngunha watharri-ku nyurna-yu warrapa-la-ku that look for-PRES snake-Ace grass-LOC-ACC karta-larta kurrjarta-ngarni. poke-FUT spear-PROP 'He is looking for the snake in the grass to stab it with a spear.'

Here the NP bears an adnominal case inside a verbal argument (Dench and Evans' "relational") case. The relational accusative case appears on all the constituents of the NP it is associated with and so appears on the dependent NP as well. This NP has a locative case that marks this lower level dependence as well. Dench and Evans document other levels besides adnominal and relational from which case can "percolate down to an NP." They note, however, that some caution is in order, because the different uses of the cases may be related diachronically rather than synchronically. In any event, like GB Case Theory and other case theories, the present framework deals only with what Dench and Evans

Compatibility of linkers

169

term "adnominal" and "relational" case: case that signals the relation of an argument to its predicator. The question still remains whether two cases on the same level, especially two relational cases, are compatible. Dench and Evans give no examples of this. One instance of "case stacking" where both cases are argument cases has been claimed in the literature for Korean (Gerdts and Yoon (1988: 164f.)) on the basis of sentences like (89): (89)

Suni-eykey-ka Chelsu-eyiyhae chaek-i cu-eci-et-ta. Suni-DAT-NOM Chelsu-by book-NOM give-pas-pst-ind 'Sooni was given the book by Chelsu.'

None of the ten or so speakers I have consulted accepts (89), but this does not rule out the possibility that it is fine for other speakers. If any of the sentences that Gerdts and Yoon (1988) cite as instances of case stacking are fine, this can be handled very elegantly within the present system. The cases that can stack must be marked, since not all speakers evidently allow all combinations if they allow one. But once a given combination of cases is allowed, there is the matter of deriving the order which may not be reversed: (90)

*Suni-ka-eykey Chelsu-eyiyhae chaek-i cu-eci-et-ta. Suni-NOM-DAT Chelsu-by book-NOM give-pas-pst-ind 'Sooni was given the book by Chelsu.'

Gerdts and Youn (1988) have a principle that case depth reflects the order of strata in RG. Thus the dative in (89) precedes the nominative, because the NP heads a 3 arc at an earlier stratum than the stratum at which it heads a 1 arc. This can be captured within the present theory by just requiring that the NP plus any case must satisfy the RC. Thus the NP plus the dative does, since there is no more restrictive applicable case than the dative. NP plus nominative does not, because the RC is violated: the more restrictive dative is not present. All other examples of case stacking also show the order predicted by the RC in this way.6 Thus, the RC can derive the order of case stacking to the extent that it occurs. To summarize, I give the following table of compatibilities, leaving open the question of the extent to which multiple linking by linkers of the same type is possible:

170 Linker interactions (91)

case case ? w.o. agreement

word order Y ?

agreement Y/N Y ?

Thus, generally it seems safe to assume that, in the unmarked case, linkers of the same type cannot multiply link an NP. Other combinations of linkers are in principle fine, subject to Obliqueness Inheritance and any language-particular constraints.

5

Icelandic

In this chapter I will provide an analysis of some central problems in Icelandic syntax within the theory developed so far. I would like to follow out one of the consequences of the theory in the last chapter, that of the possibility of double linking. If, as argued there, obliqueness and nonobliqueness (represented by [±OBL]) is introduced to NPs through the linkers that link them, then the possibility of lexical case and word order doubly linking NPs is predicted. Just this double linking, I will argue, is the best way to explain the well-known combination of rich case and relatively strict word order in Modern Icelandic. In the next chapter, I will provide a diachronic explanation of the rise of Modern Icelandic double linking. The essence of the analysis I propose is that lexical case in Icelandic does not carry the [ + OBL] feature specification and that Icelandic is type 1. If the lexical cases are not oblique, they will be non-oblique by default, and so it follows that the lexically cased NPs will behave as terms. This will be true whether or not there is a word-order position to link them as well. I will argue that preverbal or pre-VP position is the only positional linker in Icelandic. Subjects, then, will typically be doubly linked, but objects will not. Finally, the approach advocated here will provide a predictive account of the interaction of case, passivization, and coordination. 5.1

Case, agreement, and word order

Icelandic was the first language considered in Chapter 2. There I provided an analysis of case in Modern Icelandic, and illustrated the basic workings of the theory using Icelandic data. The case alternations Dative Substitution and Nominative Substitution were examined in detail and were found to be well accounted for by the restrictiveness-based theory of case. Later this theory was extended to cover languages of different types, 171

172

Icelandic

and a basic typology of pure two-case systems was presented in Chapter 3. Icelandic, like Greek, is a type 1 language, i.e. it has the characteristic type 1 values for the Limitation Parameter and the Preference Parameter: (1)

Type 1: Limitation Parameter: No Preference Parameter: Low (Top-Down)

These parameter values cause any [ARG] linkers and any default linkers to show a nominative-accusative pattern, with an elsewhere distribution of accusative. The cases are as in (2): (2)

a. b.

NOM: [ARG] ACC:

[ ]

Any [ARG] linker (nominative and, as seen below, pre-VP position) will appear on the highest argument for which it is possible by the LDC and the RC. The LDC will question the linking of lower arguments in comparison to the most restrictive possibilities of each respective type for higher arguments. For the [ARG] linkers, this means that the LDC will disallow them except in "highest available" position. The no value of the Limitation Parameter means that the any-argument case will not be prohibited on the highest argument. Recall that Icelandic had a dative of experiencers and a dative of goals, which in the present theory are cases referring to the argument categories experiencer and goal. See Chapter 2 for details: (3)

a. DAT: [ARG|EXP] b. DAT: [ARG|GOAL]

The argument category approach in essence says that morphological dative is associated with two interpretations, one as a sentient being aware of something, and the other as the entity towards which something moves, a goal. However, it is not correct that every time verbal entailments are of this nature we automatically get the dative. The implication is one way from syntactic experiencer to semantic experiencer. Some semantic experiencers were nominative. Although the use of argument categories is looser than thematic relations in useful ways, they are quite heavily constrained. First, they embody the non-trivial claim that certain uses of the dative receive a uniform interpretation, and that these uses will have syntactic consequences.

Case, agreement, and word order

173

Most importantly, the features corresponding to the argument categories are referred to by the three constraints, AC, RC, and LDC, which make a wide range of predictions. The LDC will enforce a weaker syntactic analogue of "thematic uniqueness" (see §2.6 above). If we had more than one syntactic goal per clause, then the lower one would never be able to get dative, because that would violate the LDC. Icelandic has a wide range of idiosyncratic cases as well. They are generally of the form in (4): (4)

Case:

[ARG|X|Y]

where Y = 1, 2, ... and X may be null

Each of these cases is part of the lexical entry of some verb. The idiosyncratic datives make reference to the same morphological dative paradigm as the dative of goal, but their syntactic distribution is different, i.e. they have radically different LINK values. Some of the many idiosyncratic cases needed for Icelandic are given in (5): (5)

a. b. c. d.

ACC: [ARG|EXP|1]

e.

GEN: [ARG|2]

ACC: [ARG|1] DAT: [ARG|2] GEN: [ARG|1]

(part of (part of (part of (part of able') (part of

the the the the

lexical lexical lexical lexical

entry entry entry entry

for for for for

langa 'long') reka 'drift') tyna 'lose') gceta 'be notice-

the lexical entry for sakna 'miss')

In this chapter, we will see some more idiosyncratic cases on the third argument of some ditransitives.

5.1.1 Nominative and agreement As I mentioned briefly in the last chapter, the interaction of case and agreement is somewhat special. Nominative and agreement often go together, but the correlation is not perfect. Some languages have nominative and no agreement, and others have agreement that correlates with a nominative case, often a zero case. Kiparsky (1989) proposes that nominative is the absence of case and that the agreement is the real linker. The correlation between agreement and nominative, however, is one way only, as I now show for Icelandic. In Icelandic, if an NP is the source of agreement, it must be nominative. But if an NP is nominative, it may or may not be the source of verbal agreement. Agreement is obligatory with preverbal nominatives:

174 Icelandic (6)

t>eir hafa/*hefur etid. They(NOM) have(3pl)/have(3sg) eaten 'They have eaten.'

If the highest argument is dative, the agreement is never with the dative NP despite the fact it is the highest grammatically linked argument (the subject) and despite the fact that it is the preverbal NP (Zaenen, Maling, and Thrainsson (1985)): (7)

Mer *hef/hefur/hafa alltaf J>6tt me(DAT) have(lsg)/have(3sg)/have(3pl) always thought J>eir leidinlegir. they(NOMpl) boring(NOMpl) 'I have always considered them boring.'

Agreement in (7) can be with the postverbal nominative NP, or it can be default third-person singular agreement. But the verb cannot be in the first-person singular to agree with the preverbal dative NP, although it passes the subjecthood tests. This means that agreement is not a behavioral subjecthood property. The statement that the verb agrees with the behavioral subject does not work for (7). Nor can we say that the verb agrees with the NP in preverbal position, for example by saying that INFL assigns nominative to preverbal position. The preverbal NP is neither nominative nor the source of agreement (INFL in GB). The sentence in (7) demonstrates another interesting generalization about Icelandic, namely that the verb agrees only optionally with postverbal nominatives. Putting this together with the fact that the verb agrees with the preverbal nominative, we get the following: (8)

Agreement with a preverbal nominative is obligatory but optional with a postverbal nominative.1

If we take surface evidence at face value, then the postverbal NP in the non-agreeing version of (7) must be linked by nominative not agreement. Otherwise, that NP would not be linked, thereby violating the Linking Condition (all NPs must be linked). But the sentence with no agreement with the postverbal nominative NP is fine, so the Linking Condition must not be violated. The linker must be nominative. Another question is of course why we get the pattern in (7), i.e. why agreement is obligatory but not if the nominative NP is postverbal. One account that claims to be explanatory is Van Valin (1991). On his analysis the reason agreement is optional in (7) is that the nominative NP is not

Case, agreement, and word order

175

even potentially a subject. He proposes the generalization that agreement with potential subjects is obligatory. Besides being transderivational, this cannot work for the closely related Old Icelandic, where the facts are the same, but the nominative N P is the subject (as I will show in the next chapter). Nygaard (1905: 67f.) gives abundant examples of postverbal plural subjects and singular agreement. First of all, this can happen especially if one or more words intervene between verb and subject: (9)

a. Magniisi konungi likadi ilia storgjafir Magnus king(DAT) liked(3sg) ill big-gifts(NOMpl) t>aer er Hakon konungr hafdigefit those(NOMpl) which Hakon-NOM king-NOM had given bondum til vinsaeldar ser. farmers(DAT) for friendship self(DAT) 'Magnus the king did not like the presents that Hakon the king had given the farmers for their friendship.'

(Hkr.6313) b. i Imnn tima fannst i DanmQrk kvernsteinar at that time was-found(3sg) in Denmark millstones(NOM) tveir sva miklir tWO(NOM) SO

big(NOM)

'At that time there were in Denmark two millstones so big.' (££.78.41) Nygaard further documents that optional agreement with a postverbal nominative is particularly free with the verb pykkja 'seem,' the Old Icelandic precursor of Modern Icelandic pykja 'think' in (7) above. Thus we find examples like the following (Nygaard (1905: 68)): (10)

mer fcotti vit vera i hellinum me(DAT) seemed(3sg) we to-be in hell-the(DAT) 'It seemed to me that we were in hell.' ( 0 / / . 117.28)

In Old Icelandic the nominative was the subject (see the next chapter), and so Van Valin's explanation cannot hold for it. The nominative is definitely the potential subject (Van Valin's pivot), because it is the subject. Thus the Old Icelandic agreement remains to be explained, and it seems likely that the same explanation should cover both Old and Modern Icelandic. I leave open how the agreement rule should be formulated in Modern and in Old Icelandic and what the exact differences, if any, are.

176

Icelandic

5.1.2 Word order Icelandic is notable for being one of the few rich case languages with relatively fixed word order. In the direct linking framework, it is a question to what extent this is linking by word-order position or not. Any positing of word-order linkers depends on a phrase-structural analysis of the language in question. This area of Icelandic grammar has been the subject of some controversy, especially regarding the question of whether Icelandic has a VP or not. I refer the reader to Thrainsson (1986) for a complete discussion. Nevertheless, it is clear that Icelandic is a V2 language: (11)

a. Eg

hef aldrei hitt Mariu. have(lsg) never met Mary(Acc) 'I have never met Mary.' b. Mariu hef eg aldrei hitt. Mary(ACc) have(lsg) I(NOM) never met 'I have never met Mary.' I(NOM)

The correct generalization is that exactly one constituent may precede the verb. 2 The V2 phenomenon has been a subject of intense study, and analyses of Icelandic (e.g. Holmberg (1986) and Rognvaldsson and Thrainsson (1990) and references there) differ in important respects. Most importantly for present purposes, it is possible to tell the difference between topicalized and subject NPs in Icelandic, as demonstrated at length by Zaenen, Maling, and Thrainsson (1985) and summarized in §2.1.1 above. Only subjects but not topicalized NPs can undergo Raising, be the antecedents to obligatory reflexivization, participate in inversion, be extracted, be postposed if indefinite, participate in subject ellipsis, and be the targets of control (Equi). Further, it is one of the best-known results of Icelandic syntax that the NP that bears these behavioral subjecthood properties is the NP that occupies the preverbal position in uninverted sentences. There is thus a strong correlation between subjecthood and the pre-VP or preverbal position, depending on whether we posit a VP or not: (12)

a. _VP: SUBJ b. _V: SUBJ

(13)

a. _VP: [ARG] b. _V: [ARG]

Case, agreement, and word order 111 The possibilities in (12) refer to grammatical functions and are one way to capture the relation of subjecthood and the sentence-initial position in Icelandic. The versions in (13) are the possible linkers on the present approach. Before comparing them, notice that, in inversion (lib), the subject appears right after the verb. If one accepts (12a) or (13a) with VP, then either V2 involves movement or the realization of _VP in an inversion sentence as immediate postverbal position. If one chooses (12b) or (13b), then _V should be read as "adjacency to V" rather than as "immediately preceding V." This issue is not central here. Tentatively I accept (13a), although nothing that follows depends on this choice. The main point is to establish that a positional linker with the [ARG] specification interacts with the other linkers in the proper way. The other, mediated linking approach works in this case, since it captures subjecthood and the position directly. The linking in (14) demonstrates that this is also true on the present approach: (14)

hitta ( x

y ) 'meet'

NOM: [ARG]

*NOM: [ARG]

_VP:

*_VP: ACC: [

ACC:

[ARG]

[ ]

[ARG]

]

In this plain transitive with the verb hitta 'meet,' the pre-VP position and the nominative, having [ARG], are the most restrictive linkers applicable to the highest argument. Since Icelandic has the type 1 parameter settings, the LDC will not question the linking of the highest argument, because there is no higher argument. By the RC, the top argument is linked by the nominative and the pre-VP position. This is the highest non-obliquely linked argument and so is the subject. The lower argument cannot get either of these linkers, because that would violate the LDC. The only remaining choice is accusative, which links the lower argument. This argument is a non-highest, non-obliquely linked NP and so is the object. Case and word order are in principle compatible, and, if they are both non-oblique, then the dative subject will automatically be linked by the position as in (15): (15)

Barninu batnadi veikin. child-the(DAT) recovered-from disease-the(NOM) 'The child recovered from the disease.'

The linking is as follows:

178 (16)

Icelandic batna ( X|EXP

y ) 'recover'

DAT: [ARG/GOAL]

NOM: [ARG]

NOM: [ARG]

*_VP: [ARG] ACC: [ ]

_ VP: [ARG] ACC: [ ]

In this case the dative of experiencer is the most restrictive linker that passes the AC for the highest argument and so links the NP. The nominative is not compatible, but the pre-YP position is compatible with the dative and so also links the first argument. The second argument is allowed to be linked by the nominative, since that is not the most restrictive linker of its type (case) applicable to a higher argument. Specifically, the dative is the most restrictive case for the higher argument, so the use of nominative on the other argument is fine. The use of the pre-VP position on the lower argument, on the other hand, would violate the LDC, because pre-VP position is the most restrictive linker of its type, word order, applicable to the higher argument. Because they are of different types and because there is a more restrictive case but not position for the higher argument, the second argument gets nominative but not pre-VP position by the LDC, even though they have the same LINK value, [ARG]. None of the linkers are oblique, so both arguments are grammatically (non-obliquely) linked. The highest grammatically linked argument is the preverbal dative and so is the subject. The other argument is the next highest grammatically linked argument and so is an object. By the LDC and the fact that the pre-VP position is the most restrictive linker of its type, the correlation between word order and subjecthood is captured. Turning to non-highest arguments, consider first whether we could have more than one positional linker without a VP. I will not try to answer this general question, but, before answering it for Icelandic, it will be worthwhile to consider the possibilities. If there is no VP in Icelandic, and if there is positional linking for one or both of the latter arguments in (17), then these positions must be radically different from the positional linkers of English (see the previous chapter): (17)

Eg

sagdi barninu soguna. told child-the(DAT) story(Acc) 'I told the child a story.'

I(NOM)

Case, agreement, and word order

179

If the sentence has a flat structure, i.e. no VP, there can be no linkers like _V', VP-internal position, or _VP. At most there could be V_ for the second NP. It is difficult to see how the last NP could be positionally linked unless by reference to the second NP, but this may be feasible, especially if the word order is fixed. Furthermore, if one's framework allows verb movement, NPs non-adjacent to the verb on the surface could be adjacent to the verb at some other level of structure. Just positing a flat structure is not enough to argue against positional linking for the objects in (17). More convincing evidence against positional linking comes from the low degree of fixedness of the word order in (17). The other order of the NPs in (17), given in (18a), is considered bad style but not as unacceptable as the corresponding sentence in English (18b): (18)

a. Eg

sagdi soguna barninu. told story(Acc) child-the(DAT) 'I told the child a story.' b. *I told a story the child. I(NOM)

Zaenen, Maling, and Thrainsson (1985) report that the following (based on examples of Rognvaldsson's) are both acceptable: (19)

a. Eg

gaf konungi ambattina gave king(DAT) slave-the(ACc) 'I gave the king his maidservant.' b. Eg gaf ambattina konungi I(NOM) gave slave-the(ACc) king(DAT) 'I gave the maidservant to her king.' I(NOM)

sina. self S(ACC) sinum. self S(DAT)

They note that the dative NP need not be especially heavy in order to appear last.3 For one who wanted to propose positional linking in these sentences, this difference between Icelandic and English must be explained. Even if the verb moves from some position nearer the objects in (18) and (19), the relative freedom of order of the objects is unexpected if they are supposed to bear some relation to a fixed, although invisible, verbal position in the sentence. This question of the extent of positional linking is important in that it distinguishes the analysis here from other direct linking analyses (e.g. Smith 1993), Kiparsky (p.c.)). In particular Kiparsky suggests that it should be a general principle that lexical cases are always oblique unless they are also linked by position. As we have seen, a positionally linked

180

Icelandic

NP can be oblique, as in the case of Dutch and Grisons German goals (previous chapter). On the present theory, a lexical case can simply fail to be oblique. If so, it may be linked by a non-oblique position, but it need not be. I take the relative freedom of the postverbal NPs in Icelandic ditransitives as evidence that we are not dealing with positional linking within the VP in Icelandic. If that is so, then the dative goal NP is predicted to behave as an object without being positionally linked. The dative is a non-oblique lexical case, a possibility expected on this theory. I return to this question in §5.4. In this section we have seen that there seems to be one positional linker in Icelandic. Thus in what follows the actual phrase structure (flat versus VP) is not as important as the positing of one positional linker. 5.2

Case preservation and nominative objects

In the last chapter I showed that one of the first results that falls out from the RC is the priority of lexical case over grammatical case. In this section I will extend this result to what will be seen as a closely related problem: nominative objects. 5.2.1 Case preservation Case preservation on the present account is a special case of the RC. Both kinds of lexical case, idiosyncratic and argument-category, are more restrictive than the grammatical cases nominative and accusative. First consider the by now familiar dative of the object of tyna 'lose': (20)

a. Barnid tyndi bokinni. child-the(NOM) lost book-the(DAT) 'The child lost the book.' b. Bokinni var tynt (af barninu). book-the(DAT) was lost by child-the(DAT) 'The book was lost (by the child).'

In these two sentences, the following three cases are applicable, the dative only to the logical object, the other cases to any of the arguments. Additionally, the pre-VP position can potentially apply to any argument: (21)

a. DAT: [ARG|2] b. NOM: [ARG] c. _VP: [ARG] d. ACC: [ ]

(part of lexical entry of the verb tyna 'lose')

Case preservation and nominative objects

181

Taking the active in (20a) first, we get the linking in (22): (22)

tyna ( x

y ) 'lose'

_ VP: [ARG]

DAT: [ARG/GOALJ

NOM: [ARG]

*_VP:

ACC: [

*NOM: [ARG]

]

ACC:

[ARG]

[ ]

The second argument must get dative by the RC, because it is the most restrictive linker that passes the AC and the LDC for that argument. Note in particular that the dative blocks the accusative, since the environment in (21a) properly includes the environment in (21c). Restrictiveness automatically captures the priority of lexical case. In the active, the nominative is not a possible linker for the second argument, because the LDC rules it out. The problem with the nominative for this lower argument is that nominative is the most restrictive linker of its type applicable to the highest argument. In the passive, however, there is no higher argument than the logical object, and so the nominative is not prohibited on this argument. If the nominative is the most restrictive case applicable to this logical object, as in most passives of transitives, then the nominative applies to the logical object and we get a nominative subject in the passive. With a verb that takes an idiosyncratic case for its logical object, however, the nominative, although not ruled out by the LDC for the logical object in the passive, nevertheless cannot apply to the logical object because it is not the most restrictive case passing the AC and the LDC. Rather the idiosyncratic case must apply by the RC, blocking the nominative (and the accusative): (23)

tyna ( xA

y ) lose' DAT: [ARG I GOAL]

_ VP:

[ARG]

NOM: [ARG] ACC:

[ ]

The dative in (21a) is more restrictive than the nominative in (21b) (and the accusative in (21c)) and so wins. The theory correctly predicts a sentence containing only a dative. This dative NP is the subject, because it is the highest non-obliquely linked argument. In a language where lexical case is oblique, e.g. German, the dative NP would be an oblique, and we would have a true impersonal passive. The present approach to case preservation makes all types of rule ordering including the prevalent

182 Icelandic "pre-linking" of lexical case (Yip, Maling, and Jackendoff (1987)) come without any stipulation. Case preservation is captured using the same features needed for distribution in the first place. Further, case preservation is seen to be a special instance of a very general phenomenon. The dative is "preserved" for the very same reason that nominative has priority over the accusative. The RC provides a unified account. 5.2.2 Nominative objects Closely related to case preservation is the so-called "nominative objects" problem (Zaenen, Maling, and Thrainsson (1985), Yip, Maling, and Jackendoff (1987), Sigurdsson (1989), Harbert and Toribio (forthcoming)). The problem consists in the unusual association of nominative case (and agreement) with an NP that is neither in subject position (e.g. Spec of IP in GB) nor shows behavioral subjecthood properties. The reasons why this is a problem is that many theories, and most notably GB, make nominative Case assignment by INFL correlate necessarily with subjecthood. The sentences in question are given in (24) and (25): (24)

Mer lika fcessir bilar. me(DAT) like(pl) these(NOM) cars(NOM) 'I like these cars.'

(25)

Honum voru gefnir fcessir bilar. him(DAT) were(pl) given(m.pl) these(NOM) cars(NOM) 'He was given these cars.'

Example (24) exhibits a dative experiencer subject of the verb lika 'like,' and (25) a dative goal subject of a passive on the ditransitive gefa 'give.' Each of these verbs is typical of their respective classes, psych verbs and ditransitives. For some verbs of the DAT-NOM class and ditransitive passives, the other NP can be subject, but I will deal with this question in the next section on the (a)symmetric object question. For now the focus is on sentences where the dative NP is the subject and the nominative NP is not. In (24) and (25) the dative NP but not the nominative NP shows all the subjecthood properties. They can "raise," as in the following examples (for the full range of tests see Zaenen, Maling, and Thrainsson (1985)):

Case preservation and nominative objects (26)

Hann telur mer lika Jsessir bilar. he(NOM) believes me(DAT) like these(NOM) cars(NOM) 'He believes me to like these cars.'

(27)

tel honum hafa verid gefnir believe him(DAT) have been given(m.pl) fcessir bilar. these(NOM) cars(NOM) 'He believes me to have been given these cars.'

183

Eg

I(NOM)

The preverbal dative is the subject. If there is only one subject per clause, a standard assumption, then the nominative NP cannot be the subject. It must be an object or something else. This problem carries over into infinitives as well. The case marking in infinitives behaves just as if the covert subject (PRO in GB) were assigned a case, and this is in fact the analysis of Sigurdsson (1989). Exactly when the subject of the infinitive would be dative, the object appears as nominative (29): (28)

Olafur vonast til ad kaupa bilinn. Olaf(NOM) hopes COMP to-buy car-the(Acc) 'Olaf hopes to buy the car.'

(29)

a. Konungurinn vonast til ad verda gefnar king(NOM) hopes COMP to-be given(f.pl) ambattir. slaves(NOM.f.pl) 'The king hopes to be given maidservants?' b. Ad vera gefnar ambattir var to be given(f.pl) slaves(NOM.f.pl) was mikill heidur. great honor 'To be given maidservants was a great honor.' (Zaenen, Maling, and Thrainsson (1985: (52))

The nominative objects are a problem for GB since they get nominative Case, which is assigned by INFL, without being in the Spec of IP position where it is assigned. Many solutions to this problem have been proposed, and many are summarized in Harbert and Toribio (forthcoming). The solution in LFG is quite straightforward. LFG would have the same problem as GB if the rule for nominative were "assign nominative

184 Icelandic to the subject." But instead, Zaenen, Maling and Thrainsson (1985: 466) have the following case-marking rules: (30)

Default Case Marking: the highest available GF is assigned NOM case, the next highest ACC (Universal)

This gets the nominative objects. The assignment rule is an instance of the Case Tier Hypothesis. Case is assigned to GFs (as in (30) or NPs (as in Yip, Maling, and Jackendoff (1987)) left-to-right, and the first GF (or NP) that is not pre-specified for a lexical case will get nominative and subsequent NPs accusative. When the first GF is prespecified dative, the second NP is nominative, a nominative object. Although they do not discuss case in infinitives, the Case Tier approach in (30) assumes that GFs with no corresponding NP in c-structure also get case. For the postverbal NP in (29) we might say that the non-overt subject is not available because it is non-overt, causing nominative to appear on the object. But this will not work where the covert subject would not be lexically case-marked as in (28) above. There the covert subject corresponds to a nominative NP in a free-standing clause, and the object of this lower clause in (28) is accusative, not nominative. Given the association principle in (30), the only way the ACC can be assigned to the overt object is for there to be a higher GF that gets the nominative. Otherwise, the overtly expressed GF would be the highest available and get nominative, but it should get the accusative. In the present framework, the case marking of nominative objects follows automatically from the nature of nominative and of accusative cases and the Restrictiveness Constraint. The specifications for nominative, [ARG], and accusative, [ ], were independently justified on the basis of their distribution. Further, the dative of experiencer, [ARG|EXP], and goal, [ARG|GOAL], are similarly justified on distributional grounds. Putting these together with the RC, we get both case preservation (last section) and the explanation of nominative objects. To see why, consider the linking for the passive of the ditransitive: (31)

gefa-PASs ( xA

y|GOAL

z ) 'give'

DAT: [ARG I GOAL]

NOM: [ARG]

NOM: [ARG]

*_VP: [ARG] ACC: [ ]

_ VP: [ARG] [ ]

ACC:

Case preservation and nominative objects

185

Since this is a type 1 language, the LDC questions the linking of lower arguments relative to the possibilities for higher arguments. Given this, it follows that dative, pre-VP position, nominative, and accusative are the possibilities by the AC for the goal argument. In this passive, this is the highest argument, so the LDC cannot disallow any possibilities. From the RC and the incompatibility of two morphological cases, it follows that dative will block nominative. Similarly, dative precludes accusative. Dative and pre-VP position are compatible, so both dative and the position apply. For the next argument, nominative is the most restrictive case that passes the AC and the LDC and so applies. It is just because there is a lexical case applicable to the higher argument that nominative is not the most restrictive linker of its type applicable to the higher argument. This means that the nominative is allowed by the LDC for the lower argument, where it must apply by the RC over the less restrictive accusative. The present theory directly relates the presence of the dative case for the higher argument and the possibility of nominative on the other argument. Both the priority of lexical case and the "highest available" behavior of nominative follow from the same principles, and the connection between lexical case and nominative objects is captured. The notion of "highest available" in the LFG of Zaenen, Maling, and Thrainsson (1985) and previous direct linking accounts (e.g. Kiparsky (1987)) is not needed. Zaenen, Maling, and Thrainsson use "highest available G F " and Kiparsky uses "highest available role" to restrict the nominative. Given the parameters behind the case system typology and the RC, the "highest available" aspect of the nominative's distribution follows automatically. Given that Icelandic is type 1, the LDC and the RC ensure that the nominative has "highest available" distribution. This complex distribution is ultimately derived from general principles, the parameters of Chapter 2, and the basic nature of the cases. From the specifications of the linkers, compatibility (morphological and LDC), and the constraints, the nominative objects phenomenon is directly accounted for. Note that all these elements of the account (specifications, compatibility, and constraints) are independently needed and justified, and so in essence the nominative objects account comes for free. More importantly, the approach taken here captures the relation between case preservation and nominative objects. Nominative objects occur on a lower argument when the higher argument is linked by some case more restrictive than the nominative. The nominative in turn is more

186

Icelandic

restrictive than the accusative and so gets that next argument. Case preservation and nominative objects are thus seen to be two manifestations of the same relation: restrictiveness in linking. Note that although the pre-VP position and the nominative have the same syntactic specifications, the compatibility of the word-order position with the lexical case and the incompatibility of nominative with the lexical case cause the characteristic pattern of distribution of the two [ARG] linkers. Since it is compatible with the more restrictive case linker, the pre-VP position always links the highest argument. Since the nominative is not compatible with the dative and is less restrictive, it is forced into its characteristic "highest available" pattern. Thus from the specifications, restrictiveness, and compatibility it is possible to explain the "highest" versus "highest available" behavior of pre-VP position and the nominative.

5.2.3 Nominative objects and control This analysis carries over to the infinitival constructions. This section will deal with control and the next with Raising. Approaches to control or Equi have involved relating two positions, only one of which is "overt." The original transformational rule of EquiNP Deletion related two coreferential NPs in Phrase Structure and deleted the lower of the two. PRO in GB is a coindexed non-overt NP. In lexicalist theories, the two related positions are not NP positions in phrase structure but two positions at functional structure or argument structure, only one of which corresponds to an overt NP. What will be essential to the theory of case we are developing and its relation to control is that the non-overt position receives case. Translating the LFG and HPSG approach to control (Bresnan (1982), Pollard and Sag (1987)) into the present framework, what we need to do is identify two argument positions through coindexing. Thus an Equi verb like vonast til 'hope' should have a partial lexical entry like this: (32)

[vonast til ( wi [verb ( . . . Xi . . . ) ] ) 'hope']

The two arguments are coindexed so they will be coreferential but will not be syntactically identical. Given that there are two arguments, they will be linked separately. They will both need to be linked, even though there is only one position in phrase structure. Henniss (1989) presents arguments from Malayalam that potential NPs such as PRO should also

Case preservation and nominative objects

187

get case, leading her to propose a Determinate Case Principle whose effect is very similar. Furthermore, we must ensure that the target of the control is not oblique. For Icelandic we must further be sure that the target of control is the subject (This may not be universally true (Kroeger (1993)). Lexical entries like (32) should be combined with the following constraint which ensures that the the controllee is the subject: (33)

Controllee Constraint [ verbl ( Wi [ verb2 {... x{ ...)])] D [ verbl ( Wi [ verb2 ( . . . X^-OBL] . . . ) ] > ] and there is no higher non-oblique argument of verb2

If grammatical subjecthood is not a universal condition on controllees, as Kroeger (1993) argues, then the last clause in (33) should be dropped. The specification of non-obliqueness on the controlled argument does, however, seem to be universal. To get the combined argument structure for the sentences in (29) above, we insert the argument structure of the passive of gefa 'give' into the argument structure of vonast til 'hope': (34)

[vonast til ( Wi [gefa-PASS ( xA yi|GOAL[-OBL] z ) 'give'] ) 'hope']

Now if we allow the linkers to apply to all the arguments, including both coindexed ones, we get the following: (35)

[vonast til ( wA [gefa-PASS ( xA yi|GOAL[—OBL] Z ) 'give']) 'hope'] _VP NOM ACC

DAT NOM ACC

NOM ACC

First, the goal argument of the ditransitive passive does satisfy the Controllee Constraint, since it is indeed the highest non-obliquely linked argument of the lower predicate, i.e. it is a covert subject. As for the linking of these arguments, the LDC disallows the use of a linker if it is the most restrictive linker of its type applicable to a higher co-argument. The top argument of the lower verb is not a co-argument of the top argument of the matrix verb and so the LDC does not disallow nominative. The effect is that each of the verbs defines its own case domain. The non-overt argument, the goal, gets the dative by the RC. The lower argument of the lower verb gets nominative for exactly the same reason as the nominative objects in the simple sentences of the last

188 Icelandic section. Since the higher argument of the lower predicate has dative as its most restrictive applicable case, the lower argument is free by the LDC to take nominative. As predicted, this argument even shows agreement in the infinitival construction in (29). More generally, the effect of linking covert arguments is to make the linking of the arguments of embedded predicates work exactly the same as the linking of the arguments in the corresponding simple sentence. If the embedded verb has a covert top argument that is linked by a lexical case, then the next argument will be nominative. If the covert argument is nominative, then the other argument is accusative by the LDC. Far from being caseless like PRO, the covert argument's being linked is central to the account. The present theory, in conjunction with the assumption that covert arguments are linked, captures the similarity of the linking of infinitival complements and simple clauses. The LDC and the RC get exactly the right results for the interaction of case and control, but only under the assumption that non-overt arguments are linked. This does not necessarily mean that phrase structure must contain some empty position PRO, although adoption of such an empty position is consistent with the present theory of linking. Rather, what control shows is that arguments with no phonological content must be linked like any other argument. If phrase structure does not contain any position for the non-overt argument, as in LFG where Equi verbs take VP rather then S' complements, then linking of arguments that do not correspond to NPs must be allowed for. If linking is done at argument structure, this result is hardly surprising. If, however, linking is best done at some postlexical level, then we nevertheless must allow for linking of NP-less arguments. Either way, the present theory of case provides strong motivation for the linking of non-overt arguments in control.

5.2.4 Case preservation and Raising The other infinitival construction that the restrictiveness-based theory of case must take into account is Raising. Traditionally Raising is regarded as the other "NP-Movement" type of construction along with passive. One very striking similarity between passive and Raising in Icelandic is that lexical case is "preserved" in both constructions. Whether or not passive and Raising receive similar analyses, the preservation effect must follow in each case.

Case preservation and nominative objects

189

Raising is different from control in that the argument that is shared between the higher and the lower predicates belongs semantically to the lower predicate but syntactically to the higher predicate. The following sentences are truth-conditionally equivalent, despite the fact that different arguments are raised in (36a) and (36b): (36)

a. John believes the doctor to have examined Bill. b. John believes Bill to have been examined by the doctor.

The paraphrase relation between the embedded active and passive is preserved in Raising. It does not matter which argument is raised. This contrasts with Equi where active and passive do not lead to truthconditional equivalence under embedding: (37)

a. John persuaded the doctor to examine Bill. b. John persuaded Bill to be examined by the doctor.

Most theories distinguish Raising and Equi by the role of the lower argument in the matrix clause. In both Raising and Equi, this NP belongs syntactically to the matrix clause at surface structure, but only in Equi is it semantically an argument of the matrix predicate. The verb persuade entails that the person persuaded is at least sentient. In (37a), this is the doctor, and in (37b) it is Bill. Raising verbs do not entail anything about the raised argument, which is considered to belong semantically to the lower predicate exclusively. The analysis of Equi, following LFG and HPSG, involved coindexation of two arguments. The upper argument is an argument of the upper predicate and can participate in entailment relations involving the matrix predicate. Coindexation would not work for Raising, however, because we do not want an independent argument in the matrix clause. The standard analysis in LFG and HPSG distinguishes Raising from Equi by analyzing Raising as involving structure sharing between the upper and lower predicates. In LFG the values of the grammatical function features of one argument of the higher predicate and one argument of the lower predicate are shared. Since semantic interpretation is off the functional structure, the position in the upper clause will belong to the lower clause semantically. In HPSG, the entire sign for the raised argument is shared between the lower and the higher arguments. The basic insight of these approaches is that the upper argument is essentially a placeholder and that whatever it contributes to the interpretation is only by way of the lower argument. In the present theory, NPs

190

Icelandic

are syntactic arguments if they bear the value ARG for LINK. Any NP that bears ARG is syntactically selected by the verb. What makes an argument more than a placeholder is its value for ARG. SO I will further assume that semantic interpretation will be off values of ARG. This was necessary for argument categories like experiencer and goal, where the interpretation of the argument is constrained by its membership in the argument category. Generalizing this, an argument is interpreted as belonging semantically to whichever predicate it gets its argument category from. If the matrix NP is to be treated as an argument of the lower predicate, what should be shared between the lower and the upper arguments is this value for argument category (and dependent values). A raising verb will have an entry like the following: (38) " NP [LINK [ARG [CAT [...]]]] NP [LINK [ARG [CAT [ \\\ ]]]] believe

SUBCAT

VP | SUBCAT

|p NP ,

LINK [ARG [CAT [...] [T] ]] —OBL

In (38) the verb's list of arguments is the value of a feature SUBCAT as in HPSG. The value ARG takes an Attribute-Value Matrix (AVM) containing the feature CAT (argument category) and its value in turn (e.g. DIR, GOAL). An example is [ ARG [ CAT [ DIR ]]] for an argument belonging to the default category of arguments and [ ARG [ CAT [ GOAL ]]] for a syntactic goal. This value for ARG is shared between the complement verb's argument and the upper verb's argument.4 This means that the upper argument will be interpreted just like an argument of the lower predicate. Representations like (38) are cumbersome, and for convenience I will represent this informally using a line connecting the arguments which share values of ARG: (39)

[verbl (. . . x . . .[ verb2 (. . . Z[-OBL]. . . ) ] ) ]

The lower verb, verb2, is the verb heading the VP complement. There are two argument positions in the structure, and so both must be linked. However, as in LFG and HPSG, there will be only one position in the phrase structure for an argument and that will be in the upper clause. The lower verb cannot combine with an NP that corresponds to this argument. Overall, the lower clause will correspond to a VP in the syntax. Thus the lexical entry for a verb like believe is as follows:

Case preservation and nominative objects (40)

believe ( x

y

191

[verb (. . . Z[-OBL]. . . ) ] )

i

i

The fact that the argument belongs semantically to the lower clause follows from the fact that the upper argument position shares structure, namely the value of ARG, with the lower argument, which is in turn the input to semantic interpretation of the lower predicate. As for linking, the only novel aspect about linking for arguments like the second argument of believe is that there are two sources for applicable linkers. This argument will be eligible by the AC for (i) all the linkers that apply by virtue of its syntactic selection by the verb believe ([ARG] and default linkers) and (ii) all linkers that can apply by virtue of the shared material with the argument of the lower clause. So if the lower argument belongs to some argument category that receives a special case or if it is assigned an idiosyncratic case, then this information is passed up to the higher argument and the lexical case is applicable there as well. This will be our account of case preservation in Raising. Turning back now to the Icelandic examples of Raising, we must explain the following: (41)

J)eir telja Mariu hafa skrifad ritgerdina they(NOM) believe Mary(ACc) to-have written thesis-the(Acc) They believe Mary to have written her thesis.'

First, the lexical entry for telja 'believe' will be similar to that for English believe: (42)

telja (x

y [verb (. . . z[—OBL]. . . ) ] ) 'believe'

The lower verb, skrifa 'write', has an ordinary transitive lexical entry: (43)

skrifa ( x y ) 'write'

When the larger argument structure is created out of the argument structure of telja 'believe' and the argument structure for skrifa 'write,' then we get (44): (44)

telja ( w x

[skrifa (y[—OBL] Z ) 'write' ] ) 'believe'

Now if the linkers apply to this combined argument structure, then we get the following:

192 Icelandic (45)

telja ( w NOM

VP ACC

x [skrifa ( y[-OBL]

z ) 'write' ] ) 'believe'

*NOM

*NOM

*_VP ACC

NOM

_VP

*_VP

ACC

ACC

This gets the correct cases in the correct places. To see how, consider first the upper predicate. As usual, the highest argument gets nominative over accusative and the pre-VP position is compatible. For the next argument, the linkers that might apply (pass the AC) are those that it would get by being a syntactic argument of the upper predicate (nominative, pre-VP position, accusative). No lexical cases refer to the shared material, so no extra linkers apply. By the LDC neither nominative nor pre-VP position is applicable. For the lower predicate, the highest argument is linked as highest arguments usually are, but in the phrase structure there is no place for it, and so it cannot show its linking overtly. An alternative that must be considered is that the lack of nominative in the last example is due to the incompatibility of agreement and infinitivals. On this alternative, if we take agreement to be the real linker, we could say that, since the lower verb is infinitival, there can be no agreement and so no nominative, and similarly there is no preverbal position available. Or raising might be said to absorb nominative case. However, these approaches would have the additional problem of the nominative objects again. If an argument that corresponds to a lexically case marked NP in the active is "raised," then the other argument gets nominative (example from Andrews (1982: 464)): (46)

Hann telur barninu i barnaskap sinum hafa he(NOM) believes child-the(DAT) in foolishness self s to-have batnad veikin. recovered-from disease-the(NOM) 'In his foolishness he believes the child to have recovered from the disease.'

It cannot be true that raising in any way absorbs agreement-nominative, since the lower verb agrees with the nominative NP in (46). The present analysis of (46) is as follows, with the lexical entry for batna 'recover' combined with the lexical entry for telja 'believe': (47)

batna ( X|EXP y ) 'recover from'

Case preservation and nominative objects (48)

telja (w

193

x [batna (y|EXp[-OBL] z) 'recover from'] ) 'believe'

NOM

DAT

DAT

NOM

_VP

*NOM

NOM

*_VP

ACC

*_VP

_FP

ACC

ACC

ACC

The only difference here is the applicability of the experiencer dative. Since argument category is shared from the lower argument to the higher argument, the dative of the argument category experiencer is applicable in both clauses. The LDC does not rule out identical linking across a predicate boundary, so both the lower argument and the higher argument are dative. As in other nominative object constructions, the LDC allows nominative for the last argument, and this is the correct result. We can thus explain why the nominative is available for the non-raised argument just in case the raised argument would get non-nominative case (46) and why the lower argument gets accusative when the raised argument would get nominative (41). On the present analysis this is because the raised argument does get two cases. This two-case analysis contrasts with the HPSG analysis where the entire sign for the raised NP is structure-shared. Because the case on the second argument of the lower predicate in the present theory depends so much on what case the higher argument gets, we have no choice but to say that only the argument specifications are shared but not case. Values of ARG are shared, but the values for linkers (REAL) are not. Case is assigned to subcategorized NPs, and, there being two of these in the matrix SUBCAT list, there are two Unkings. The matrix argument is a syntactic argument of both predicates, but its ARG path comes from the lower predicate, and so it is only a semantic argument of the lower predicate. Since case depends on the specifications of the arguments and the specifications are shared, the same cases will be applicable to the arguments in the sharing relationship. Which linkers actually apply depends on the LDC and the RC in both clauses. We can explain why dative but not nominative is "preserved" under raising. Unlike in HPSG, case is not shared directly. Preservation only occurs when the linker is applicable by referring to the shared material, but type 1 nominative does not refer to the value of [ARG] but [ARG] itself. Lexical cases are preserved because they both pass the LDC (no higher argument in the upper clause has this linker as a possibility), and the RC requires it as the most restrictive linker passing the AC and the LDC. 5

194 5.3

Icelandic Is Icelandic symmetric or asymmetric?

In this section I will address the question of whether Icelandic is a symmetric- or an asymmetric-object language. Symmetric-object languages are those in which either of the two lower arguments of a ditransitive shows primary-object behavior, e.g. the ability to be the subject of a passive. In asymmetric-object languages, only the higher of the internal arguments can be primary object. Icelandic shows elements of both patterns. As is well known, lexical case in Icelandic interacts with the ability of a ditransitive to passivize on the lowest argument. One class, the NOM-DAT-ACC class, represented by segja 'tell' below, allows either object to passivize, while all other ditransitives only allow the goal to passivize. Therefore, the question of whether Icelandic is a symmetric- or an asymmetric-object language has no ready answer. Icelandic seems to be symmetric for the segja 'tell' class but asymmetric for the others. The proper question to ask is: is Icelandic basically symmetric, asymmetric, or neither. I will present evidence that Icelandic is basically symmetric, but that the symmetric-object question is more complex than the setting of a parameter for the whole language. Icelandic has the symmetric value for the symmetric-object parameter for ditransitives, but the assignment of idiosyncratic case blocks symmetric-object behavior for the non-NOM-DAT-ACC classes. It also shows symmetric behavior with some two-argument verbs. This mixed behavior requires a more fine-grained approach to symmetric objects than a single parameter setting. I will show how the treatment of the symmetric-object parameter in the present theory predicts the full range of phenomena in Icelandic. Before trying to locate Icelandic within the range of possibilities that languages have for ditransitives, it is necessary to lay out the three main types of language. Ditransitive passives are of interest, because they contain two arguments after the suppression of the highest argument of the active counterpart. The first division among languages with respect to ditransitives is between what Bresnan and Moshi (1990) call "symmetric-" and "asymmetric-object" languages. In the symmetric-object languages, either of the two lower arguments of a ditransitive can show primaryobject behavior. Most strikingly, either can be the subject of a ditransitive passive, leaving the other argument behind as some type of object. In asymmetric-object languages, on the other hand, only one of the two lower arguments shows primary-object behavior, and only that

Is Icelandic symmetric or asymmetric?

195

argument can be the subject of a ditransitive passive. Of languages we have examined already, Japanese is symmetric, and German is asymmetric with only the non-goal showing primary-object behavior. One dialect of English is asymmetric with only the goal as primary object. Japanese is typical of symmetric-object languages. In Japanese ditransitives, the goal is marked by the dative particle -ni and the lowest argument receives the accusative (-o or _V): (49)

Daitooryoo ga gakusei ni kunsyoo o ataeta. president NOM student DAT medal ACC gave 'The president gave a medal to the student.'

As indicated in §3.1 above, the proper treatment of lexical cases in Japanese in the present theory involves specifying them [-HA] like the accusative (-o) and the V-adjacent position (_V). The dative will by hypothesis be as follows: (50)

DAT: [ARG|GOAL|—HA]

The ditransitive in (49) forms two passives: (51)

a. Kunsyoo ga daitooryoo ni (yotte) gakusei medal NOM president by student ni atae-rare-ta. DAT give-PASS-PAST

'A medal was given to the student by the president.' b. Gakusei ga daitooryoo ni kunsyoo o atae-rare-ta. student NOM president by medal ACC give-PASS-PAST T h e student was given a medal by the president.' As evidenced by (51), Japanese is a symmetric-object language. On the present scheme, a necessary condition for one argument to outrank another in GF is for it to outrank it on the argument structure. The fact that either of two arguments can outrank the other in the passive of a ditransitive in Japanese suggests that each ditransitive is associated with two argument structures differing minimally in the order of the lowest two arguments. A conventional approach on ordered argument theories is to allow an operation permuting the lower two arguments of a ditransitive (see, e.g. Uszkoreit (1987)): (52)

Symmetric Object Operation (Optional) verb ( x y z ) -» verb ( x z y )

196 Icelandic The Symmetric-Object Parameter here is the presence or absence of this operation in a language. Symmetric-object languages have (52); asymmetric-object languages do not. This operation permuting arguments is not associated with any morphology, making it quite unlike the passive. Furthermore, as we will see, a major constraint on such permutation seems to be that only adjacent arguments may flip. This suggests that the operation, if it is an operation, is not one of permutation but rather one which suspends the order of two arguments. I will return to this possibility later. Turning back to the ditransitive in Japanese, we see first that the goal passive works as follows: (53)

give-PASS ( xA

y|GOAL NOM: [

z) ]

ACC: [ARG/—HA] NOM:

[ ]

Japanese, being type 2, has the low value of the Preference Parameter, so that the LDC will question the linking of higher arguments in the light of the most restrictive linkers of each type applicable to lower arguments. On the lower argument, the accusative and nominative compete, and the LDC says nothing about the linking, because there is no lower argument to compare with. The accusative is more restrictive than the nominative and so applies. On the goal argument, the accusative and the dative cannot apply, because the goal is the highest argument, and those cases have [-HA] (similarly for V-adjacent position which I am leaving out for expositional ease). The nominative, although the least restrictive, is the only possible case. The other passive is the result of passivizing the verb whose argument structure has undergone the Symmetric-Object Operation. In this other passive, we get roughly the same linking: (54)

give-PASS ( xA

z NOM: f

y|GOAL ) ]

DAT:

[ARG/GOAL/—HA]

ACC: [ARG|-HA] NOM:

[ ]

This time the goal is lowest argument and so all three cases might apply by the AC, and the LDC has nothing to say about the linking of this argument, because there is no lower argument. The dative applies by the RC, since it is most restrictive. For the highest argument, i.e. z, the

Is Icelandic symmetric or asymmetric?

197

accusative this time is not applicable, since it has [—HA]. The only remaining possibility is nominative. One dialect of English is asymmetric. In a ditransitive passive, only the goal may be the subject of the passive: (55)

John gave Mary the book.

(56)

a. Mary was given the book, b. *The book was given Mary.

Unlike Japanese, the ditransitive passive with the "theme" in subject position is ungrammatical (56b) for such speakers. This is expected if the goal is consistently higher than the other remaining argument, which it will be if this dialect lacks the Symmetric-Object Operation (52). In the passive the goal is the highest grammatically linked argument and so the subject (see Chapter 3 for details). German also appears to be an asymmetric object language, but it only allows the other, non-goal, argument to passivize. I will argue that this type of asymmetric-object language really is not asymmetric. The appearance of asymmetric behavior stems from the obliqueness of the lexical cases including the goal dative. In a ditransitive, the goal may not be the subject: (57)

Der Schiiler hat dem Lehrer einen Apfel gegeben. the(NOM) pupil has the-DAT teacher a(ACc) apple given T h e pupil gave the teacher an apple.'

(58)

a. Dem Lehrer ist ein Apfel gegeben worden. the(DAT) teacher ist an(NOM) apple given become(pAss) 'An apple was given to the teacher.' b. Ein Apfel ist dem Lehrer gegeben worden. an(NOM) apple is the(DAT) teacher given become(pAss) 'An apple was given to the teacher.'

Although two word orders are possible for the German ditransitive passive, both orders reflect the same functional structure. In both (58a) and (58b), the nominative NP is the subject, unlike in Icelandic. Recall from Chapter 1 that Icelandic and German differ in the status of their lexically case-marked NPs. Icelandic is well known for its false obliques: lexically case-marked NPs which behave nonetheless as terms. German, however, does not have this property. So in the dative experiencer construction, the dative NP was not the subject in German (but was in Icelandic). The

198

Icelandic

lexically case-marked NPs fail all the subjecthood tests. So too the dative goal in the ditransitive passive does not behave as the subject, e.g. it cannot be the target of control: (59)

*Er hofft einen Apfel gegeben zu werden. he(NOM) hopes an(Acc) apple given to become 'He hopes to be given an apple.'

But the other NP can be the target of control: (60)

Der Sklave hofft dem Konig gegeben zu werden. the(NOM) slave hopes the(DAT) king given to become 'The slave hopes to be given to the king.'

Thus, German appears to be asymmetric with only the non-goal passivizable. But consider the definition of subject, which any subject of a ditransitive passive or any other passive must satisfy: it must be the highest grammatically linked argument of its clause. Since syntactic goals (argument category GOAL), are linked by an oblique case, dative, the NPs corresponding to GOAL arguments are never terms and hence never highest terms, i.e. never subjects. The "asymmetric" behavior of ditransitives in German follows directly from the status of lexical case in German. Ditransitive verbs in Icelandic take a wide range of case frames. I illustrate these case frames in (61) (Thrainsson (1979: 21-2)): (61)

a. t>eir leyndu Olaf sannleikanum. they(NOM) concealed Olaf(Acc) truth-the(DAT) 'They concealed the truth from Olaf.' (NOM-ACC-DAT) b. Jon bad mig bonar. John(NOM) asked me(ACc) favor(GEN) 'John asked me a favor.' (NOM-ACC-GEN) c. Eg sagdi J>er soguna. I(NOM) told you(DAT) story(ACc) 'I told you a story.' (NOM-DAT-ACC) d. Olafur lofadi Mariu ^essum hring. Olaf(NOM) promised Mary(DAT) this-DAT ring(DAT) 'Olaf promised Mary this ring.' (NOM-DAT-DAT) e. Maria oskadi Olafi alls gods. Mary(NOM) wished Olaf(DAT) everything(GEN) good(GEN) 'Mary wished Olaf everything good.' (NOM-DAT-GEN)

Is Icelandic symmetric or asymmetric? 199 The fact that the subject is nominative follows from the system above. Icelandic is type 1 which will result in the [ARG] linkers both applying to the highest argument, as long as there is no more restrictive lexical case to displace the nominative. As for the lowest argument, it can get any of the three remaining cases. At the very least, we need special datives and genitives for (61a,d) and (61b,e), respectively. Finally, the middle NP is a goal in all of the sentence types in (61), and so the datives in (61c-e) follow directly from the general dative for goals. The accusative second arguments in (61a,b) are goals semantically without belonging to the syntactic argument category goal. Recall that syntactic goals are semantic goals but not necessarily vice versa. We have already seen accusative goals in Classical Greek which behave as arguments of default category (direct).6 Putting these pieces together, we get the following lexical entries for the verbs in (61): (62)

a. leyna ( x y z ) 'conceal' DAT: [ARG13]

b. bidja ( x y z ) 'ask for' GEN: [ARG 13] c. segja ( x y|GOAL z ) 'tell' d. lofa ( x y|GOAL z ) 'promise' DAT: [ARG13]

e. oska ( x y|GOAL z ) 'wish' GEN: [ARG13]

Note that the entries in (62) make some correct predictions about the relative frequency of the different classes of ditransitives, if we assume that relative markedness should predict relative frequency. In particular the NOM-DAT-ACC pattern should be most common, since it is the least marked. The class in (62c) both has the least number of specifications and is subject to fewer of the more restrictive cases. Yip, Maling, and Jackendoff (citing a verb list in Maling ms.) show that the NOM-DATACC pattern is by far the most common with more than seventy-five entries, compared to twenty-five for the NOM-ACC-DAT pattern which is the next highest. The other verb classes have smaller numbers, about which it is harder to draw statistical conclusions. Following previous authors I take the lack of a NOM-ACC-ACC frame to be an accidental gap. 7 Some of the lexical entries in (62) contain idiosyncratic cases which have scope over only that entry. Some contain arguments of the category

200

Icelandic

goal, to which dative is regularly assigned. Putting these cases together with the other linkers proposed for Icelandic, we have the following inventory of linkers applicable with ditransitives: (63)

a.

DAT: [ARG|3]

b.

GEN: [ARG|3]

(part of the lexical entry of leyna 'conceal,' lofa 'promise,' etc.) (part of the lexical entry of bidja 'ask for,' oska 'wish,' etc.)

c. DAT: [ARG|GOAL]

d. NOM: [ARG] e. _VP: [ARG] f. ACC: [ ] In connection with (63) the natural question arises whether we need multiple dative cases in (63a,c) or whether they could be collapsed. All the datives are morphologically the same, including all allomorphy. All this may mean is that each of the syntactically distinct datives makes reference to the same dative paradigm. This is symbolized with the label DAT. It is a separate question whether the datives are syntactically unitary or distinct. The present theory provides some justification for making a synchronic syntactic distinction between the datives. If the LDC holds in Icelandic, which by hypothesis it should, then we are actually forced into the position that the two datives in (6Id) are syntactically distinct. If goal datives and idiosyncratic datives came from the same syntactic case (or, alternatively, were outputs of the same rule), then (6Id) is a clear violation of the LDC. To the extent that the LDC is motivated both cross-linguistically and within the grammar of Icelandic, there is solid motivation for the separateness of the two datives. Pushing this point further, the most obvious way the two datives are different is that one marks goals and the other does not. Thus we have yet more evidence for the argument category GOAL or some representation of goalhood at argument structure. Ditransitives in Icelandic show two patterns with respect to the passive. The NOM-DAT-ACC class allows either object to passivize and become subject. If the goal is passivized, the dative case remains (Zaenen, Maling, and Thrainsson (1985: 460)): (64)

a. Konunginum voru gefnar ambattir. king-the(DAT) were given(f.pl) slaves(NOM.f.pl) 'The king was given maidservants.'

Is Icelandic symmetric or asymmetric?

201

b. Ambattin var gefin konunginum. slave-the(NOM.f.sg) was given(f.sg) king-the(DAT) 'The maidservant was given to the king.' That the preverbal NP in both (64a) and (64b) is a subject is now well known. For complete details, see Zaenen, Maling, and Thrainsson (1985: 461f.). Of their seven tests I will reproduce just one, subject-verb inversion. Subjects but not topics show inversion, making inversion a test for behavioral subjecthood. For both of the sentences in (64), the preverbal NP shows inversion: (65)

a. Voru konunginum gefnar ambattir? were king-the(DAT) given(f.pl) slaves(NOM.f.pl) 'Was the king given maidservants?' b. Var ambattin gefin konunginum? was slave-the(NOM.f.sg) given(f.sg) king-the(DAT) 'Was the maidservant given to the king?'

Thus, as is true for all passives in symmetric object languages, NOM-DATACC ditransitives in Icelandic allow either passive. The other classes in the passive all show asymmetric-object behavior: only the immediately postverbal NP in the active passivizes. The NOMDAT-DAT class in (61d) is representative. Zaenen, Maling, and Thrainsson (1985: 459) demonstrate the behavior of this class with skila 'return'. The active has the NOM-DAT-DAT pattern: (66)

Eg skiladi henni peningunum. I(NOM) returned her(DAT) money-the(DAT) 'I returned her the money.'

As with all of the remaining verbs, skila 'return' only allows the postverbal NP, the goal, to passivize: (67)

a. Henni var skilad peningunum. her(DAT) was returned money-the(DAT) 'She was returned the money.' b. *Peningunum var skilad henni. money-the(DAT) was returned her(DAT) 'The money was returned (to) her.'

Furthermore, only the goal in the passive in (67) can raise, showing that it alone is the subject:

202 (68)

Icelandic a. Eg

tel henni hafa verid skilad I(NOM) believe her(DAT) to-have been returned peningunum. money-the(DAT) 'I believe her to have been returned the money.' b. *Eg tel peningunum hafa verid I(NOM) believe money-the(DAT) to-have been skilad henni. returned her(DAT) 'I believe the money to have been returned (to) her.'

Only the second but not the third argument of all ditransitives of these other classes can become the subject in a passive. This asymmetric-object behavior is characteristic of all but the NOM-DAT-ACC verbs. Since, for all the verbs in the asymmetric class, the non-passivizing argument is idiosyncratically case marked, it is tempting to say that idiosyncratic case prevents these NPs from passivizing. The simplest way to implement this would be to say that these idiosyncratic cases are oblique. This simple solution cannot work since the lowest argument can passivize if it is the only overt postverbal argument: (69)

a. Eg skiladi peningunum. I(NOM) returned money-the(DAT) 'I returned the money.' b. Peningunum var skilad. money-the(DAT) was returned 'The money was returned.'

Further, if the goal surfaces as a prepositional phrase, then the "theme" also can passivize (Zaenen, Maling, and Thrainsson (1985: 460)): (70)

a. Eg skiladi peningunum I(NOM) returned money-the(DAT) 'I returned the money to her.' b. Peningunum var skilad til money-the(DAT) was returned to 'The money was returned to her.'

til hennar. to her(GEN) hennar. her(GEN)

The sentences with the PP goal (70) behave exactly as the sentences without an overt goal (69), with the dative NP passing the behavioral subjecthood tests (Zaenen, Maling, and Thrainsson (1985)). In light of (69)

Is Icelandic symmetric or asymmetric?

203

and (70), it is not possible to make the inability of the lowest argument of the non-NOM-DAT-ACC ditransitives dependent on the nature of the idiosyncratic case. If one said that the dative for the lowest argument of skila 'return' is oblique, one would wrongly predict that the dative NP in (69) and (70) should be non-subjects, which is incorrect. The correct generalization for the data so far is that the lowest argument of a ditransitive can passivize only if a) it is not lexically case-marked or b) it is the only NP eligible to be subject in the passive. If both conditions fail, the lowest argument cannot passivize. Only if at least one of these conditions holds can the lowest argument passivize. Thus accusative lowest arguments passivize (condition (a) holds), and solitary dative and genitive argument NPs passivize (condition (b) holds, e.g. ((69), (70)). I would now like to show how the split properties of Icelandic grammar with respect to the symmetric-object parameter follow from the theory being developed. The basics are that Icelandic has the Symmetric-Object Operation permuting the lower two arguments of ditransitives in general, but that the designation for lexical case blocks the additional argument order. Recall that I am assuming a hierarchical argument structure. Thus for a basic ditransitive, the argument structure is as follows: (71)

verb ( x y|GOAL z )

If Icelandic has the symmetric-object operation, then we have the following order in addition: (72)

verb ( x z y|GOAL )

If a ditransitive is formed on this argument structure, then the "theme" will be the subject. By hypothesis, Icelandic is a symmetric-object language, i.e. it has statement (52). In the present framework we are almost forced to this hypothesis because there is no other way to account for the symmetric behavior displayed by the unmarked DAT-ACC verbs. Thus for the verb segja 'tell,' the lexical entry is as in (62c), repeated here: (73)

segja ( x y|GOAL z ) 'tell'

This entry qualifies for the operation in (52) and so allows both orders: (74)

a. segja ( x y|GOAL z ) 'tell' b. segja ( x z y|GOAL ) 'tell'

204

Icelandic

The order in (74a) results in (64a), and the one in (74b) results in (64b). Thus if (74a) is chosen and passive is done, suppressing the top argument, then the goal is subject as demonstrated in (65a). This order (74a) is universally available. The other order (74b) is also available, since Icelandic by hypothesis is basically symmetric for the ditransitives, i.e. has the Symmetric-Object Operation. The extra order in (74b) in the passive will have the "theme" as subject as in (64b). The two possibilities are presented in (75): (75)

a. segja-PAss ( xA

y | GOAL

z ) 'tell'

DAT: [ARG/GOAL]

NOM: [ARG]

NOM: [ARG]

b. segja-PAss ( xA

*_VP: [ARG] [ ]

_ VP: [ARG] ACC: [ ] z

ACC:

NOM: [ARG]

DAT: [ARG I GOAL]

_ VP: [ARG] ACC: [ ]

*NOM: [ARG]

y | GOAL ) 'tell'

*_VP: [ARG] ACC: [ ]

Note that in the first of the pair, the goal is the highest non-obliquely linked argument and so is the subject. In the other order (75b), the nongoal argument is the subject. Generally the lower two arguments of a ditransitive permute. This leads to the symmetric-object behavior, since either of these arguments can be the primary object (subject in the passive), depending on whether the permutation is carried out or not. The other class of ditransitives, the class with an idiosyncratic case on the lowest argument in the unmarked order, do not show symmetric-object behavior. The generalization is that verbs that assign idiosyncratic case do not allow the permutation of their lowest arguments. The idiosyncratic case is assigned to one of the arguments that might permute, so the generalization seems to be the following: (76)

An argument assigned idiosyncratic case may not permute.

This might be left as a possibly language-specific stipulation or it may reflect a general principle. I return to this question after showing how (76) would work as a stipulation. Note that as a stipulation it woulcj have much the same effect as Zaenen, Maling, and Thrainsson's (1985) language-specific rule assigning the lowest available GF to case-marked

Is Icelandic symmetric or asymmetric?

205

themes. Case-marked themes will, then, never outrank goals. In a ditransitive passive, the goal must be the subject, if there is a goal. Case-marked themes may only be the subject of the passive if there is no other argument. The behavior of the quirky case ditransitives follows from (76). For these verbs the lexical entry is like that in (77) for skila 'return': (77) skila ( x y|GOAL z ) 'return' DAT: [ARG|3]

This argument structure does not allow permutation, because one of the arguments that would be permuted is assigned idiosyncratic case. Since an entry with the reverse order of the lower two arguments is out, the grammar will not generate passives with the z argument of these verbs as subject. We only get the following goal passive, not the theme passive: (78)

a. skila-PASs ( xA

y| GOAL

z ) 'return'

DAT: [ARG/GOAL] NOM: [ARG]

*_VP: [ARG] ACC: [ ]

NOM: [ARG]

_VP: [ARG] []

ACC:

b. *skila-PASs ( xA z

y| GOAL ) 'return'

NOM: [ARG]

DAT: [ARG I GOAL]

_ VP: [ARG] [ ]

*NOM: [ARG]

*_VP: [ARG] [ ]

ACC:

ACC:

The only well-formed passive will have the goal as the subject. As for the sentences in which the goal is not expressed (69) or is expressed by a PP (70), the passivizability of the remaining lexically case-marked argument also follows. The linking for (69) is as follows. Unlike for Japanese implicit arguments, we must assume that the goal argument is not linked. One way this can be done is to follow Zaenen, Maling, and Thrainsson (1985) in saying that verbs like skila have the goal as an optional argument. Thus the revised lexical entry for skila 'return' is as follows: (79)

skila ( x

(y|GOAL)

z

) 'return'

GEN: [ARG|3]

The linking for (69) is as follows with the non-goal expansion:8

206 (80)

Icelandic skila ( xA

z

) 'return'

GEN: [ARG/3]

_ VP:

[ARG]

NOM: [ARG] ACC:

[

]

The genitive is the most restrictive and so links the non-suppressed argument. The pre-VP position is compatible and so also applies. This argument is the highest non-oblique argument (it is the only argument) and so is the subject. The PP in (70) might be said to arise in one of two ways. Either it is an adjunct added to (80), or it is linked by the preposition as an argument (cf. English to). The latter, however, is unlikely, since the preposition til does not generally link goal arguments. So I take the PP to be an adjunct PP. The interesting possibility is that the interaction of case and the symmetric-object parameter, stipulated on a language-specific basis in (76), actually follows from general principles on the present theory. What is noteworthy is that the argument permutation and idiosyncratic case both refer to one of the arguments of a ditransitive verb. The permutation operation refers to ditransitive verbs in general, while the idiosyncratic case refers to one verb in particular. It has scope over only that entry. It is possible that the blocking of the permutation by the presence of the idiosyncratic case is an instance of Elsewhere-type blocking. In this case, it would follow from a general condition which would have the RC as a special case. This general condition would compare not only LINK values but other environmental specifications. The specifications of a lexical entry with an idiosyncratic case would then properly contain those of the permutation operation: (81)

a. verb ( x y z ) - ^ verb ( x z y ) b. skila ( x (y|GOAL) z ) 'return' GEN

The idiosyncratic case will block the symmetric-object operation because it refers to the lowest argument but in a more specific way. I will not formulate exactly what should count as the environmental specification but will leave this open for further work. There is in fact good evidence from Icelandic that the permutation of arguments can be lexically stipulated, making permutation and lexical case more similar than they might otherwise appear. The present account

Is Icelandic symmetric or asymmetric?

207

also extends to the class of verbs first noticed by Bernodusson (1982: 38) which allow either the dative or the nominative to be the subject. (82)

a. t>etta this(NOM) This loan b. ]3etta this(NOM) 'Sven had

Ian hafdi Sveini stadid loan(NOM) had Sven(DAT) stood had been on offer to Sven.' Ian hafdi stadid Sveini loan(NOM) had stood Sven(DAT) this loan on offer.'

til boda. for offer til boda. for offer

In (82a) the dative NP Sveini immediately follows the finite verb like a subject, but in (82b) it follows the past participle like an object. Other subjects cannot follow the past participle (Bernodusson (1982: 38)): (83)

a. Pali hafdi Paul(DAT) has T h e teacher has b. *Pali hafdi Paul(DAT) has T h e teacher has

kennarinn hjalpad. teacher-the(NOM) helped helped Paul.' hjalpad kennarinn. helped teacher-the(NOM) helped Paul.'

Here the nominative NP is the subject, and if the object is topicalized as in (83), the subject must immediately follow the finite verb. However, with standa til boda 'be on offer,' the NP Sveini can appear either immediately following the finite verb or after the participle. To further test whether this verb, standa til boda 'be on offer', really has two possible GF assignments, we can try the subject-verb inversion test. Recall that Zaenen, Maling, and Thrainsson (1985) showed that subjects, both nominative subjects and lexically case-marked subjects, do participate in inversion but that topics do not. For standa til boda 'be on offer,' either NP can invert like a subject: (84)

a. Hefur honum nokkurn tima stadid !>etta has him(DAT) any(Acc) time(ACc) stood this(NOM) til boda? for offer 'Has he ever had this on offer?' b. Hefur Jsetta nokkurn tima stadid honum has this(NOM) any(ACc) time(ACc) stood him(DAT) til boda? for offer 'Has this ever been on offer to him?'

208

Icelandic

Other verbs that seem to work this way, i.e. allow either the dative or the nominative to be subject, are sdrna 'wound,' hugnast 'like,' among others. What this means on the present approach is that argument permutation is not just a parameter at the level of entire grammars but rather can be lexically stipulated. The verbs like standa til boda 'be on offer' and hugnast 'like' permute their arguments on an idiosyncratic basis. The small class of verbs which do this apparently has no semantic unity. Near synonyms can show different behavior. For instance, lika 'like' allows only the dative to be subject, but hugnast 'like' allows either the dative or the nominative to be subject. One generalization is that the verbs that allow either to be subject are a small class including many rare and literary sounding verbs. The way to handle this in the present approach is to say that some verbs exceptionally call for the additional symmetric argument order. So for hugnast 'like' we have the following: (85)

hugnast ( X|EXP y ) 'like' -> hugnast ( y X|EXP ) 'like'

The important conclusion to draw is that the "symmetric-object parameter" needs to admit lexical exceptions. A language cannot just select one value once and for all as in current LFG. To sum up, Icelandic is basically symmetric for ditransitives, because it has the permutation operation. Lexical case blocks this operation, possibly by a generalized RC. Furthermore, argument permutation can be stipulated by individual verbs, just like idiosyncratic case. What Icelandic demonstrates is the need for a more fine-grained approach to the symmetric-object parameter.

5.4

Coordination

One way to distinguish the predictions made by the present account of Icelandic case and previous accounts is in coordination. I first sketch out two other approaches and then contrast them with the present one on the basis of coordination data. One reason Icelandic presents such a challenge to linguistic theories is that its grammatical function behavior does not match well with its case morphology. This problem is most acute for direct linking which seeks the tightest possible connection between Case and case. If grammatical functions did not correlate with the surface morphosyntax, then it would

Coordination

209

be necessary to abandon direct linking in its strongest form. In this section, I will review some previous analyses of Icelandic. It will become apparent that the behavior of Icelandic has led to the exploitation in various theories of precisely the freedom direct linking does not allow: a separation of grammatical functions and surface morphosyntax. In the classic analysis of Zaenen, Maling, and Thrainsson (1985), the following principles for Icelandic are proposed: (86)

Icelandic Association Principles a. If there is only one thematic role, it is assigned to SUBJ; if there are two, they are assigned to SUBJ and OBJ; if there are three, they are assigned to SUBJ, OBJ, 2OBJ. (Universal) b. AGENTS are linked to SUBJ. (Universal) c. Case-marked THEMES are assigned to the lowest available GF. (Language Specific) d. Default Case Marking: the highest available GF is assigned NOM case, the next highest ACC (Universal)

This will mean that in ditransitives there will be two ways of doing case assignment (87), unless the theme is case marked (88). In a ditransitive, the lowest available GF is a 2OBJ SO the case-marked theme is a 2OBJ: (87)

(88)

gefa 'give

skila 'return'

agent

theme

goal + DAT

SUBJ

OBJ

2OBJ

SUBJ

2OBJ

OBJ

agent

theme + DAT goal

SUBJ

2OBJ

OBJ

The crucial clause is that case-marked themes are associated with the "lowest available GF." In the case of a monotransitive verb like tyna 'lose,' the lowest available is OBJ not 2OBJ: (89)

tyna 'lose'

agent

theme + DAT

SUBJ

OBJ

The theme is correctly allowed to passivize. Likewise in the event the goal is left off or expressed as a PP (possibly an adjunct) for the NOM-(DAT)DAT type, the lowest available GF is OBJ not 2OBJ, correctly allowing the passive (which converts OBJS but not 2OBJS into subjects). Interestingly, the "lowest available" GF solution is not available in new LFG. This is because the GFs are built up monotonically step by

210

Icelandic

step in Lexical Mapping Theory. There is no "lowest available" GF, only the features [ ± o] and [ ± r]. What is needed is some way to ensure that the theme of a ditransitive is [ + o] just in case there is a higher internal argument. This could be done by reference to the other argument, but this would be a unique type of classification.9 Note, finally, that despite the fact that case usually depends on GFs in LFG, the analysis of Icelandic involves the reverse: the function of the lowest argument of ditransitives depends on the case marking. Thus Icelandic seems to call for a move in the direction of direct linking. On the other hand, other GF assignments are blind to case. Case marking for most NPs in LFG depends on GF, not vice versa. More recently in direct linking, Kiparsky (p.c.) has proposed the following generalization: lexically case-marked NPs are always oblique unless linked by word-order position or argeement. Word-order position thus overrides uniformly oblique lexical case. This follows on Kiparsky's theory, because grammatical linking has priority over "semantic" linking, i.e. linking which makes reference to thematic relations. On the present approach, lexical linkers, which correspond roughly to Kiparsky's semantic linkers, may be oblique or non-oblique. Kiparsky's approach and the present one get the same results for the subject but not for the other arguments. On the override approach, the reason there are lexically case-marked subjects in Icelandic is that the lexically case-marked arguments are linked by the VP-external position. On the present approach, the non-oblique position may apply, because the lexical case is not oblique. The difference between the approaches comes in what is predicted for objects. On the override approach, there are two choices for the lowest argument in a ditransitive. One is that it is a true oblique and not linked by position, and the other is that it is linked by position and is not oblique. As mentioned earlier, the evidence for a positional linking in ditransitives is rather thin compared with English. If we accept that there is no positional linker for the third argument in a ditransitive, then the theme should be oblique with verbs like skila 'return' when the dative goal is overt. The present approach allows linking by a non-oblique case in the absence of positional linking. The lowest argument is predicted to be a plain object and not an oblique. In contrast to the LFG solution, this object is not distinguished from the other object by the label OBJ versus 2oBJor OBjtheme. Rather we have two (derived) objects, one higher and one

Coordina tion 211 lower on the argument hierarchy. In the case of the ditransitives that assign idiosyncratic case to the lowest argument, that lowest argument must be the lowest on the list. However, it will be an object just like the goal object higher on the list. To sum up, the present theory predicts that the lower object of a ditransitive should pattern like other objects. It should not behave like a 2OBJ as in LFG, nor should it behave like an oblique, as it would on the override view without a positional linker. To test the various predictions, I will now consider some evidence from coordination that the case-marked objects are objects. They are certainly not obliques and do not seem to be OBjes either. Rognvaldsson (1990) proposes the generalization that for deletion under identity in coordinate VPs, the two NPs must match in both grammatical function and case. First, an object cannot coordinate with a subject. I use the letter e to indicate the gap without a commitment as to whether there is a position there in the phrase structure or not: 10 (90)

*Eg

sa myndinai og ei gerdi mig reidan. saw movie-the(ACc) and made me(ACc) angry(Acc) 'I saw the movie and (it) made me angry.'

I(NOM)

(91)

Jon tok bokj ur hillunni og John(NOM) took book(Acc) from shelf-the(DAT) and gaf mer e^ gave me(DAT) 'John took a book from the shelf and gave me (it).'

In (90), the NP and its gap disagree in both case and GF. It would be better to take an example where the two differ in GF only and not in case, and indeed such examples are also bad, confirming GF matching as a necessary condition: (92)

*Eg hjalpadi honumj og ei batnadi I(NOM) helped him(DAT) and recovered-from veikin. disease-the(NOM) 'I helped him and (he) recovered from the disease.'

Here the first is a dative object and the second is a dative subject. For object gaps too, identity of GF and case are necessary conditions. In (92)

212

Icelandic

the object gap is dative, but the antecedent NP is accusative, and in (93) we similarly have a genitive object gap and an accusative NP object: (93)

*Eg

J>ekkti Peturi vel og hjalpadi ei oft. knew Peter well and helped [him(DAT)] often 'I knew Peter well and helped (him) often.'

I(NOM)

(94)

*Eg

elskadi £>essa stelpui og sakna QX loved this(Acc) girl(ACc) and miss [her(GEN)] mjog mikid. very much 'I loved this girl and miss (her) very much.' I(NOM)

What we want to know is whether the lowest object of the ditransitives bears the same or a different GF than normal objects. To test this we can try to coordinate such a lowest ditransitive object with regular objects. To control for case, we take dative- or genitive-case-assigning monotransitive verbs for one conjunct and have a ditransitive gap with the same case in the other. Such coordinations are well-formed, indicating that the lowest object of the ditransitive and the regular object match in GF. First, dative and genitive ditransitive lower objects can precede dative object gaps: (95)

a. Eg

het Olafi Jseinii en tyndi I(NOM) promised Olaf(DAT) them(DAT) and lost ei sidan. [them(DAT)] afterwards 'I promised Olaf them and lost (them) afterwards.' b. Eg unnti Olafi JsesSi en sakna I(NOM) not-begrudged Olaf(DAT) that(GEN) but miss ei mi. [it(GEN)] now 'I did not begrudge Olaf that but now miss (it).'

The lower ditransitive object can also be the gap, as demonstrated in the following: (96)

Eg

tyndi J>eim en het lost them(DAT) but promised Olafi ^ t>ratt fyrir J)ad. Olaf(DAT) [them(DAT)] anyway 'I lost them but promised (them) (to) Olaf anyway.' I(NOM)

More on the lexical accusative 213 If the generalization that the gap must match in GF with the overt coreferential NP is correct, then the lower objects of ditransitives must be objects, not second objects or obliques. There may be ways to change the various analyses of coordination or monotransitives in order to cover this data. For instance, we might say that monotransitives like tyna 'lose' can optionally take an OBJ0 or an oblique. However, there is no other evidence to support this. Without modifying existing analyses of gapping or monotransitives, the facts do support only the present approach. The idiosyncratic case-marked objects of ditransitives are neither 2OBJS (or OBJ6S) and definitely not obliques. Icelandic does have an [ARG] linker which is both non-oblique and compatible with any lexical case, i.e. the pre-VP position. On the analysis above, lexical cases are not oblique. The result is that subject and pre-VP position seem to correlate. This is quite common cross-linguistically. Kiparsky (p.c.) proposes as a universal that lexical case is always oblique unless overridden by word order or agreement. The present approach, however, allows functionally non-oblique NPs (terms) linked by lexical cases and that are not linked as well by position. Because of a positional linker that has [ARG] (but "highest role" in Kiparsky (1987)) and that is compatible with any case, one cannot tell the two approaches apart on subjects. It is not clear from Icelandic whether the fixed position is a necessary condition for the behavioral subjecthood of lexically casemarked NPs as Kiparsky predicts, or not. 11 In Icelandic, the only way to test these divergent predictions was on the lower objects of idiosyncratic case-taking ditransitives. I argued that, under the standard view of gapping and of the objects of monotransitives, the lower objects of ditransitives are plain objects rather than being obliques (or OBJ0S). I will return to the question of the interaction of lexical case and configuration in the next chapter in the discussion of diachronic change. 5.5

More on the lexical accusative

With the present theory we can also partially account for the presence or absence of lexical accusative. The peculiarity of lexical accusative is that it appears like syntactic accusative but shows different linking properties. Like other lexical case, it is preserved under passive and raising. In this section I would like to show how the restrictiveness-based approach accounts for the presence or absence of lexical accusative in certain morphologically related verbs.

214 5.5.1

Icelandic Restrictiveness and the evaluation metric

The notion of restrictiveness can do more than decide for the grammar which linker has priority or which linkers dissimilate. Restrictiveness by hypothesis also functions on a higher, metagrammatical level as a key component of the evaluation metric. The link between restrictiveness and the evaluation metric will need to be something like the following, which correlates restrictiveness of the linker for an argument in the unmarked case to complexity: (97)

Optimality Hypothesis: If Grammar A links a given argument by a less restrictive linker than Grammar B in a simplex clause, Grammar A is simpler than Grammar B relative to that argument.

The more restrictive the linker applying to a given argument, the less favored the grammar. If two grammars are observationally adequate they generate the same desired set of strings - the partial metric in (97) provides a way to decide between them. This is quite necessary, since it allows us to rule out many undesired grammars. If Grammar A has a syntactic nominative, but Grammar B achieves the same distribution of nominative with many lexically stipulated nominatives, then some redundancy has not been factored out. The metric in (97) captures this, favoring the less redundant syntactic nominative. In general, less restrictive cases are less redundant and organize more information. Ultimately, a statement such as (97) may be derivable from an understanding of human language processing, or (97) may turn out to reflect psychological reality. For present purposes, all that has to be kept in mind is that (97) is a hypothesis based on case-marking data and is formulated independently on the basis of the formal properties of linkers. 5.5.2 Lexical accusative in related verbs In Chapter 2 we saw that even Dative Substitution does not violate the generalization that verbs do not assign the disjunction of two lexical cases directly. And it was shown that the present theory can handle complex case alternations by means of simple optionality. Zaenen and Maling (1984) document a similar situation with morphologically related verbs. These morphologically related verbs assign at most one "quirky" (idiosyncratic) case to a given argument, and they demonstrate that morphologically related verbs have the option of assigning that case. They cannot assign different quirky cases. Thus one class of verbs, represented

More on the lexical accusative 215 by fcekka 'decrease,' shows lexical dative on the "theme" argument in both the active and the related unaccusative (Zaenen and Maling (1984: (11)): (98)

a. Flugfelagid faekkadi ferdunum um Jmdjung. airline-the decreased trips-the(DAT) by one-third T h e airline decreased the number of trips by one third.' b. Ferdunum faekkadi um fcridjung. trips-the(DAT) decreased by one-third T h e number of trips decreased by one third.'

In (98) dative is assigned to the object in the first sentence and to the subject in the second. However, in both sentences, the dative is assigned to the "theme." On the basis of evidence like this, they propose that lexical case is sensitive to thematic role, although this is individual thematic role, not thematic role type. Maling and Zaenen note that no pair of morphologically related verbs call for two different quirky cases on the same role. So there are no pairs like (98) with dative for the theme in the transitive but genitive for the theme in the related unaccusative. Thematic roles, however, are not the only way to index arguments. Without thematic roles, the generalization is that no argument may be assigned more than one idiosyncratic case. The second type of verb pair has the idiosyncratic case with only one member of the pair. Zaenen and Maling use sokkva 'sink' as a typical example: (99)

a. Skipstjorinn sokkti skipinu. captain-the(NOM) sank ship-the(DAT) T h e captain sank the ship.' b. Skipid sokk. ship-the(NOM) sank T h e ship sank.'

In the transitive the theme object gets the dative, but in the related unaccusative the theme appears as a nominative. To capture these two types of pairs Zaenen and Maling (1984: (13)) propose the following lexical entries:

216 (100)

Icelandic a. faekka ( (Agent), Theme ) 'decrease' DAT b. sokkva ( Agent, Theme ) <=> sokkva ( Theme ) 'sink' DAT 'sink'

For the first type (100a), the agent is simply optional, but, with either option, the theme is still dative. In the second type (100b), the two verbs of the pair are related by redundancy rule. The dative is assigned only on one of the pair. In the present framework, these entries can be reconstructed at argument structure as follows: (101)

a. faekka ( (x)

y ) 'decrease' DAT

b. sokkva ( x y ) 'sink' <=$ sokkva ( y ) 'sink' DAT

Maling and Zaenen provide evidence that the first class involves one lexical entry and the second two entries. The justification is that in general pairs in the first class are morphologically identical, whereas most verbs of the second class are related but not identical: sokkva, so'kk, sokkid (intransitive) versus sokkva, sokkti, so'kkt (transitive). One of the main questions about these lexical case alternations centers around the behavior of the lexical accusative. Interestingly, even unaccusatives that do show the lexical accusative and so should be compared With, fcekka 'decrease' show regular nominative subject passives (Zaenen and Maling 1984: (16)): (102)

a. Stormurinn bles strompinn af husinu. storm-the(NOM) blew chimney-the(Acc) off house-the(DAT) 'The storm blew the chimney off the house.' b. Strompinn bles af husinu. chimney-the(Acc) blew off house-the(DAT) 'The chimney blew off the house.' c. Strompurinn var blasinn af husinu. chimney-the(NOM) was blown off house-the(DAT) 'The chimeny was blown off the house.'

Zaenen and Maling are quick to point out that this peculiar behavior cannot be attributed to any putative ability of the passive to absorb accusative case, since lexical accusative (but not syntactic accusative) is

More on the lexical accusative 217 preserved under raising and subsequent passivization (Zaenen and Maling 1984: (17,18)): (103)

a. Batana hefur brotid i spon. boats-the(Acc) has broken in pieces(ACc) 'The boats have broken into pieces.' b. Allir telja batana hafa brotid all(NOM) believe boats-the(ACc) to-have broken i spon. in pieces(Acc) 'Everybody believes the boats to have broken into pieces.' c. Batana er talid hafa brotid i spon. boats-the(Acc) is believed to-have broken in pieces(ACc) 'The boats are believed to have broken into pieces.'

But the syntactic default accusative in (104b) on krakkana 'kids' is not preserved under passivization: (104)

a. Krakkarnir hafa brotid batana i spon. kids-the(NOM) have broken boats-the(ACc) in pieces(Acc) 'The kids have broken the boats into pieces.' b. Allir telja krakkana hafa brotid batana all(NOM) believe kids-the(ACc) to-have broken boats-the(Acc) i spon. in pieces(ACc) 'Everybody believes the kids to have broken the boats into pieces.' c. Krakkarnir eru taldir hafa brotid batana kids-the(NOM) are believed to-have broken boats-the(ACc) i spon. in pieces(ACc) 'The kids are believed to have broken the boats into pieces.'

Zaenen and Maling propose to capture this alternation by breaking cases into features. Their framework is LFG, and the features are functional features, [±oBL(ique)] and [±OBj(ective)], features very much like the [ ± r] and [ ± o] of current Lexical Mapping Theory. The lexical accusative and the syntactic accusative both get the same feature-value pairs: [-OBL, + OBJ]. Lexical cases are assigned to arguments on the basis of thematic role. This assignment, however, is subject to a matching requirement with the grammatical functions: only marked values of case features get realized. For the OBJ function this means that the one lexical case that

218

Icelandic

gets no realization is the accusative. Thus if a theme is associated with OBJ and it is assigned lexical accusative, that case is not realized and the default ACC applies. If, however, this theme is associated with the subject function then it is realized since ACC [—OBL, + OBJ] is marked with respect to the subject function, because [ + OBJ] conflicts with the [-OBJ] value of subject. What this proposal crucially relies on for (102) is the earlier LFG rule of passive. The difference between (102c) which does not preserve accusative in the passive and (103c) which does is the initial grammatical function associated with the theme. In (102c) this is the OBJ, but the passive rule changes this to SUBJ. Since the initial GF is OBJ, the ACC is unmarked with respect to it and so will not be realized. In (103c), however, the theme is initially a SUBJ. Since the accusative is marked with respect to this GF, it will be realized. In current LFG, however, there are no initial GFs but rather underspecified ones at argument structure. The passive does not change GF labels but changes the way the GFs are built up. So at argument structure there is no longer a contrast between (102c) and (103c). One possibility in Lexical Mapping Theory would be to make lexical accusative realizable only on arguments specified [-o] at argument structure. This, however, would run into problems with the Dative Substitution verbs like vanta 'lack' which has a lexical accusative (105) or dative (106): (105)

a. Mig brestur kjark. me(ACc) lacks courage(ACc) 'I lack courage.' b. Mennina f>rytur mat. the-men(Acc) lacks food(Acc) 'The men lack food.' c. Mig vantar hnif. me(ACc) lacks knife(Acc) 'I lack a knife.'

(106)

a. Mer brestur kjarkur. me(DAT) lacks courage(NOM) 'I lack courage.' b. Honum ]3raut Jsrottur. him(DAT) lacked strength(NOM) 'He lacked strength.'

More on the lexical accusative 219 c. Mer vantar hnif. me(DAT) lacks knife(ACc) 'I lack a knife.' Note also in (106c) that vanta still takes an accusative on the second NP. This lexical accusative wins against a competing nominative, as demonstrated by (106a,b). In the Lexical Mapping theory, the intrinsic classification of the second argument of vanta is either [ + o] or [—r], with respect to neither of which the accusative is marked. But if this accusative is syntactic, it is strange that it is giving way to the nominative, since syntactic accusatives and nominatives usually do not alternate. In the present theory, it is only possible for a lexical accusative to block the nominative here, because syntactic accusative is less restrictive than the nominative. If, on the other hand, Lexical Mapping Theory tried to use later stages of GFs for the assignment of lexical case, the distinction between initial objects and subjects, which Zaenen and Maling rely on, is obscured. That is, at f-structure (102c) and (103c) do not differ. First, take the intransitive-transitive pairs represented by fcekka 'to decrease' and sokkva 'sink' above in (100). Zaenen and Maling seem to assume that the intransitive is derived from the transitive since the intransitive sometimes lacks the lexical case that the transitive has. The intransitive is said to "preserve" the lexical case marking sometimes and sometimes not. However, the transitive member of the pair is the one with regular morphology. If we take this as a sign of its derived status, then it is best to regard the transitive as derived. So in (101b) the transitive adds the specification for dative. If we then apply this approach to the accusative pairs, we get (107) with the transitive derived from the intransitive: (107)

blasa ( y ) 'blow' -> blasa ( x ACC

y[-Acc] ) 'blow' ACC

The transitive is marked as an exception to the accusative, but how this is done is a matter of execution.12 The main point is that the transitive does not share the ability to assign idiosyncratic case. If the transitive is derived from the intransitive, then the lexical accusative assigning ability must be blocked or taken away. We will need exception features elsewhere, so their use in (107) is not new. Further, the exception feature adds information rather than removing it.

220

Icelandic

The question now is why a grammar with the exception feature (or its equivalent) should be favored. The difference between this entry and the one for sokkva is that, for this pair, the idiosyncratic case is associated with the intransitive member of the pair. The transitive on the right will show no case preservation since in effect there is no lexical case to preserve. The intransitive on the left will always show preservation, since it does have the idiosyncratic case which is more restrictive than the nominative. This is descriptively adequate, but the question arises why we get the pattern in (107) exactly for the lexical accusative. The answer may lie in trends that I will discuss at greater length in the next chapter. Over time, the case applying to a given argument of a given verb decreases in restrictiveness. Thus the addition of an exception feature to a grammar, which is an apparent complication, is motivated if its effect is to minimize the restrictiveness of cases applying to the argument with the exception feature. Thus the exception feature in (107) is motivated diachronically and in language acquisition, because it minimizes restrictiveness. In simplex clauses, there is no difference in the surface transitive whether it has a lexical accusative or a default accusative. Only in more complex clauses involving passive and raising do the two accusatives make a difference. The change, as I will argue, is based on minimizing the restrictiveness of applying linker in underived clauses, as stated in the Optimality Hypothesis. In simple transitive clauses, the accusative on the object could be lexical or syntactic, but the optimal grammar is one in which the default accusative's application is maximized. So despite the larger number of symbols in (107) than in an entry without the exception feature, the optimal grammar has the exception feature to maximize application of the default. The simplicity metric is correlated not just with symbol content but also with restrictiveness. The optimal grammar is the one that minimizes restrictiveness of applying cases consistent with the surface patterns of simple clauses, and such a grammar will have exactly the behavior of lexical accusative that Icelandic displays.

6

Changes in linking

In this chapter I would like to turn to the diachronic consequences of the theory outlined in the previous chapters. I intend to show that the approach to linking based on restrictiveness provides an elegant way to state generalizations about the relative likelihood of changes. The Optimality Hypothesis introduced in the last section of the last chapter defines what a local simplification would be: the lowering of the restrictiveness of linkers applying to a given argument. If change in linkers is local simplification, then we can, by way of the Optimality Hypothesis, identify decreases in the restrictiveness of linkers applying to arguments as the likely changes. The organization of this chapter is as follows. First, I will consider some of the issues raised in some previous work on change in linking. Then, I will return to the topic with which I began Chapter 2, case alternations. Like much, if not all variation, the case alternations are changes in progress. In §6.3 I will lay out the predictions the restrictiveness-based approach to linking makes for linking change. Then I will show that the diachronic side of the case alternations is exactly what is expected on this view. I will then turn to a more general consideration of the consequences of the theory and provide an account of linking changes in Scandinavian in particular. Finally, I will consider the general character of changes in linking.

6.1

Previous work on linking change

Previous work in case change has raised three major issues that I would like to address in this chapter in the context of the restrictiveness-based theory of linking. These are the directionality of the changes, their gradualness, and the order of certain changes before others. 221

222

Changes in linking

6.1.1 Directionality Much work has been done in the last century on case and word-order change. One of the most significant discoveries in this area is that many such changes have a directionality. The names most associated with this type of work on case change are Meillet (1912) and Kurylowicz (1964, 1965). Many changes involve the change of a lexical item into a grammatical formative. Changes from lexical item to grammatical formative were studied throughout the nineteenth century. This very general type of change is often called "grammaticalization," and it is Meillet (1912) who first proposed this term (see Traugott and Heine (1991) for references to more recent work on grammaticalization). One very well-known way that grammaticalization effects changes in linking is when some independent lexical item becomes a postposition and then a case marker (for a lengthy discussion, see Lehmann (1985) and references there). Kurylowicz (1964) focuses on a particular aspect of grammaticalization: the changing functions of a given case form. Basic to Kurylowicz' view of case change is the distinction between the relations that a case expresses and the form of the case itself. The relations that cases express come in two varieties, "concrete" and "grammatical." Case forms often express more than one relation. This should already be familiar from previous chapters. For example, in Icelandic a morphological accusative is the case of the "object" and an idiosyncratic subject case. Kurylowicz distinguishes two kinds of cases: semantic and grammatical. Semantic cases express "concrete" forms as their primary function. Grammatical cases express non-concrete, grammatical relations as their primary function. The generalization Kurylowicz proposes is that semantic functions give way over time to grammatical functions. Or as Lehmann (1985), building on Kurylowicz, phrases it, grammaticalization is "the process which turns lexemes into grammatical formatives and makes grammatical formatives still more grammatical." Very interestingly, Meillet also proposed that a use of a word-order position could be the result of grammaticalization. Meillet realized that word order often performs the same function as grammatical cases. For instance, in English, pre-VP position correlates with subjecthood just as does agreement. Word-order is not a morphological item, so that at first it might seem strange to speak of grammaticalization involving word order. But, as we have seen, word order is a morphosyntactic device in the same sense as case and agreement. Word-order positions are a dif-

Previous work on linking change 223 ferent type of linker, but they are linkers. Most importantly, they share the same morphosyntactic behavior as cases and agreement. In the present theory, this means word-order positions - or more strictly speaking the features like _VP assigned to arguments which constrain the positional appearance of those arguments - have LINK values just like other linkers. I have even argued that positional linkers can even be oblique, providing they refer to lexical information (see the discussion of Dutch and Grisons German in §4.4 above). Since they have LINK values like other linkers, positions will behave the same way as other linkers, because morphosyntactic behavior is determined by the three constraints, AC, LDC, RC, which all three refer to LINK values. In the following, the account of change will be about the syntactic behavior of linkers and so will be stated on LINK values. The changes will be much more likely in one direction, decreased restrictiveness of applying linker(s). Since the predictions are stated over LINK values, word-order position will behave diachronically like grammatical case.

6.1.2 Unity of different types of changes As noted in the last section, the shift from semantic to syntactic case is not the only change that could be described as a change in linking. First of all, linking change can involve agreement and word order, the other two main types of linkers, as well as case. Changes in different types of linkers show certain similarities, which raises the question of how to provide a unified explanation for the various types of linking change. There have been at least two major approaches to the question of how to provide a unified account of linking change. I will term them the cluster view and the tendency view. The first view, the cluster view, seeks to explain patterns in the co-occurrence of multiple changes. If an individual change can be shown to be one of a group of changes which necessarily co-occur, then the individual change is no longer an isolated unexplained fact. The explanation for the group of changes lies then in the trigger for the whole group. Changes in case, agreement, and word order have been approached from this point of view by Jespersen (1927) and more recently by Lightfoot (1979, 1981, 1988, 1991). Jespersen seeks to unify the loss of case and agreement and the fixation of SVO word order. The reanalysis is illustrated with the following hypothetical examples:

224 (1)

Changes in linking a. ]}am cynge licoden peran the(DAT) king liked(3pl) pears b. the king liceden peares c. the king liked pears d. he liked pears.

In the first stage, there is a distinctive oblique case, and the verb agrees with the postverbal NP identifying it as the subject. In the second stage, the NPs are not distinguished by case, but the verb still agrees with the postverbal NP. When, in the next stage, agreement is lost, the subject and object are potentially confused. This leads to the reanalysis of the construction, with the preverbal NP as the subject. This leads to nominative pronouns appearing in this position. More recently, Lightfoot (1979) has proposed a similar reanalysis account based on a precursor to the Principles and Parameters framework. Essential to his (1979) account is the Transparency Principle. This is a metagrammatical principle which defines the upper bound of derivational complexity. When such a bound is reached, reanalysis must take place. In the example above, when the inflections are lost and a DStructure verb-object order arises, the derivation of the sentence involves movement of both NPs. The derivation is simplified by fixing SVO word order, and no movement is necessary beyond movement to get Case. Later Lightfoot (1981) uses the Trace Erasure Principle to motivate the reanalysis. At the stage where case is lost and the D-Structure order is verb-object, both NPs must move from their D-Structure positions. The theme must move from preverbal position to postverbal position, and the experiencer must move from inside the VP to a position before it. The problem is that the experiencer NP will come to occupy the position of the trace of the theme NP's movement, thereby "erasing" it. Since such erasures are disallowed by the Trace Erasure Principle, the reanalysis is forced. The cluster view often locates the ultimate cause of the clusters of syntactic changes in the phonological loss of case endings. This general view has become very popular in handbooks, where case loss and rigidification of word order are usually attributed to ultimate phonological factors. The other view, the tendency view, seeks the common denominator of changes. All such changes can then be explained together by appeal to the common denominator. The common denominator replaces individual accounts of each of the changes. This view is both looser and stricter

Previous work on linking change 225 than the cluster view. It is looser in that the various changes in a group of changes need not cause one another, nor need there be a single trigger for every group of changes. The tendency view is stricter than the cluster view as far as it requires there to be some element common to the changes. For one change to cause another, they do not need to share any property, but for two changes to be instances of some general tendency, there must be some common element, the common denominator, between them. One difference between the cluster view and the tendency view is in how sudden the changes in question are predicted to be. The essence of the cluster view is that what is to be explained are co-occurring changes. These changes could in theory co-occur and extend over a long period of time. But if a cluster-type explanation is sought through a reanalysis mechanism or a parameter change, then the changes should be characterized almost defmitionally by rapid S-curves of innovation, as Lightfoot (1991) puts it. On the other hand, under the tendency view, there is no such requirement that changes be rapid. Changes can both be instances of some general tendency without co-occurring and without being quick. A related difference between cluster and tendency explanations is in what the unit of linking change should be. If linking change is a form of reanalysis, then we expect linking changes to occur primarily at the level of the construction or of whole languages. If a change is the resetting of a parameter, then the effects of the change should be as general as the effects of the parameter setting that has been changed. If, on the other hand, linking change is relativized to individual elements like cases in Kurylowicz (1964, 1965), then the tendency can be fulfilled in a very local way. An individual case changing from semantic functions to syntactic functions fulfills Kurylowicz' generalization. There is no expectation that such change should characterize whole languages, because the shift from semantic to syntactic functions is relative to forms like prepositions and cases. At the level of whole languages there is a sort of equilibrium. The shift of individual adpositions and cases from semantic functions to syntactic functions is offset by the creation of new adpositions and cases to express the semantic functions (and the loss of earlier syntactic cases). The tendency is a tendency of individual morphosyntactic items. Both types of change no doubt occur. Reanalyses are familiar, and so are unidirectional changes, where we have some idea why change in the one direction is more "natural" than in the other. In the following, I will offer a tendency account of linking change based on the theory developed

226

Changes in linking

in the previous chapters. The focus will be on the changes in the linking of arguments. Along the way I will point out other types of case change that may be reanalyses. The account of change will be relativized to particular arguments. At the end of this chapter, I will consider how the present approach to changes in argument linking fits into the more general picture of grammaticalization and analogy.

6.2

Diachronic aspects of case alternations

I will briefly present some further facts of English and present Lightfoot's most recent (1991) analysis as the latest example of the "cluster approach" to the loss of the impersonals in English. Since the loss of the impersonals is taken to be the result of the loss of lexical case marking, one expects these changes to cluster. However, they dramatically do not, as we shall see. It was noted in passing in Chapter 2 that the variation involved in DS and NS has a diachronic dimension as well. DS is a shift toward dative, NS a shift toward nominative. A complete account of the case alternations should include a general historical explanation as well. In fact, the diachronic aspect of these changes is of particular theoretical significance because what I have called "Nominative Substitution" reflects one of the most studied changes in the history of English, the shift from impersonal to personal constructions. The shift from impersonal to personal constructions in English has been intensively investigated by many (van der Gaaf (1904), Jespersen (1927), Visser (1963), Lightfoot (1979, 1981, 1988), Elmer (1981) Fischer and van der Leek (1983, 1987), Allen (1986), von Seefrantz-Montag (1984), J. Anderson (1986), Denison (1990a, b), and Traugott (1992)), and, in the Principles and Parameters framework, by Lightfoot (1991). I will briefly present the facts of English and present Lightfoot's analysis as an example of an explanation of the loss of the impersonals as the result of the loss of lexical case marking. An often cited set of examples involve ofhreowan 'pity' in (2): (2)

a. him ofhreow £>aes mannes him(DAT) was-pity the(GEN) man(GEN) 'He was sorry for the man' (ACHom I 13 192.16)

Diachronic aspects of case alternations

227

b. se maessepreost fcaes mannes ofhreow that(NOM) priest that(GEN) man(GEN) pitied 'the priest took pity on the man' (JELS (Oswald) 262) Like many verbs in Old English, ofhreowan 'pity' could be impersonal or personal. 1 Eventually, for the Old English impersonals, the dative or accusative gave way to nominative on the experiencer NP as with hreowan 'regret': 2 (3)

a. 3if him his yfel ne if him(DAT) his evil(NOM) not 'if his evil doesn't grieve him' b. Hreow hine swide pitied him(ACc) very-much 'It grieved him very much that c. He rues his deed.

hreowd rue (s£.Boeth.37J) Jsaet . . . that ...'

(GenA,B 1276)

Lightfoot follows many others in reducing the change in (2-3) to the loss of morphological case. Lightfoot seems to claim that presence of lexical case is both necessary (p. 140) and sufficient (p. 137) to trigger impersonals like (la-b), as in Old and early Middle English. When morphological dative was lost in Middle English, lexical case was not triggered for the child learner, and the impersonals had to drop out of the speaker's grammar. Interestingly, Lightfoot (1991) claims that a parameter resetting approach handles obsolescence, such as that of the impersonals, in a uniquely adequate way. He contrasts his approach with a form of lexicalist approach in which the obsolescence of an item involves the addition of a lexical exception feature. Lightfoot correctly points out that this use of exception features involves positing the use of negative evidence in language acquisition, which by a common hypothesis does not occur. So, according to Lightfoot, lexicalism would lead to unattainable grammars. But the assumption that the obsolescence of an item is to be captured by the isolated adding of a feature (in contrast to adding one that maximizes application of a default; see Section 5.5.1) is by no means a necessary one for lexicalism, and, as we will see, a lexicalism without this assumption is quite capable of handling obsolescence better than Lightfoot's own account. Lightfoot himself notes that "a striking property of this change is its gradualness" (p. 137). This is a problem for parameter-setting, since the

228

Changes in linking

latter is expected to occur abruptly. Instead, to save the parameter setting account of obsolescence, he posits a "diglossic" situation with two coexisting grammars. Lightfoot criticizes the use of optionality in grammar but never explains why diglossia is more constrained. After all, each "option" of two or more options can just be surrounded with a grammar and the grammars placed in a diglossic relationship with each other. The loss of morphological case cannot, however, be a general explanation of the NS type of shift. NS in both Icelandic and German are partial shifts of this sort, and in German the shift is complete with some verbs.3 One verb involved is ahnen 'apprehend' (for other verbs see, e.g., Curme (1922), Behaghel (1924), Dal (1966)). In Middle High German, the corresponding verb an(d)en 'apprehend' could, according to all sources, take either accusative (4a) or dative (4b): (4)

a. mich andet me(ACc) apprehend 'I apprehend . . . ' b. mir anet me(DAT) apprehend 'I apprehend . . . '

{Tristan 236.1)

(Herbort 9592)

In the eighteenth century, an archaizing writer like Klopstock (17241803) uses accusative, but Lessing (1729-1781) uses dative. Goethe (1749-1832) uses the dative, but also commonly nominative: (5)

Ich

ahnt' es! apprehended it 'I apprehended (felt) it.'

I(NOM)

(9,138,4 Geschw.)

In present-day German, for most if not all speakers the verb ahnen 'perceive' takes either nominative or dative. Thus for this verb (as earlier for zweifeln 'doubt'), the period in which German shows both impersonal and personal constructions for this verb extends over at least several centuries.4 For Old Icelandic, the handbooks usually do not mention the possibility of a dative with any accusative impersonal (e.g. Nygaard (1905)). Halldorsson (1982), in a search for the earliest occurrence of DS and NS with a variety of verbs, finds that in Old Icelandic only one verb, skorta 'need, lack,' ever shows dative (6), in addition to the usual accusative (7) (Halldorsson (1982: 175), see also Cleasby and Vigfusson (1957: 555)):

Diachronic aspects of case alternations

229

(6)

a. EN allt \>at er honom scort/r i . . . but all that REL him(DAT) lacks in . . . 'but all that he lacks . . . ' (Grdg (Kon.): 242) b. EN alt Ipat er honom scort/r i . . . but all that REL him(DAT) lacks in . . . 'but all that he lacks . . . ' (Grdg (Stb.): 92)

(7)

mik skortir eigi hug me(ACc) lacks not courage(ACc) 'I do not lack courage'

(Eg 719)

An identical example can be found in the same text, but Halldorsson (1982: 175) gives reason to believe that this isolated example in (6) does not reflect native Old Icelandic usage. 5 Note, anyhow, that skorta 'need, lack' too is a psych verb, conforming to the generalization that DS involves experiencers. Note, too, that this NP is most probably oblique, since lexically case-marked NPs generally do not exhibit subjecthood behavior in Old Icelandic. 6 Other verbs which later show DS do not show it in Old Icelandic:7 (8)

a. eigi brestr mik araedi not lacks me(Acc) courage(Acc) 'I do not lack courage.' (Fs. 62) b. illt er J)at ef fodur minn Jsrytr drengskapinn ill is that if father(Acc) my(ACc) lacks courage-the(ACc) 'It is bad if my father lacks the courage.' (LvA 1) c. vantadi \>k eigi hesta ok adra lacked them(Acc) not horses(ACc) and other-pl(ACc) hluti things(ACc) 'They lacked neither horses nor other things.' (ims.iii.77)

These are the Old Icelandic counterparts of the Modern Icelandic verbs in (23-24) of Chapter 2. In Old Icelandic, only the original accusative is possible. DS sentences like those in (24) of that chapter are not attested, nor are they attested with other accusative-experiencer verbs aside from the one possible example with skorta 'need, lack' in (6). As in German, DS in Icelandic itself has not always correlated with the special subject status of oblique case marked NPs. At least one possible oblique NP in Old Icelandic may have undergone DS.

230

Changes in linking

NS, however, is more common in Old Icelandic.8 Halldorsson (1982) documents the use of nominative with a variety of accusative and dative experiencer psych verbs, as illustrated in (2) of Chapter 2 for Modern Icelandic and in the following for Old Icelandic: (9)

mi er sa dagr kominn er ver hofum now is that day come REL WC(NOM) have allir langat til all(NOM) longed for 'Now that day has come that we have all longed for' (Fms. viii.220)

This is an Old Icelandic analogue of the Modern Icelandic sentence in (2) of Chapter 2.9 For further comparison, let us turn briefly to the situation in Gothic, which does not show DS during the period for which we have records. Gothic experiencer verbs mostly have accusative or nominative, rarely dative. The cognates of the DS verbs uniformly take accusative. The class includes gredon 'be greedy, be hungry,' huggrjan 'hunger,' and paursjan 'thirsts': (10)

J)ana gaggandan du mis ni huggreij) jah ]sana that(Acc) going(ACc) to me NEG hungers(3sg) and that(ACc) galaubjandan du mis ni t)aursei£> luanhun believing(ACc) to me NEG thirsts(3sg) ever 'He who comes to me never hungers and he who believes in me never thirsts' (Jo. 6.35)

The accusative experiencer verbs show no tendency toward dative, i.e. the dative is not attested with these verbs.10 At least one verb, pugkjan 'seem' took a dative experiencer, but it shows only dative or nominative. In (11) we have the dative: (11)

Imgkeit) im auk ei in filuwaurdein seinai seems them(DAT) also that in much-talking their andhausjaindau hear-PASs 'It seems to them also that in their talking much they are heard' (Mt. 6.7)

The use of the dative is probably native Gothic here for two reasons. One is that it goes against the Greek original (as does the accusative in (10)),

Diachronic aspects of case alternations

231

which has a personal (nominative) construction. The second is that the older Germanic languages generally have dative, which suggests that the dative with this verb goes back to Proto-Germanic. 11 Returning to German and Icelandic, the diglossic approach cannot work for two reasons. First, since Lightfoot seems to deem loss of lexical case both necessary and sufficient for the loss of the impersonals, the loss of impersonals in Icelandic and German should be accompanied by a loss of lexical case. But Icelandic does not show much, if any, loss of lexical case since Old Icelandic times, and so this cannot be the trigger for a reanalysis there. Nonetheless, Icelandic shows some of the shift. Even more clearly, German has relatively healthy dative case but has completed the shift with some verbs. Lightfoot might retreat from identifying loss of lexical case as a necessary cause of the loss of the impersonals. That leaves us, however, with the very unremarkable proposition that if morphological dative is lost, then there can be no syntactic dative and so no verbs that assign it. More importantly, German and Icelandic point up a major flaw in Lightfoot's account of English. One implication of the diglossic account is that "once a grammar without dative case (and thus without truly impersonal verbs) is attested, one might find any of the impersonal verbs with nominative and accusative NPs" (p. 137). In other words in the diglossic period all of the formerly impersonal verbs should manifest both the impersonal and the personal forms. Not all forms are actually attested because "given the haphazard and fragmentary attestation for any given period, one could not expect to find all the relevant verbs attested in both construction types at any given time" (p. 137). Given that modern German and Icelandic sources of data have no similar limitations, we expect that if Lightfoot's diglossic account is to work there, the lexical case verbs should all alternate with nominative. This is patently false. For example, at any given stage, only a proper subset of the impersonal verbs in German can take nominative and accusative. For instance German gefalien 'please, like' is ungrammatical with nominative and accusative: (12)

a. Mir gefallt der Tisch. me(DAT) pleases the(NOM) table 'I like the table/The table pleases me.' b. *Ich gefalle den Tisch I(NOM) like the(ACc) table

232

Changes in linking

Clearly Lightfoot cannot say that German is diglossic like Old English. Similarly, Icelandic speakers with NS do not generally have it across the board. But then we are left with no explanation of the obsolescence of (morphologically healthy) lexical case with any of the impersonal verbs in German and Icelandic. Furthermore, Lightfoot (1991: 139) claims that one advantage of his account over Fischer and van der Leek's (1983, 1987) is that they incorrectly predict personal passives with all the "impersonate" as soon as lexical case assignment is optional. Leaving aside the question of whether there are good examples of personal passives of these verbs in Old English, it is clear nevertheless that Lightfoot's diglossic solution fares no better. As soon as the diglossia arises, we should expect that the lexical case-less grammar should be generating plenty of personal passives of the old impersonal class. Lightfoot is incorrect in asserting that his theory predicts that impersonal verbs would "occur in such constructions [personal passives] only after the loss of lexical case" (p. 138). Rather, as soon as the diglossic situation arises we expect to find them, long before lexical case has been completely lost. A conjunction of two grammars in diglossia presumably generates the union of what the two grammars would generate on their own. Optionality and diglossia are not as different as Lightfoot supposes. In predicting too wide a range of data, Lightfoot's account actually shares a basic problem with Fischer and van der Leek's original solution. The problem is that both he and they expect lexical case assignment to be uniform. Lexical case is either uniformly obligatory or uniformly optional for all verbs that can assign it. For the impersonate we then expect lexical case alone, either lexical case or nominative, or no lexical case across the class of "impersonate." But German and Icelandic show this to be an unrealistic expectation: one verb may be impersonal only (gefallen 'please, like') while the next is varying (ahnen 'perceive') and another used to be impersonal but no longer is (zweifeln 'doubt'). This is not so surprising once one recalls that what is involved here is lexical case marking. Since generalizations about case marking are lexical, we expect them, like other lexical aspects of grammar, to be characterized by exceptions. Lightfoot is quite right that capturing a historical change as the addition of lexical exception features is not very enlightening. But there are other better ways to handle lexical change, along the lines of a relatively well understood type of change, analogy.

A theory of analogical change in linking

233

In explaining DS, it is necessary to avoid appeal to properties of any one language not shared by the others. Otherwise the explanation will not generalize and we will need a separate, and otherwise superfluous, explanation for another language. It has been suggested (Collberg (1986: 90ff.)) that DS is related to the functional subjecthood of oblique NPs in Icelandic. As should be apparent by now, this cannot generalize, since German does not have oblique subjects but does show DS. Similarly, case syncretism and fixation of SV word order have been taken to "explain" NS (Jespersen (1927), Lightfoot (1979, 1981, 1988)). However well this works for English (see especially Fischer and van der Leek (1983) and Allen (1986) for criticism), it cannot work for Icelandic which does not show much, if any, case syncretism since Old Icelandic, nor for German which has relatively healthy case and does not have fixed SV order. Nonetheless, Icelandic and German both show large amounts of NS and from a very early stage. Even Gothic arguably shows NS. Further, as we have seen, at any given stage DS and/or NS is a property of verbs, not whole languages. At any given time, only some of the verbs that might show DS or NS do so. 6.3

A theory of analogical change in linking

It was demonstrated in the previous section that the variation involved in DS and NS has a diachronic dimension. DS is a shift toward dative, NS a shift toward nominative. A complete account of the case alternations would include a general historical explanation as well. This section will provide a diachronic theory from which will follow an account not only of DS and NS as historical trends but also a wide range of other changes in linking. The main prediction is that the restrictiveness of the case applying to a given argument should decrease over time, a prediction which is fulfilled by both DS and NS. One of the main generative approaches to explanation in historical linguistics has been to locate the motor of grammatical change in simplicity (see Kiparsky (1982)). The notion of possible or likely change is tied to some version of an evaluation metric. Such an approach leads one to identify as relatively likely those changes which the metric labels as local simplifications. Equally importantly, if the metric labels a change as an increase in cost, it should not be likely. The success of this method is the extent to which actually occurring changes fall into the former class and the extent to which the changes labelled as unlikely do not occur.

234

Changes in linking

The restrictiveness-based view of case leads to a diachronic explanation of the case alternations as well. Each case can be thought of as a rule and so it is natural to ask whether the changes above may be rule-based analogy rather than reanalyses as proposed by Lightfoot. What makes such a view possible is the further assumption that restrictiveness correlates inversely with simplicity, as argued in the previous chapter. The Optimality Hypothesis, restated here, is based on this assumption: (13)

Optimality Hypothesis If Grammar A links a given argument by a less restrictive linker than Grammar B in a simple clause, Grammar A is simpler than Grammar B relative to that argument.

If analogy is taken to be simplification as in Kiparsky (1982), then the expectation is that the preferred direction of a case change will be toward decreased restrictiveness: (14)

Restrictiveness Change Prediction Analogical change in case frames will be relatively favored in the direction which decreases the restrictiveness of the linker applying to a given argument.

Analogy in case frames is relatively more likely to result in a decrease in the restrictiveness of the case applying to an argument than in an increase.12 As we will see, a wide range of case changes can be explained in this way. 13 Both Dative Substitution and Nominative Substitution fall in the class of changes predicted as relatively likely. Both involve the optionality and loss of a more restrictive case in favor (synchronically and diachronically) of the next most restrictive case. Synchronically, the present theory treats Dative Substitution as a quadruple, Nominative Substitution as a triplecase competition for certain arguments. The competing cases are listed again in (15): (15)

a. ACC: [ARG|EXP|1] (lexical entries) > (Dative Substitution) b. DAT: [ARG|EXP] > (Nominative Substitution) c. NOM: [ARG] d. ACC: [ ]

I have indicated the changes on the right. Dative Substitution is the optionality and then loss of the lexical accusative with certain verbs.

A theory of analogical change in linking

235

Synchronically, whenever the option of leaving out the lexical accusative is taken, the dative automatically applies as the next most restrictive case. Diachronically, the no-accusative option is taken with greater frequency, until the lexical accusative option is lost for that verb. The change follows the preferred direction, that the restrictiveness of the case applying to a given argument will decrease. The dative is less restrictive than the accusative it replaces. Thus the direction in Dative Substitution with the great majority of verbs is the expected one. Similarly, Nominative Substitution has a diachronic dimension. As with Dative Substitution, the alternation is between a more and a less restrictive case for a given argument. By the prediction in (14), the variation should arise in favor of the nominative, and the change should generally move toward the nominative. This is because the nominative is less restrictive than the dative or the lexical accusative, the cases it replaces. The optionality this time is of the experiencer classification rather than a lexical case. Nevertheless, the alternation and change conform to the prediction: the trend is toward the less restrictive case. Both Dative Substitution and Nominative Substitution receive a unified account as manifestations of the same basic predicted trend. Furthermore, like lexical analogies, the shift is in a preferred but not exceptionless direction. For example, English shows a shift from strong to (the more regular) weak verbs, e.g. help/helped, used to be helpan/ healp. Most verbs have undergone the strong to weak shift, but the reverse occurs occasionally, e.g. dive/dived > dive/dove for some speakers. Similarly, all verbs in English and many in Icelandic and German have undergone or are undergoing the impersonal to personal shift, as predicted: the new case, nominative, is the next less specific case to the disappearing dative. Less frequently, we get the reverse. For example in Middle English non-nominative experiencers become more frequent for a time before dying out altogether. We can make two predictions about reverse analogy. First, like all analogy, reverse analogy is predicted to be rule based. The new analogical form conforms to a subregularity in the language. The form dove is not totally idiosyncratic but is part of the remaining strong verb class. Second, such analogy is less favored than the analogy in the reverse direction. Many more verbs in English have shifted from strong to weak than vice versa. Consonant with the present view, case frame analogy exhibits both these characteristics: reverse analogy is rulebased and rare relative to regular analogy. The nearest thing to excep-

236

Changes in linking

tions to DS are the Modern Icelandic verbs hlakka til 'look forward to' and kvidafyrir 'be anxious or apprehensive about' which earlier showed only nominative and which now can take nominative, dative, and accusative (see Svavarsdottir (1982)). I know of no verbs that start out taking only dative and wind up taking only accusative. The present view does not claim that exceptions to DS are impossible, only that they are relatively unlikely. Similarly, for NS the only exceptions that have come to my attention are those in Dal (1966: §124). She gives two examples, mangeln 'lack' which originally took a nominative experiencer but became impersonal after the beginning of the New High German period and brauchen which anyway changed in meaning from 'use' to 'need.' On the present approach, it is expected that such exceptions should be rare, which they are. For present purposes, it is sufficient to note that in the large number of verbs involved in the languages considered, there are many more verbs which undergo DS or NS than undergo the reverse changes, making DS and NS true trends, as predicted. Note that the approach just sketched fits in well with the assumption of no negative evidence in language acquisition. The assumption that the loss of a form leads to addition of an exception feature is simply unnecessary. If the form being lost is less regular than another applicable form, then the loss of the first form is feature loss. This is true in boththe strong/weak verb shift and the dative/nominative change. The type of analogy involving the strong/weak verb distinction is one of the better understood aspects of language acquisition and is often cited as support for the proposition that children do not use negative data. A child does not need direct positive evidence to come up with a form like bringed, but rather uses a rule or affix entry (e.g. the -ed rule/entry) that she/he has set up. Seeing case frame changes like DS and NS as analogical regularization in fact ties in very nicely with the standard picture of language acquisition. 6.4

Acquisition of termhood

The change from one case to another is a change in the coding properties of some NPs. If the change is to nominative case or to subject agreement, then the NP in question has acquired a coding subject property. Coding subject properties usually but do not always coincide with behavioral subject properties, such as ability to raise and be the target of control. This led Keenan (1976) to propose the behavioral versus coding subject

Acquisition of termhood

237

property distinction in the first place. More recently, Cole et al. (1980) have proposed a diachronic generalization using this distinction: that NPs acquire behavioral subject properties before or simultaneously, but not after, coding properties. 6.4.1

Acquisition of behavioral versus coding subjecthood

Cole et al. (1980: 720) propose three stages in the development of subjecthood for a given NP type: (17)

Stage A: The NP in question displays no subjecthood properties, either behavioral or coding. Stage B: The analogous NP exhibits behavioral, but not coding, properties. Stage C: The analogous NP displays both behavioral and coding properties.

The claim is that diachronic change, if it happens to a given NP, will be from Stage A to Stage C, possibly by way of Stage B but never by way of a putative Stage B': (18)

* Stage B': The analogous NP exhibits coding, but not behavioral, properties.

For example, a lexically case-marked NP may start to participate in Raising and Control, be the antecedent for certain kinds of reflexives, etc., before it acquires nominative case, but not vice versa. We get no nominative NPs that behave as obliques. In the following I will show how this very powerful generalization follows from the theory so far. The key is in the relation of obliqueness to linkers, discussed at the end of Chapter 4. There it was argued that obliqueness (reconstructed in the theory as the feature-value pair [ + OBL]) should depend on the linkers rather than vice versa. Obliqueness is a property that NPs receive by virtue of the nature of the linker that links them, rather than being some property inherent in the NP. The generalization that Stage B' is ill-formed follows from the condition on obliqueness that only linkers specified beyond [ARG], i.e. lexical linkers, may be oblique. The generalization about the acquisition of subjecthood in Cole et al. (1980) raises the further question of directionality. All the languages and language families considered by Cole et al. (Germanic, Polynesian, and Georgian) show the acquisition of subjecthood, but Cole et al. leave open

238 Changes in linking the possibility of de-acquisition of subjecthood: the passage of an NP from Stage C to Stage A by way of Stage B (but not B'). I will argue in the following that the acquisition of subjecthood is unidirectional in the sense of §6.1.1. This statement is subject to one major proviso: that grammaticalization (e.g. lexeme to postposition to case marker) can renew the stock of argument linkers. The acquisition of subjecthood may also be extended to acquisition of termhood. NPs seem to acquire objecthood in a fashion parallel to acquisition of subjecthood. The problem with observing this trend is that there are fewer good tests of objecthood than subjecthood. The acquisition of termhood follows in the theory of Kiparsky (1987), where grammatical linking takes priority over semantic linking. The acquisition of termhood is the acquisition of grammatical linking. In the following, I will try to extend the generalizations about acquisition of termhood, unifying the acquisition of termhood with other changes like the change of which the alternation called Dative Substitution is a slice. The prediction that linking changes should involve decreases in restrictiveness of the linker applying to an argument is a "tendency" explanation which will unify the acquisition of termhood with changes like Dative Substitution which do not involve termhood at all. The present approach leads to an extension of Cole et al.'s (1980) acquisition of subjecthood. In Nominative Substitution we saw our first example of the acquisition of coding properties. It was clear that some of the NPs acquiring nominative already had behavioral subject properties, but others did not. Modern Icelandic lexically case-marked NPs are at Cole et al.'s Stage B: they display behavioral subjecthood properties but not coding subject properties. In NS, these NPs acquire coding subjecthood as well. In German, on the other hand, lexically marked NPs behave as obliques. They are at Stage A, showing neither behavioral nor coding subjecthood properties. NS for these NPs is the simultaneous acquisition of both types of subjecthood properties. 6.4.2 Acquisition of objecthood In this section and the next I will show how the Restrictiveness Change Prediction handles the various changes. Recall that changes should be of the following general type: (19)

a. LINKER A: [X|Y] where Y is non-null b. LINKER B: [X]

Acquisition of termhood

239

I have already noted that DS and NS receive a unified account by the Restrictiveness Change Prediction (RCP). Further, the acquisition of objecthood should fall under the RCP, because the loss of an idiosyncratic case in favor of the accusative is also an instance of restrictivenessdecreasing change relative to an argument: (20)

a. Case A: [ARG|2] b. ACC: [ ]

(part of a lexical entry)

This type of change is extremely common. This change is a strong tendency even for languages like German which still retain lexically casemarked behavioral obliques. This change has actually occurred for the Middle High German verbs begern 'desire,' beginnen 'begin,' hueten 'watch,' swern 'swear,' verge^en 'forget,' zihen 'pull' which all took genitive theme logical objects in Middle High German but which by Modern German have replaced the genitive with an accusative. Consonant with the view that such change is characterized by lexical diffusion is the survival of some genitive object verbs into Modern German. An example is Middle High German bedurfen 'need' which as Modern German bedurfen 'need' still takes the genitive. And entbehren 'do without' (Middle High German entbern) shows this change in progress: it takes either genitive or accusative where earlier it took only genitive. Such shifts to accusative have occurred and are occurring in many other Indo-European languages (Classical Greek, Russian, English, Latin and Romance, etc.). 14 Another interesting example of the acquisition of objecthood comes from Lardil (Klokeid (1978)) a PamaNyungan language of Australia. The reason this change is particularly interesting is that it seems to involve reformulation rather than loss, as with the genitive in German. Klokeid demonstrates that the case called accusative in Lardil is cognate with the dative in the related language Yukulta, and that the accusative in Lardil derives historically from a dative like that in Yukulta. Klokeid argues that the dative in Yukulta and the accusative in Lardil are both underlyingly /-intha/. The case called "accusative" in Lardil has this form: (21)

Ngawa pe-th-ur niwe-ntha-r. dog(ABs) bite-VM-FUT him-ACC-FUT 'The dog will bite him.' (Klokeid (1978: 603))

240

Changes in linking

Yukulta shows an ergative-absolutive pattern. The object is normally absolutive but can under special circumstances (involving negation and person) be dative: (22)

a. Ngawu-ya-dngu-kanta paaja. dog-ERG-them(ABs)-Aux bite 'A dog bit them.' b. Ngawu-thu-yinka paaja. dog(ABs)-DAT-Aux bite 'A dog bit me.' (Klokeid (1978: 588))

The /-intha/ case is otherwise the case of goals in Yukulta. Klokeid provides a wide range of facts supporting the reconstruction of a common ancestor with the ergative-absolutive pattern of Yukulta and the dative pattern of Yukulta. That pattern has a case for goals and for other arguments under certain conditions of negation and person. Klokeid analyzes this last use of the dative as 2-3 retreat. He then shows that for phonological reasons the ergative ~ya was lost in Lardil. He argues that word order was fixed, and the 2-3 retreat construction was generalized so that subject could be distinguished from object. In our terms what has happened is the following change: (23)

a. -intha: [ARG|GOAL|— HR]

b. -intha: [

]

c. -intha: [ARG|— HR]

The dative has not been lost but has been reformulated as an accusative of either type 1 (23b) or type 2 (23b). I do not have evidence pointing to which accusative. If the change is to (23b), then the goal winds up with a less restrictive case. Likewise if the shift is to (23b), the change is also most probably restrictiveness decreasing. As I noted above, it is not initially clear whether the feature [—HR] should be counted as part of LINK values and hence for restrictiveness. If not, then the shift is clearly restrictiveness decreasing. In any case it is not clear whether the dative was already [—HR] (as in (23a)) or not. Lardil under Klokeid's analysis provides an example of dative to accusative shift, which is probably an instance of reformulation of the kind expected under the RCP. 6.4.3 Acquisition of behavioral termhood So far we have seen how the RCP predicts the acquisition of coding termhood and unifies this with the shift to dative in DS. The RCP,

Acquisition of termhood

241

under a slight extension, is also able to predict the acquisition of behavioral termhood properties. The reconstruction of obliqueness versus non-obliqueness of an NP in the present theory is whether that NP bears [ + OBL] or [-OBL]. AS argued in §4.4 above, the unmarked value of this pair for argument linkers is [—OBL]. It is filled in by default: (24)

[

]^[-OBL]

All linkers that are specified [ + OBL] are prohibited from getting [-OBL] by the default, since the values are contradictory. The RCP states that as far as LINK values are concerned, arguments tend to be linked by less restrictive linkers over time (subject to renewal through grammaticalization). More default-like linkers win out over time. If the same principle, suitably generalized, applies to the default that assigns [—OBL] in (24), then we get the prediction that linkers (and hence their NPs) should lose their obliqueness over time: (25)

Linkers become [-OBL] over time.

Since NPs inherit their obliqueness or non-obliqueness from their linkers, we get the following corollary: (26)

NPs become [—OBL] over time.

This predicts the acquisition of behavioral subjecthood and behavioral objecthood as well. The data in Cole et al. (1980) are confirming instances of this prediction. In the next section I will examine Scandinavian in more detail. Finally, we ideally should have a way to capture Cole et al.'s (1980) most significant generalization, that coding subjecthood (and similarly termhood) properties are not acquired before behavioral subjecthood (termhood) properties. The impossibility of a Stage B' where an NP shows coding termhood properties but not behavioral termhood properties, is related to the issues of obliqueness raised in Chapter 4. There it was argued that obliqueness is best regarded as a feature of linkers, and more specifically of (some) lexical linkers, and this led to interesting results about the interaction of case and configuration. To capture these general facts, I proposed the Obliqueness Constraint, a Feature Co-ocurrence Restriction (FCR):

242 Changes in linking (27)

Obliqueness Constraint (FCR) Linker: +OBL D Linker: [ARG|X] where X is non-null + OBL

This simply says that in order for a linker to be oblique, it must be lexical. Lexical linkers (idiosyncratic, argument category) may be oblique or nonoblique, but non-lexical linkers (the [ARG] and [ ] linkers) must be nonoblique or (27) is violated. This leads to the following: (28) lexical non-lexical

oblique Y(A) N(B')

non-oblique Y(B) Y(C)

A "coding termhood" linker is non-lexical, i.e. nominative, accusative, ergative, absolutive, the [ARG] and default linkers. No such linkers can be oblique. Note that this means that a Stage B' for an NP is automatically impossible. To be oblique, an argument NP must be linked by a lexical linker, but, if it is linked by such a linker, it does not have coding termhood properties. In (28) I have indicated which Stage an NP would be at if it were linked by the type of linker represented by that cell. All the occurring Stages A-C are allowed ('Y' for yes) by the Obliqueness Constraint, while B' is exactly the one ruled out. The non-existence of Stage B' synchronically and diachronically can be unified with the facts about case and configuration in §4.4 above. 6.4.4 Reanalyses I have argued that changes like those behind DS and NS in the Germanic languages are best regarded as changes which follow from the Restrictiveness Change Prediction rather than from the reanalysis or parameter-setting account. I am not, however, claiming that no case changes involve reanalysis. First of all, it is clear that phonological change could effect a case change, although clear examples are difficult to find. But reanalysis does not depend on there being a phonological trigger. In this section I will point out three types of change for which reanalysis has been proposed and seems to be the best analysis. First, the change of nominative-accusative languages into ergativeabsolutive ones is a change that is relatively well understood. Anderson (1977) proposes that this change can arise as the result of the reanalysis of two constructions in nominative-accusative languages: the passive and the perfect. The passive in a nominative-accusative system is reanalyzed

Acquisition of termhood

243

as an active. The by-phrase marker is the new ergative marker, and the old nominative is the new absolutive. In the present theory, such changes are indeed best treated as reanalyses. In Chapter 2 it was argued that twocase systems show a surprising uniformity cross-linguistically. The difference between any two languages of the four-way typology is in the setting of one or, at most, two parameter settings. This means that as far as the two grammatical cases (or other linkers) go, the systems are equivalent. Thus the change from one to the other is less of a change than appears at first sight. While individual arguments may change in the restrictiveness of their linker, overall there is no change. The shift is not from one linker to another for a given argument, the type of change the Restrictiveness Change Prediction deals with, but rather a change from one system to another. There is, however, one aspect of the rise of ergativity that is an instance of the acquisition of subjecthood and so should follow from the RCP. This is the fact that when a marker changes from being a by-phrase marker to being a syntactic ergative case, along the way somewhere it changes from [ 4- OBL] to [—OBL], the default value for obliqueness. This aspect of the change is predicted by the RCP and is documented by Cole et al. (1980). An open question is at what point in the rise of ergativity this acquisition of subjecthood occurs. But it does follow from the RCP that this acquisition of subjecthood occurs and that the reverse change from ergative case to by-phrase marker does not occur. Another type of reanalysis that may occur is the rise of idiosyncratic case. Recall that the change of dative to nominative in NS with a verb was taken to reflect the change of the verb from taking a syntactic experiencer to taking a direct argument. The direct argument is no longer eligible for the experiencer dative. If this change happens with enough verbs, then the motivation for the syntactic experiencer dative might become unclear. If the dative appeared with a small enough group of psych verbs, the dative might become idiosyncratic: (29)

DAT: [ARG|EXP] > DAT: [ARG|1] (part of a few lexical entries)

This change would be a reanalysis based on the opacity of the experiencer dative. Further, this change would be an apparent counterexample to the RCP, because the NPs undergoing the changed dative would be linked by a morphologically similar dative which is more, not less, restrictive. Note, however, that this change results from the general fulfillment of the RCP on the majority of original dative experiencer verbs. This would mean

244

Changes in linking

that the RCP's prediction that change in the direction of decreased restrictiveness is more likely than the reverse is not violated. I leave it to further work to ascertain how often this type of change occurs. Reanalysis may also be behind the lexical accusative phenomenon of Icelandic. If the lexical accusative in Icelandic and its relatives was originally a syntactic accusative, then the NPs linked by the lexical accusative seem to have undergone a change in the direction opposite to that predicted by the RCP. Since the idiosyncratic accusative and the default accusative are identical morphologically, it is desirable to find a common diachronic origin for them. Delbruck (1893: §15) suggests that these verbs arose from a shift from a personal construction to a construction with only an object in order to emphasize the passive role of the experiencer. One way for this to happen would be through the rise of a null subject which would force the experiencer to appear as object. Later the verb might become a subjectless impersonal. Delbruck also reports on the other view that these verbs were once causatives (e.g. something causes me thirst) which later became impersonal. In any case, an object accusative might be reanalyzed as an idiosyncratic accusative. This type of change is evidently not very common, as evidenced by the lack of idiosyncratic accusatives outside of Germanic and traces in some other IndoEuropean languages. Finally, it is worth pointing out that some of the most intensively studied syntactic changes, changes in word order, bear only indirectly on the RCP. The well-known shift from OV to VO order in the history of English (see van Kemenade (1987) and Lightfoot (1991) for recent discussions) and Scandinavian (e.g. Sigurdsson (1988) on Icelandic) do not necessarily change the inventory of linkers. This is because it is not necessary that word order be a linker in languages undergoing such shifts. However, word order sometimes does have a linking function. As seen in §4.4 above, word-order linkers do not usually refer to lexical information and therefore are usually not oblique. Thus when a word-order linker is adopted it is usually non-oblique and less restrictive than some existing linker. Thus a trend towards decreasing restrictiveness in the linker applying to a given argument would favor the rise of such word-order linkers. On the other hand, if word-order linkers are usually non-oblique, it is clear that the theory outlined in earlier chapters provides a way for them to be lost as well. If a new oblique case is grammaticalized in (e.g. from a preposition or a clitic), then that case is likely to refer to lexical information. In this situation a new oblique lexical case competes with a less

Scandinavian

245

restrictive non-oblique word-order linker. The new oblique case must apply, precluding the positional linking. Therefore, by itself, the RCP says nothing about which word orders shift to which other word orders, nor does it predict any overall direction in the long run towards or away from rigid word order. Whether the RCP makes verifiable predictions about which cases give way to which word-order linkers and vice versa remains a subject for further study. 6.5

Scandinavian

In this section I compare the Scandinavian languages in greater depth. Here there is great diversity in form and functions of case among a set of closely related languages. The oldest attested stage, Old Icelandic (Old Norse), will be seen to represent the situation inherited from Common Scandinavian. This seems doubly reasonable, since earlier Old English, Gothic, and all stages of continental High German are quite similar in the most important features of their case systems to Old Icelandic. Thus Old Icelandic also represents the continuation of the situation in ProtoGermanic. The acquisition of termhood by certain NPs in Scandinavian and the fixation of word order will be seen to be aspects of restrictiveness-decreasing analogy.15 6.5.1 Old Icelandic Old Icelandic exhibits extensive use of lexical cases which are oblique. These cases are dative, genitive, and accusative (there is also a syntactic accusative). The cases I will argue for are the following:16 (30)

Old Icelandic a. GEN: [ARG|2] (part of the lexical entry of hefna 'avenge,' etc.) + OBL

b.

DAT: [ARG|2]

(part of the lexical entry of sceta 'waylay,' etc.)

+ OBL c. DAT: [ARG|GOAL]

+ OBL d. e.

NOM: [ARG] ACC:

[

]

The nominative and the accusative are the same as in Modern Icelandic for the same reasons. Icelandic has been type 1 at all stages. All of the constructions cited in Chapter 1 as evidence for Icelandic's status as a

246

Changes in linking

type 1 language are present in Old Icelandic. The main difference between Old and Modern Icelandic is that the lexical cases are truly oblique in Old Icelandic, whereas they are not in Modern Icelandic. Since the dative and the genitive are more restrictive than the nominative and the accusative, the prediction is that the lexical case takes priority. And indeed one always gets the preservation of lexical case in the passive in Old Norse. 17 Thus, the actives in (31) have lexical cases: (31)

a. hann saetti J)ar skipum he(NOM) waylaid there ships(DAT) 'He waylaid ships there' (/fAr.iii.336) b. eigi vil ek lit ganga t>vi at ek em madr gamall not want I out to-go that for I(NOM) am man old ok litt til buinn at hefna sona minna and little for prepared to avenge sons(GEN) mine(GEN) 'I do not want to go out, because I am an old man and little equipped to avenge my sons' (Nj. 129)

Old Icelandic sceta 'waylay' with the dative and hefna 'avenge' with the genitive demand idiosyncratic case for their logical objects. As predicted, the dative- or genitive-marked NP does not become nominative in the passives: (32)

a. var ^eim saett was them(DAT) waylaid(N.sg) 'They were waylaid' (Fms. viii. 380) b. Hafgrimr sagdi at ]3essa mundi hefnt verda Hafgrim(NOM) said that these(GEN) would avenged become 'H said that these things would be avenged' {Flat.1.127.5)

There is arguably no semantic regularity to this use of the dative, as with idiosyncratic case in Modern Icelandic. The datives and genitives above in Old Icelandic are best regarded as idiosyncratic, i.e. associated with the second argument of the verb hefna 'avenge' and sceta 'waylay' as in (30a,b). For other verbs there should be other idiosyncratic genitives and datives. Two other datives are for experiencers (33) and goals (34) as in Modern Icelandic:

Scandinavian

247

(33)

mer J>ykkir J)eir merkligstir menn me(DAT) seem they(NOM) most-remarkable(NOM) men(NOM) 'I think they are most remarkable men/they seem to me most remarkable men' (Fs. 19)

(34)

hann gaf ]3eim godar gjafir he(NOM) gave them(DAT) good(ACc) gifts(ACc) at skilnadi at parting(DAT) 'He gave them good gifts at parting'

(GisL9)

Since dative links goals, the case is as in (30c). In (35), the goal dative in Old Icelandic is preserved in the passive, i.e. there is no alternation with nominative: (35)

Njall bad konu til handa Hogna Njal(NOM) asked wife to hands(GEN) Hogna(Acc) ok var hon honum gefin and was she(NOM) him(DAT) given 'Njal asked for the hand of Hogna, and she was given to him' (Nj. 120)

Example (35) is the passive of the same verb as in (34), gefa 'give.' Psychological experiencer verbs such as (33a) do not have passives, as in Modern Icelandic. Just as in Modern Icelandic, lexical cases, both idiosyncratic and argument category, have priority over both the nominative and the accusative by the RC. I hypothesized in (30) above that lexical cases in Old Icelandic are [ + OBL]. All genitive and dative NPs in Old Icelandic are obliques; that is, no NP marked dative or genitive patterns with NPs that are prototypically subject and/or object. The dative-marked NPs fail (on the basis of negative evidence) all the behavioral tests for subjecthood such as obligatory antecedence of reflexives and ability to be the Equi controlee. Claims of possible behavioral subjecthood of oblique NPs in Old Icelandic are reviewed and a new claim added in Rognvaldsson (1991). He correctly discounts as unreliable all the evidence for behavioral subject oblique NPs that comes from tests like Conjunction Reduction, Reflexivization, and Raising. We will begin by considering what he takes to be a more convincing source of evidence: inversion. Rognvaldsson (1991: 374) cites the fact that in sentences with inversion

248

Changes in linking

and an auxiliary, the oblique N P follows the auxiliary like a subject and never in his corpus follows the main verb: (36)

Ei m u n J)ig her mat skorta not will you(Acc) here food lack 'You are not going to lack food here' (Svarfdcela saga, p. 1806)

However, this is not conclusive either, since in the unmarked German word order, the dative experiencer comes after the inverted verb without being a subject: (37)

Gestern hat mir dieser Aufsatz nicht yesterday has me(DAT) this(NOM) essay not 'Yesterday I did not like this essay.'

(38)

Gestern hat dieser Aufsatz yesterday has this(NOM) essay

mir nicht me(DAT) not

gefallen. pleased gefallen pleased

One would expect sentences like (38) to be rare in a corpus of German sentences. But recall from Chapter 2 that German dative NPs do not behave as subjects; they are behavioral obliques. Nonetheless, in inversion sentences, the preferred word-order position of these behavioral obliques is immediately after the finite verb. Thus immediate postverbal position in inversion does not necessarily distinguish behavioral obliques from behavioral subjects. The Icelandic data from inversion cited by Rognvaldsson is consistent with the hypothesis that dative NPs are behavioral obliques, not behavioral subjects, in Old Icelandic. The other sources of evidence for behavioral subjects in Old Icelandic are even less persuasive. Cole et al. (1980) search for behavioral subject properties for these dative N P s in Old Icelandic and come up largely empty-handed. The one kind of N P they claim shows grammatical subject properties is the experiencer of the verb pykkja 'think/seem' (cf. me thinks from earlier English). Their argument rests on the evidence in (39): (39)

honum Jsottir J>u hafa haft vid him(DAT) seemed(2sg) you(NOM) to-have had against sik fjgrrad self death-plot 'He thought you to have had a death plot against him' (JV/.109)

Scandinavian

249

As Cole et al. (1980) themselves admit, the ability to be the antecedent of the reflexive is not good evidence for subjecthood in Old Icelandic, since objects can optionally control reflexivization. I will now show that examples like (39) show nothing about subjecthood in Old Icelandic. Smith (forthcoming) argues that the reflexive pronoun sik in sentences like (39) is best regarded as logophoric, as Thrainsson (1976), Maling (1984), and others have demonstrated for the Modern Icelandic long-distance anaphor sig. First of all, the longdistance reflexives have always occurred with the same class of verbs in Icelandic. In Modern Icelandic it can be shown that only verbs of communication and psych verbs allow long reflexivization into their complements. All of the examples I am aware of from Old Icelandic involving long reflexivization have the reflexive within a complement of a verb of saying or a psych verb. Sentence (40) is an example with a verb of saying, spyrja 'ask,' and (41) has a verb of mental state, hugsa 'think': (40)

(41)

jDorbjonii spyrr ef Samr vildi ngkkura lidveizlu veita Thorbjorni asks if Sam wants some support to-give seri selfi 'Thorbjorni asks if Sam would want to give him; some support' (Hrafn. 66) Ipk hugsadi Jx>rri f>at - ef hannA kvaemisva i then thought Thorj that if hei came thus to faeri at sla hannj et J>ridja hggg - at aldri opportunity to strike himj a third blow that never skyldi hannj sja sikj sidan. would hej see selfi after 'Then Thori thought that if hei got the chance to strike himj a third blow, hej would never see hinii again' (Snorra Edda 30,9)

Further, it can be shown that some reflexives must be long reflexives, because a sentence boundary intervenes between the antecedent and the reflexive: (42)

Jsat that at that var was

var eina nott t>a er MagniiSi la i hvilu sinnii was one night when MagnuSi lay in bed self Si hannj dreymdi ok f>6ttisk staddr !>ar sem he dreamed and seemed-to-self located there who-REL fadir hansi inn helgi Olafij konungrj [s ok J>6tti father hisi the holy Olafj (the) kingj [s and seemed

250

Changes in linking hanrij maela vid sikj ] s hej to-speak to selfi ] s 'That was one night, when Magnus^ lay in hisi bed, that hei dreamed, and hisi father, St. Olafj the kingj, who was hisi (dead) father, appeared standing there, and hej seemed to speak to hinii . . . ' {Harolds Saga Sigurdarsonar, p. 105)

The antecedent of sikt is hannt. Here the second ok is either an adverb meaning 'also' or it is a simple coordinating conjunction with VI, as the verb-second word order of ok potti hann shows. In either case there is a sentence boundary being crossed, and so these reflexives cannot require a subject antecedent in the same clause, since they do not require the antecedent to be in the same clause at all. It still might be thought that a logophoric long-distance reflexive requires a subject NP as antecedent, even if it is not in the same clause. Cross-linguistically it is not true that antecedents of logophoric pronouns must be subjects. Latin, for example, shows a wide variety of oblique NPs as antecedents of logophoric reflexives (Clements (1975: 143f.)). Thus (39) is in no way evidence of a behavioral subject property for the dative NP. The dative NP in (39) is the discourse antecedent of logophoric sik 'self and is not required to be a grammatical subject. In the absence of evidence for its subjecthood, I take the datives not to be grammatical subjects. This is also cross-linguistically the unmarked situation and arguably the situation in earlier Germanic. So on our account, a dative or genitive NP inherits obliqueness from the lexical case which links it. Thus honum 'him' cannot be a term in the syntax and so cannot be the grammatical subject. In (39) the second-person singular nominative pronoun, pu 'you' is the grammatical subject of the matrix verb pottir 'seems,' which agrees with it as well. Thus, as in Modern German, in Old Icelandic what looks oblique really is oblique: with respect to both coding and behavioral properties, genitive- and dative-marked NPs are obliques.18 The RCP predicts acquisition of objecthood as well as subjecthood. Acquisition of subjecthood is evidenced by lexically case-marked NPs not passing the subjecthood tests in Old Icelandic (based on negative evidence) and later the analogous lexically case-marked NPs passing the subjecthood tests in Modern Icelandic. Acquisition of objecthood is less easy to prove, because there are fewer good tests for objecthood. One possible test is the ability to become the subject in the passive, but

Scandinavian

251

this reduces objecthood to latent subjecthood. Fortunately, there is evidence in Icelandic, independent of passivization, that lexically case marked NPs have acquired objecthood as well. The reflexivization test extends to objects. Thrainsson (1979) shows that within simplex sentences, if the antecedent of a pronoun is a subject, then reflexivization is obligatory, but if the antecedent is an object, reflexivization is optional. Some speakers do not allow object controlled reflexives. This is shown in the following paradigm based on Thrainsson (1979: 291) and Maling (1990: 280): (43)

(44)

Eg

bardi SiggUi med diikkuni hennari/%sinnii. I(NOM) hit Sigga(Acc) with doll her(DAT)/%refl(DAT) 4 I hit Siggai with heri/self s{ doll.' Eg

syndi Haraldii fot a hanni/%sigi showed Harold(DAT) clothes(ACc) for him(Acc)/refl(Acc) 'I showed Haroldi clothes for himi/himsehV

I(NOM)

Dative objects pattern like accusative objects in being optional reflexive antecedents for those speakers who allow object antecedents at all. The lexically case-marked NPs are best regarded as terms in general, not just as subjects. This more general acquisition of termhood which has occurred in the history of Icelandic is exactly what is predicted by the RCP. Finally, Modern Icelandic has subjects in preverbal position, whereas Old Icelandic shows greater variety in the position of the subject (Nygaard (1905), see also Allen (1984) for similar changes in English). In the present theory, this means that the following positional linker has been acquired: (45)

_VP: [ARG]

Since this is a non-oblique position, it can only apply after the lexical cases lose their oblique status. This is because the lexical cases are more restrictive than the _ VP linker, and so use of the pre-VP position linker in the absence of the lexical case would violate the RC. But because the lexical cases are oblique, the non-oblique positional linker must not be used or the NP will inherit both [ + OBL] and [—OBL]. Thus loss of obliqueness on the lexical cases will allow a less restrictive linker to apply, lowering the average restrictiveness of the linkers applying to such arguments. One issue that deserves further work is whether lexical

252

Changes in linking

cases may lose obliqueness before the rise of positional linking. Kiparsky (p.c.) proposes that lexically case-marked NPs do not acquire behavioral termhood unless they are positionally linked. This is the diachronic dimension of the issue discussed in Chapter 4 about the relation of obliqueness and position. At the very least, it seems that non-oblique lexical case and position often correlate, and this may be related to the commonness of a pre-VP linker. Whether the correlation of non-oblique lexical case goes beyond what is predicted by the commonness of the _VP linker, I leave open. One final point of note about Icelandic is that the type of change seen here for Icelandic cannot be handled by the phonological view of case change. Case has hardly changed phonologically at all from before 1100 to the present. Phonologically there is no case change. But syntactically, as far as linking goes, there are vast changes between the two stages. Icelandic has moved from Stage A to Stage B: a wide range of NPs shifts from showing no subjecthood properties (behavioral or coding) to showing behavioral subjecthood properties. This is a purely grammatical change which has nothing to do with phonological form, and it is predicted as likely on the present account. The change is in the direction of the default [—OBL].

6.5.2 Mainland Scandinavian I turn now to Norwegian, which is also to some extent a daughter of Old Icelandic.19 In terms of case, Norwegian is like Swedish and Danish in having no semantic cases; only prepositions are [ + OBL]. Like English, these languages have at most nominative and accusative, and these only in the pronouns. Word order is fixed (e.g. Holmberg (1986)), so I propose the following linkers: Norwegian, Danish, Swedish (46)

a. _V: [ARG|GOAL]

b. c. d. e.

NOM: [ARG]

_V": [ARG] [ ] _V: [ ]

ACC:

The sentences that correspond to the ones we have been examining for the other languages are given in (47):

Scandinavian (47)

Han liker henne. he likes her(ACc) 'He likes her.'

(48)

Han ga henne boken. he gave her(Acc) book-the 'He gave her the book.'

253

There are of course no verbs with idiosyncratic dative, since there is no dative. The order of the objects is fixed in (48) and the subject precedes the verb. These languages should have the same inventory of linkers as English (as argued already in direct linking by Wechsler (1995)). The dative was lost, but there has been renewal in the rise of the verb-adjacent word-order position for goals. There may have been an intermediate stage like Grisons German (see §4.4 above) where the goal is doubly linked by an oblique dative and an oblique position, or the obliqueness of the dative may have been lost before the rise of the positional linking (as in Icelandic). Idiosyncratic case has been lost on logical objects as well. Reinhammer (1973) provides a very detailed account of the loss of idiosyncratic dative in a wide variety of Norwegian and Swedish dialects. What is very interesting to the RCP is that he finds that the datives with more motivation like the dative of goal are lost after idiosyncratic datives, like the dative with kasta 'throw.' First of all, this provides a further justification for treating argument category case and idiosyncratic case differently with respect to LINK values. Furthermore, less restrictive cases (like the dative of goal) are more stable than more restrictive cases (like the idiosyncratic dative with kasta 'throw'). In Standard Norwegian (also Swedish and Danish), idiosyncratic case has been completely lost. Thus a verb like hjelpe 'help' takes accusative although its relatives take dative in Old and Modern Icelandic, Faroese, Old English, and many other stages of various Germanic languages. In this, the modern Scandinavian languages are again parallel to English: (49)

Han hjalp henne. he helped her(ACc) 'He helped her.'

Thus for Modern Norwegian there has been a complete loss of datives, allowing the less restrictive syntactic cases to apply. The dialects and older languages (e.g. Old Swedish (Noreen (1903-1907)) show partial

254

Changes in linking

loss and are at the intermediate stages. The change that has happened in the mainland Scandinavian languages is that lexical cases are first partially lost, and later completely lost. Finally, since dative never marks in the mainland Scandinavian languages (since it does not exist anymore), it is not [ + OBL] either. 6.5.3 Summary of Scandinavian It is now possible to summarize the case changes in Scandinavian we have seen thus far. When one examines the languages and the linkers, the patterns stand out quite clearly (cf. predictions in (27) and (28)): (50)

Old Icelandic

Modern Icelandic

a. DAT: [ARG|GOAL]

a. DAT: [ARG|GOAL]

+ OBL

b.

NOM: [ARG]

c.

ACC:

[ ]

b. NOM: [ARG] c. _VP: [ARG] d. ACC: [ ]

*Stage B'

Norwegian et al.

a.

NOM: [ARG]

a. _V: [ARG|GOAL]

b.

ACC:

+ OBL

[ ]

b.

NOM: [ARG]

+ OBL

c. _VP

[ARG]

e.

[ ]

ACC:

Acquisition of fixed word order is the addition of non-oblique linkers. From Old Icelandic onward there is trend towards less restrictive linker(s) applying to given arguments, subject to renewal through grammaticalization of new argument category linkers. The main analogical trend in Scandinavian case is towards decreased restrictiveness. 6.5.4 A note on reconstruction If the RCP is correct in predicting the unidirectionality of these changes in linking, then the RCP has the potential for becoming a useful tool in syntactic reconstruction. Reconstruction in any area of grammar is much easier if it is known which changes are relatively likely or natural and which changes are less so. In reconstruction, the preference should be for positing likely changes relating the proto-language and the daughter

Conclusions 255 languages. The RCP labels changes which decrease the restrictiveness of the linker applying to a given argument as likely, and so reconstruction of the behavior of arguments at earlier stages should not normally involve restrictiveness-increasing changes. I leave this for further work.

6.6

Conclusions

This section considers the wider context in which the changes predicted by the RCP are located. I first consider the question of morphologically related linkers. I end with a consideration of what class of changes the RCP-type changes belong to. 6.6.1 Homophonous linkers There are some cases and prepositions that serve both as argument and as adjunct linkers and some that serve only one function. I illustrate this with the prepositions meaning to and datives from some Germanic languages. In German zu 'to' can be used for locations but not with verbs of transfer: (51)

a. Er geht zum Tisch. he(NOM) goes to-the(DAT) table 'He is going to the table.' b. *Er geht dem Tisch. he(NOM) goes the(DAT) table 'He is going to the table.'

(52)

a. Sie gibt ihm den Mantel. she(NOM) gives him the(ACc) coat 'She gives him the coat.' b. *Sie gibt den Mantel zu ihm. she(NOM) gives the(ACc) coat to him(DAT) 'She gives him the coat.'

The dative, like the immediately postverbal position in English, cannot be used for purely spatial goals. The preposition zu cannot be used for argument goals of possession. In English and Mainland Scandinavian, the preposition to or til respectively can be used for either adjuncts or goal arguments: John gave Bill the book and John went to the store. Diachronically, prepositions seem to start out as independent

256

Changes in linking

predicators and become argument cases (see §6.1). This is happening with to in English. It is happening with til in Mainland Scandinavian but not in Icelandic, which is like German. In Romance, Latin could use ad 'to' only as an independent predicator, but in French a 'to' can, like to in English and til 'to' in Mainland Scandinavian, be used both for adjuncts and arguments. Thus the diachronic picture is as follows: (53)

predicator ( NP ) predicator ( NP ),

LINKER: [ARG|X]

LINKER: [ARG|X]

I II

> >

III

Latin ad, Icelandic til, and German zu are at stage I; English to, Mainland Scandinavian til, and French a are at Stage II; and the Latin, Icelandic, and German datives, and the postverbal position in English and Mainland Scandinavian are at Stage III. Stage II shows mixed behavior. The X on the end of the specifications for the argument use of the case usually receives an interpretation related to that of the independent predicator. This relation may be purely diachronic, as Kurylowicz argued. Certainly the relatedness of the interpretation of argument categories and the spatial prepositions can be explained diachronically. If the two are supposed to be related synchronically, as on "locational" approaches, the question is what more one explains by positing such a synchronic relationship. Demonstrating a morphological identity and semantic relatedness is not sufficient to justify treating homophonous linkers as a unified syntactic category. The synchronic status of semantic generalizations about linkers is particularly acute for idiosyncratic cases. The idiosyncratic genitive in Icelandic is a good example. Some verbs like bida 'await' take a genitive case on their second argument: (54)

J)eir bida !>in. They(NOM) await you(GEN) 'They are waiting for you.'

Three approaches are possible. The first is to follow J. Anderson's (1986) approach to the Old English genitive and call it a genitive of cause: (55)

a. Him daes sceamode him(DAT) that(GEN) shame3sg 'He was ashamed of that'

Conclusions 257 b. Menn scama]3 for godan daedan swydor men(DATsg)/(ACcpl) shame3sg for good deeds more donne for misdaedan than for misdeeds 'A man is more ashamed of his good deeds than his bad ones' Synchronically there is evidence in Old English that the genitive has a cause meaning, since it was freely addable to a clause as an adjunct. Second, as demonstrated by (55b), the genitive NP alternated with a for-PP (a newer device to express cause), showing that the phrase synchronically has a causal meaning. In modern English this for is less semantically transparent and 'wait for' is more like a transitive and is not synonymous with 'wait because of. One can wait because of John without waiting for John (or awaiting John). Interestingly, Old English has the genitive with bidan 'await,' cognate of Icelandic bida 'await': (56)

bldaj) Dryhtnes domes await-3pl Lord's judgment(GEN) 'They await the Lord's judgment'

(Exon.23a)

A cause phrase is most probably the diachronic source of the genitive in Icelandic, but it is hard to tell if the genitive is synchronically a genitive of cause. One weak piece of evidence that it is not is that no independent cause phrases get the genitive in Modern Icelandic and new verbs do not assimilate to the NOM-GEN pattern, even if they plausibly involve causes. 6.6.2 Grammaticalization and analogy in linking Changes in linking share properties with two of the major classes of syntactic change: grammaticalization and analogy. If a linker is lost or is reformulated as less restrictive, it is in some sense becoming more grammatical. It applies more blindly than before. The rise of new argument category linkers like the [GOAL] linkers of the last section is a form of grammaticalization. An independent lexcial item is now a grammatical marker of syntactic goals. In Chapter 1, when introducing the notion of syntactic argument categories, I used an analogy with the category AUX. The category AUX (represented by the feature value pair [AUX + ]) was interpreted in a constrained way (tense, aspect, and/or modality), but it is not true that whenever one of these notions is expressed, we get AUX behavior. The category AUX and the argument categories are subcategories of verbs and arguments, respectively, that are interpreted in

258

Changes in linking

a constrained way. Auxiliaries have been regarded as grammaticalized verbs (Croft (1991: 142ff.)), because they are verbs on their way to being grammatical markers. So too the argument categories may be half-way between the predicators like to from which they derive and on their way, via the path predicted by the RCP, to the status of grammatical case ([ARG] or default) as in Lardil or to oblivion. Argument categories are analogous to auxiliaries in the diachronic dimension as well. On the other hand, the changes predicted by the RCP involve loss or reformulation in the direction of more "rule-like" linkers. The linkers in conjunction with the three constraints are like rules. The replacement of one linker by another is also a form of the familiar rule-based analogy (Kiparsky (1982)). The changes predicted by the RCP bear certain resemblances to both grammaticalization and analogy. Most importantly, the RCP allows a unified account of a wide range of changes involving case, agreement, and word order.

7

Case semi-preservation

Faroese and Classical Greek show optional preservation of lexical case in the passive. This phenomenon is not accounted for by current theories of case assignment. This chapter will show how the theory developed so far predicts the exact range of preservation facts and their diachronic evolution. 7.1

Case (non-)preservation

7.1.1 Case preservation Icelandic and most case languages have lexical cases which mark the object and which are preserved in the passive: (1)

NPi-NOM V-ACTIVE NP2-DAT <£> NP2-DAT V-PASSIVE (by-NPi)

This pattern is illustrated in (2) and (3) with the active in (a) and the passive in (b): (2)

a. Barnid tyndi bokinni. child-the(NOM) lost book-the(DAT) T h e child lost the book.' (active, dative object) b. Bokinni var tynt (af barninu). book-the(DAT) was lost by child-the(DAT) 'The book was lost (by the child).' (passive, dative subject)

(3)

a. Eg

sakna £>in. miss you(GEN) 'I miss you.' b. t>in er saknad. you(GEN) is missed 'You are missed.' I(NOM)

(active, genitive object)

(passive, genitive subject)

In (2) and (3) the verb selects a lexical case for its object, dative for tyna 'to lose' and genitive for sakna 'to miss.' In the passives (b), each of these 259

260

Case semi-preservation

cases appears on the corresponding subject. That such NPs really are terms (show grammatical functionhood behavior) in Icelandic was first established by Andrews (1976), and the arguments were extended by Thrainsson (1979) and Zaenen, Maling, and Thrainsson (1985). In response to data like (2) and (3), most theories make this casepreservation effect a necessary property of any grammar (see §7.1.5 below). The effect must be to give the dative and genitive priority over the accusative and nominative. These last two cases should only appear, as in (4), when neither dative nor genitive can: (4)

a. Hann lamdi mig. he(NOM) hit me(Acc) 'He hit me.' (active, accusative object) b. Eg var laminn (af honum). I(NOM) was

hit

'I was hit (by him).'

by him(DAT)

(passive, nominative subject)

Normally, objects get accusative and subjects nominative (with wellknown exceptions). By ordering the dative and genitive rules before the nominative and accusative rules, the NP will get the lexical case regardless of whether it surfaces as subject or object. The case-preservation effect follows automatically. This approach runs into problems, however, if the case-preservation effect is not obligatory. Faroese and Classical Greek show optional preservation of lexical case in the passive. I will now present the different behavior that these two languages exhibit with respect to casepreservation. 7.1.2 Case semi-preservation Another option is non-preservation, where a lexically case-marked object in the active corresponds to a nominative subject in the active: (5)

NP r NOM V-ACTIVE NP2-DAT & NP2-NOM V-PASSIVE (by-NPi)

Faroese and Classical Greek show optional preservation of lexical case in the passive; they exhibit both patterns, (1) and (5). Although Faroese is the language most closely related to Icelandic, it shows optional case preservation, i.e. case semi-preservation, rather than full case preservation. Faroese has almost completely lost the genitive with verbs but it does have datives with verbs. The verb hjdlpa 'help' takes the dative on its second argument:

Case (non-)preservation (6)

261

a. Teir hjalpa honum. they(NOM) help him(DAT) 'They help him.' b. Hann vard hjalptur. he(NOM) became helped 'He was helped.' c. ?Honum vard hjalpt. he(DAT) became helped 'He was helped.'

The dative is the only possibility for the active in (6a); the accusative would be ungrammatical. In the passive, however, the dative may be preserved as in (6c) or not, appearing as nominative as in (6b). The difference between (6b) and (6c) is purely of form. There is no corresponding semantic difference. In (6c) the idiosyncratic dative is preserved, but in (6b) it is not. The variants in (6b) and (6c) are in free variation. However, (6b) is considered more colloquial while (6c) is considered more formal. One possibility with a periphrastic passive like that in Faroese and Icelandic is that it may be adjectival. If the passive in (6b) with a nominative subject is adjectival, then all we have is simple predication. However, it seems to be verbal. First, Icelandic in fact does have an adjectival passive with nominative subjects: (7)

a. Dyrnar voru lokadar. doors-the(NOM) were closed(3pl) 'The doors were closed.' b. Dyrunum var lokad. doors-the(DAT) was closed(3sg) 'The doors were closed.'

(stative)

(dynamic)

Benediktsson (1980: 116f.) argues that only the nominative passive with the stative interpretation is adjectival, since only it can take the prefix -6 'un-,' just as a large class of verbs in English can take the prefix un- in the adjectival but not the verbal passive (see Wasow (1977) and references there): (8)

a. Dyrnar voru olokadar. doors-the(NOM) were unclosed(3pl) 'The doors were unclosed.'

(stative)

262

Case semi-preservation b. *Dyrunum var olokad. doors-the(DAT) were unclosed(3sg) 'The doors were unclosed.'

(dynamic)

According to Lockwood (1955: 134), Faroese has both types of passive, adjectival and verbal, and both can have nominative subjects. The distinction seems to parallel that in German (where the sein 'be' passive is stative and adjectival, and the werden 'become' passive is verbal and dynamic). The Faroese passive with vera 'be' is adjectival and receives a stative interpretation, and the passive with verda 'become' (or, more archaically, bliva 'become'), with which we are concerned (e.g. (6)), receives a dynamic interpretation and is verbal: (9)

a. Seglini eru slitin. sails-the(NOM) are torn 'The sails are torn.' u c r • [verda] r . b. Seglini I , slitin.

(stative)

[bliva J sails-the(NOM) become torn 'The sails are being torn.' (dynamic) The passive in (9b) belongs to this latter type and is most probably verbal. The focus in this chapter is on the verbal ("become") passive in Faroese, in its case-preserving and non-preserving variants (e.g. (6)). The passive in Classical Greek is affixal rather than periphrastic (and so very unlikely to be adjectival), but it too shows the same semipreservation effects as in Faroese. Verbs that take only dative or genitive objects normally form personal not impersonal passives. In Classical Greek, a dative-object verb like epibouleud 'plot against' normally loses case in the passive: (10)

a. kai proteron epeboiileuse soi and before plotted-against you(DAT) 'and he plotted against you before' (Aristoph.£#.984) b. hemeis hup' Athenaion epibouleuometha we(NOM) by Athenians(GEN) plot-against(lpl.PASs) 'we are plotted against by the Athenians' (Thuc.I.82.1: 32)

Similarly a genitive-object verb such as katapsephidzomai 'vote against' normally shows the non-preserving passive:

Case (non-)preservation (11)

263

a. hoti mou katepsephisasthe that me(GEN) voted-against-2pl 'that you voted against me' (Plato.^4/>.36al) b. kai ekeinos men katepsephisthe kai apothneiskei and that-one(NOM) PTCL was-condemned and dies 'and so that man was condemned and dies' (XQn.HelLY.236: 27-28)

Note that (1 lb) is a passive of a middle, and that the passivized genitivetaking verb is conjoined with an active verb. The non-preserved pattern is normal in the passive just as in Faroese (6b). As in Faroese, preservation seems to be possible but not favored. This is even more true in Greek. Although in Greek impersonal passives of all verbs are rare (see Schwyzer (1966: 11.239) for discussion), preservation of lexical case does occur. The preserved dative is shown in (12), the preserved genitive in (13): (12)

nomidzon t§i te peniai aut&n believing the(DAT) PTCL poverty(DAT) they(GEN) epikekourSsthai an to-have-been-aided PTCL

'(I) believing . . . that their poverty would be relieved' (Xen. vert. 1.1: 8-10) (13)

Lakedaimonioi antelegon me dikaios sphOn Lacedaemonians(NOM) replied not justly they(GEN) katadedikasthai to-have-been-judged-against 'the Lacedaemonians replied that they had been unjustly judged against' (Thuc.5.49.2: 22-24)

If case preservation is really optional in Greek, then the expectation is that the same verb will show both patterns. Although evidence is sparse for Greek, it does seem that a verb can show preservation and non-preservation. Corresponding to the passive in (13) with preserved genitive, we get (14) with no preservation:

264 (14)

Case semi-preservation ean de tis apheiretai ten arkhen ten if PTCL anyone(NOM) obstruct the(ACc) court(ACc) the(ACc) katadikasasan katadikastheis . . . judging-against(ACc) judged-against(NOM) 'If anyone, judged against, should obstruct the court which judged against (him) . . . ' (Plato.Afora.958c: 2-3)

Better evidence comes from (15), where the same author in the same text has both types of passive for the genitive-object verb kategoreo 'accuse, charge': (15)

a. ou dikaios kategoroumai not justly accused-1 sg.PASs 'I am unjustly accused' (Antiphon.7Wr.3.2.7) b. eph' hois kategoreitai mou in those-things charge-PASS me(GEN) 'in those things (with respect to which) I am charged' (Antiphon.r^r.1.2.13)

The dative hois 'those things' is "attracted" from the accusative into the dative of the main clause (for a discussion of the phenomenon of case attraction in Classical Greek relative clauses, see Smyth (1956: §2532)). Before continuing, it should be emphasized that verbs showing nonpreservation always take the lexical case in the active. That is, the phenomenon considered here is not a case of the dative being optional in all contexts. A few verbs in both languages allow the accusative as an option in the active and seem to be shifting towards accusative, but the vast majority do not allow this: (16)

a. logon d' akousai tis blabe; words(GEN) PTCL to-hear what harm 'To hear his words, what harm?' (S.OC.l 187) b. ho d' hos akoiiei phthoggon eksaiphnes he(NOM) PTCL as hears-3sg sound(ACc) suddenly pikron . . . bitter(Acc) 'he, as he suddenly hears their bitter cry . . . ' (S.0C.161O)

For verbs like this, the passive with a nominative subject is unsurprising, since the accusative variant of the active (16b) can be taken as the "source" of the non-case-preserving passive. This "natural" explanation

Case (non-)preservation

265

is not available for the verbs presented earlier in this section, since they never allow accusative objects in the active. Thus the non-preservation of case for the verbs above is a separate issue from the optionality or loss of a case across the board exemplified in (16). A theory of case should be able to explain both patterns. 7.1.3 Exceptions Compounding the problem of case semi-preservation is the fact that not all datives can correspond to nominatives in the passive. In both Faroese and Greek certain datives are always preserved. The uniformly preserved datives are goals, i.e. they necessarily denote a participant to which something moves. The discussion in Barnes implies that datives of goal are preserved with some if not all verbs. Thus in Faroese we seem to have the following:1 (17)

a. Hann seldi bondanum eina kiigv. he(NOM) sold farmer-the(DAT) a(ACc) COW(ACC) 'He sold the farmer a cow.' b. Bondanum vard seld ein kiigv. farmer-the(DAT) became sold a(NOM) COW(NOM) 'The farmer was sold a cow.' c. *B6ndurinn vard seldur eina kiigv. farmer-the(NOM) became sold-m.sg. a(ACc) COW(ACC) T h e farmer was sold a cow.'

Only the preserved version in (17b) but not the nominative version in (17c) seems to be allowed. Greek also normally preserves the dative of goal according to Kiihner (1898): (18)

ekeinoi d' haiite he khora edothe him(DAT) PTCL this(NOM) the(NOM) land(NOM) was-given 'this land was given to him' (Xen.ift?//.3.1.6: 19-20)

The version of (18) with the first NP in the nominative is attested nowhere according to all the grammars (Kiparsky (1987) notes some lexicalizations of the theme and verb into a complex verb, where we appear to get passives of dative goals). The question then for Faroese and Greek is why datives of goal should be particularly prone to full preservation even in semi-preservation languages. On the other hand, the Ob-Ugrian languages Ostyak and Vogul seem to show very widespread case semi-preservation (Lavotha (1960),

266

Case semi-preservation

Kulonen (1989)). Here non-objects marked in cases other than the accusative seem to alternate with the nominative in the passive. The accusative is not possible in the active. Thus we get the active-passive pair in (19) from Ostyak (Kulonen (1989: 157)): (19)

a. ew9 tapat woja penta jo%tot girl(NOM) seven elks' path(LAT) came 'The girl came to the path of the seven elks.' b. tapat woJ9 pent ewans joxtaj seven elks' path(NOM) girl(LOc/by) came-PAss T h e path of the seven elks' was come (to) by the girl'

It is not clear whether the impersonal variant is also grammatical (it is, at any rate, rare according to Kulonen (1989: 259)), but since impersonal passives (including those of unaccusatives) occur generally in both Ostyak and Vogul, the expectation is that the case-preserving variant is grammatical too. The main difference between Ob-Ugrian and the other languages above is that case semi-preservation is extremely general. Most lexical cases show the pattern of (19). Further, there must be some way for the adjuncts to be promoted to arguments; verb agreement with adjuncts is rare or non-existent. If the passive subjects are still adjuncts we would expect that verbs do not put restrictions on which adjunct can be "subject." But Kulonen implies that not just any adjunct can be the subject of the passive. For instance, the passive with subject-location phrase tends to occur with certain verbs, e.g. jox^t 'come' (Kulonen (1989: 158)). Once again, since Kulonen is considering occurrences in a corpus rather than grammatically, it is not clear just how restricted the passive is. I leave this question open.

7.1.4 Isolated case semi-preservation Some languages show a pattern in between preservation and semipreservation. In these languages a small class of verbs shows semipreservation, and the rest of the verbs show obligatory preservation. Two such languages are French and Old English, where the semi-preservation phenomenon has been treated as a puzzle in the literature, since normally nominative passive subjects correspond to active accusative objects.

Case (non-)preservation

267

7.1.4.1 French The first traditional puzzle has been discussed at length by Postal (1982, 1986: 35 and references there). Many verbs in French take a logical object marked by the preposition a 'to,' for instance telephoner 'to telephone': (20)

Paul telephonera *(a) son pere. Paul will-telephone to his father 'Paul will telephone (to) his father.'

The verb must take the preposition; leaving it out leads to ungrammaticality. Similarly, the object may be a dative clitic but not an accusative one: (21)

Paul lui/*le telephonera. Paul to-him/him will-telephone 'Paul will telephone (to) him.'

Not surprisingly, such sentences correspond only to impersonal (22) but not personal passives (23): (22)

II a ete telephone au capitaine. It has been telephoned to-the captain 'The captain was telephoned.'

(23)

*Le capitaine a ete telephone, the captain has been telephoned 'The captain was telephoned.'

A small class of verbs behaves exactly the same as this large class, forming actives only with a objects, except in the passive. This class includes obeir 'to obey,' desobeir 'to disobey,' pardonner 'to pardon,' and for some speakers, consentir, 'to consent to,' etc. First, the actives, parallel to (20) and (21), can only have an a object: (24)

Paul obeira *(a) son pere. Paul will-obey to his father 'Paul will obey his father.'

(25)

Paul lui/*le obeira. Paul to-him/him will-obey 'Paul will obey him.'

268

Case semi-preservation

Once again the direct object is impossible in the active. But in the passive, the obeir class forms not only the impersonal passive (26) but also the personal passive (27): (26)

II a ete obei au capitaine. It has been obeyed to-the captain T h e captain was obeyed.'

(27)

Le capitaine a ete obei. the captain has been obeyed 'The captain was obeyed.'

The impersonal passive of obeir (26) is parallel to the impersonal passive of telephoner (22), but obeir also allows the personal passive (27), whereas the more normal telephoner did not allow the personal passive (23). This is case semi-preservation. The obeir class is behaving exactly like the Faroese and Greek verbs in §7.1.2. The only difference between French and Faroese is that the class of semi-case-preserving verbs is quite small in French, whereas in Faroese and Classical Greek it is by far the norm. 7.1.4.2 Old English The status of the passive rule in Old English has been a matter of great controversy. One aspect of this controversy has been whether the personal passive of dative objects was possible. Visser (1973) claims that at least some dative object verbs could form personal (non-case-preserving) passives in Old English. Mitchell (1979, 1985) reviews the evidence for such passives in Old English and comes up almost - but not quite empty-handed. First, many of the verbs that take a dative object in the active corresponding to a nominative subject in the passive actually took an object in either the dative or the accusative. These verbs thus do not show case semi-preservation but full optionality of the lexical case. One such verb is blissian 'to gladden, delight' where the dative object is italicized: (28)

Sum sceal on heape blissian aet beore one(NOM) shall in company(DAT) delight at beer(DAT) bencsittendum benchsitters(DAT) 'One shall in company delight the bench-sitters at beer'

(ExonMa)

Case (non-)preservation

269

But as Mitchell points out, Visser himself cites examples with the same verbs taking an accusative object: (29)

He sarig folc blissade he(NOM) sorrowful(Acc) people(ACc) gladdened 'He gladdened the sorrowful people' (Ps.Th. 106.32)

Thus, these are to be compared with the Greek pair in (16) above: the object varies between dative and accusative and corresponds to a passive subject in the nominative as in the following: (30)

Da waes Gudlaces gaest geblissad PTCL was G.'s guest-NOM delighted 'At that G.'s guest was delighted'

(GuthA 722)

Thus, if Old English has a true passive, it is usually limited to making subjects of NPs that can be accusative in the active. According to Mitchell, there is one verb where a nominative passive subject corresponds to a dative object in the active. This verb is (ge)fultumian 'help someone or something.' This verb is never cited with an object in anything other than the dative. More precisely, no object of this verb is ever unambiguously accusative. Morphologically the Old English dative and accusative overlapped in their paradigms. The verb (ge)fultumian always takes an object that is either clearly dative or one of the ambiguous forms. A clearly dative example is the following: (31)

Det hi him that they(NOM) them(DAT) 'that they would aid them'

fultumedon would-aid (C/zr.868; Erl.13.22)

Mitchell (1979) lists six examples of personal passives with this verb, one of which is: (32)

ac he waes godcundlice gefultumed but he(NOM) was divinely helped 'but he was divinely helped'

(Bede 342.14)

One solution to this problem is to say that Old English was beginning to show case semi-preservation for this verb. 7.1.5 Comparisons As I have already mentioned briefly, the way theories handle casepreservation in the first place is to make the lexical cases apply prior to

270

Case semi-preservation

the syntactic cases in some sense to be discussed shortly. I will now show that regardless of how this ordering is effected, the case semi-preservation facts require a grammar of a problematic sort. First, none of the possible simple orderings of lexical case before syntactic case will do: (33)

DAT, GEN > ACC > N O M DAT, GEN > N O M > ACC DAT, GEN > NOM, ACC

The orderings in (33) capture the case-preservation effect because lexical case has priority over the accusative in the active and over nominative in the passive. The problem is semi-preservation, because there dative and genitive must be able to apply before or after nominative but before accusative. One solution for theories with extrinsic ordering (and so not an option for GB) would be to substitute for (33) a more partial ordering reminiscent of Anderson's (1969) local ordering: (34)

DAT, GEN > ACC

The lexical cases would be ordered before the accusative but unordered with respect to the nominative. A grammar incorporating the partial ordering in (34) would capture the Faroese and Greek facts: the active would show the lexical case only, because dative and genitive come before accusative. And the passive would show either the lexical case (preservation) or the nominative (non-preservation), because the lexical cases and the nominative are not ordered with respect to each other. The problem here is not descriptive but explanatory: the notation in (34) makes it equally easy to state the opposite state of affairs: (35)

DAT, GEN > NOM

Here the dative and the genitive would alternate with the accusative in the active but always block the nominative in corresponding passives. We would get variation in the active but the lexical case only on the subject of the. corresponding active (so He lost U-DAT, He lost U-ACC but only U-NOM was lost, not It-DAT was lost). This anti-preservation does not seem to occur. Thus the partial extrinsic ordering solution is not sufficiently explanatory, because this type of device contains the power to write as many grammar types for which we have no use as ones for which we do. In the absence of some other way to rule out anti-preservation on

Case (non-)preservation

271

principled grounds, incorporating rule ordering in the case component makes the false prediction that semi-preservation and anti-preservation should both occur. Although GB does not order Case rules, Case in Government and Binding Theory works in a similar way, and so similar problems arise. Since the lexical cases are associated with particular theta roles, these cases are assigned at D-Structure (Chomsky (1981)). Nominative and accusative, on the other hand, are assigned at S-Structure. The priority of idiosyncratic case over syntactic case falls out of the priority of D-Structure over S-Structure: (36)

GB view: D-Structure GEN, DAT S-Structure ACC if not, then movement to get NOM

Whenever the lexical cases are assigned, the accusative (in the active) or the nominative (in the passive) automatically are not, assuming for the moment that NPs can receive only one Case and that case is a good guide to Case. The problem of the priority of cases reduces, then, to two crucial factors. One is the motivation for having Cases on their respective levels, e.g. idiosyncratic Case at D-Structure and accusative at S-Structure. The second factor is how one level comes to be derivationally "prior" to another, e.g. why S-Structure processes as a class may depend on D-Structure processes but not vice versa. Once the motivation for assignment to levels and the relative priority of levels is accomplished, case preservation follows no matter how those two problems were solved. However, as I will show, these problems become more serious when it comes to case semi-preservation. The fundamental problem with case semi-preservation is that we want the lexical cases to have priority over the accusative but not the nominative. There are three main types of solution within the GB framework which I will term the morphological solution, the unordered-level solution, and the ordered-level solution. In each case, I will show that the solution leads to undesirable consequences. The first solution, the morphological, exploits the distinction between morphological case and abstract, syntactic Case. The lexical cases of a language like Icelandic, where morphologically oblique NPs behave as subjects and objects, are simply said to be case but not Case (for

272

Case semi-preservation

Icelandic, this has been proposed by Cowper (1988)). A verb assigns a feature, e.g. +DAT or +GEN, to its internal argument. This feature remains with the argument, and no matter what abstract Case its NP receives, the feature causes a morphological spell-out as dative or genitive, etc. This of course captures case preservation, since the spell-out features are assigned before abstract Case and the actual spelling out of case. As for semi-preservation, the problem begins when we consider the question of the priority of one morphological case over another. This question is completely identical to the case-priority problem already discussed. Solutions which necessarily make features like + DAT and + GEN block realization of an NP as nominative or accusative cannot differentiate between nominative and accusative. But in semi-preservation we want the possibility of realization as nominative but not accusative. The alternatives once again involve excessive power. The features + DAT and + G E N must come with instructions that they can be overridden by nominative but not accusative. However, we could just as easily have stipulated that accusative overrides these features but nominative does not, if the facts went that way. But they do not: we do not get antipreservation. So we are back where we started, in need of an explanation. The second, unordered-level solution would make semi-preservation a property of individual cases. Individual lexical cases would vary between being idiosyncratic and structural. When idiosyncratic, they would be preserved and when structural they would not. This solution has initial attraction, since in some languages, e.g. Japanese, lexical case is not preserved (Kuno (1973), Marantz (1984)): (37)

a. John ga Mary ni soodansita. John NOM Mary DAT consult-PAST 'John consulted Mary.' b. Mary ga John ni soodans-(r)are-ta. Mary NOM John DAT consult-PASS-PAST 'Mary was consulted by John.'

Marantz (1984) proposes structural dative for Japanese. The structural dative is assigned to the OBJ function at S-Structure by the verb. If the logical object is assigned SUBJ (in the passive), then the dative does not appear. One problem with this is that it is not clear what it means for a verb to assign structural dative. Idiosyncratic Case is usually thought to be assigned at D-Structure precisely because it refers to thematic role (or at least individual arguments). Structural Case is supposed to be blind to

Case (non-)preservation

273

identity of the verb and hence is assigned at S-Structure. An idiosyncratic but structural dative violates the motivation for locating cases on their proper levels. If, anyway, we extend Marantz's approach to semipreservation, we must say that the verb assigns either idiosyncratic dative at D-Structure or structural dative at S-Structure. In other words, we have a disjunction over two levels, a very powerful device. In fact, a condition to the effect that a verb may assign idiosyncratic dative if and only if it is passive is quite similar. And such a condition would get us anti-preservation. Before turning to the final option, th,e ordered-level approach, one might be tempted in the present theory to analyze the semi-preserved cases as being optionally [—HA]. When the [-HA] option is not taken, the case is preserved. When the [—HA] option is taken, then the lexical case may not appear on the highest argument. Such a solution, however, has the by now familiar drawback of not being explanatory. First, [—HA] is nowhere else optional. Second, a feature [ + HA] (or even [—LA] needed otherwise) might be able to get (the unwanted) anti-preservation by in effect requiring the idiosyncratic case to be on the subject in order to appear. The final approach in GB would be the ordered-level approach, which is most conventional. Preservation is captured as the across-the-board priority of lexical case. As noted earlier, on the ordered-level approach, the problem reduces to the motivation for assigning the cases to particular levels and to the priority of the levels. Consider the latter problem. There are two ways for one level to be ordered before another. One is that it is temporally or, more properly, derivationally prior. This is extrinsic ordering of levels. Once again, the problem is that in semi-preservation we want the lexical cases to vary with the nominative but not the accusative. If the lexical cases have absolute derivational priority, this is not possible. If we add exceptions, we get global statements of the following form: (38)

Assignment of dative to an argument at D-Structure is obligatory if it would be assigned accusative at S-Structure and optional if it would be assigned nominative at S-Structure.

Such a constraint is both global and transderivational. It is easy to see that with this device it is possible to write an anti-preservation grammar. All that is needed is a statement like the following:

274

(39)

Case semi-preservation

Assignment of dative to an argument at D-Structure is obligatory if it would be assigned nominative at S-Structure and optional if it would be assigned accusative at S-Structure.

This condition is identical to the one in (38) except for the exchange of nominative and accusative. Another way to make D-Structure prior to S-Structure is for the processes at D-Structure to be inherently more specific than those at S-Structure. Then the priority of D-Structure can follow from a general principle like the Elsewhere Condition and need not be stipulated. This of course is ideal and has been widely advocated in theory if not in practice. Making such an approach work for semi-preservation depends crucially on the processes at D-Structure that need priority. Specifically, it should follow from some property of the lexical cases and nominative and accusative that the lexical cases can, in some languages like Faroese, vary with the nominative but not the accusative. The solution proposed in the following will aim to provide the properties of the cases from which semi-preservation but not anti-preservation will follow. The questions raised in this section are: How can one capture optional preservation in a constrained way, especially without predicting the possibility of antipreservation? How can one explain the exceptions to case semi-preservation? How can isolated semi-preservation be captured (French, Old English)? What historical change leads to case semi-preservation? Ideally a case theory should explain all these facts and others. I now turn to how the present theory solves these problems and makes further correct predictions.

7.2

"Scattered cases" and semi-preservation

The solution to semi-preservation in the present theory is extremely simple. It follows from the recognition in case theory that some cases may be "scattered." A scattered rule in phonology is one which must apply both before and after some other rule, depending on what undergoes the rule (Robinson (1976)). In semi-preservation a lexical case applies both "before" and "after" the nominative. In the present theory,

"Scattered cases" and semi-preservation

275

as we will see, a scattered case would be one which has more than one value for LINK. Without revising a single condition or constraint of the theory, the adoption of multiple LINK values for certain cases captures not only the case semi-preservation phenomenon but makes correct predictions about case dissimilation as well. First of all, the idiosyncratic cases bear a special relation to the nominative in both Faroese and Greek. These languages, like Icelandic, are type 1, and so their nominative is the [ARG] case. The type 1 status of Classical Greek was discussed in Chapter 3. Faroese also shows type 1 default behavior for the accusative. It is very similar to Icelandic. A Faroese example of a path NP in the accusative is given in (40): (40)

Vit fara loftvegin. we(NOM) go air-way-the(ACc) 'We are going by air.'

For Faroese, (41a,b) demonstrate the accusative as used for expressions of extent of time and similarly (41c) for extent of space: (41)

a. Tey voru verandi hja honum tann daginn. they(NOM) were being with him(DAT) ten days(Acc) 'They were with him ten days.' b. Hesin fiskur hevur gingid eina dag. this(NOM) fish(NOM) has gone one(Acc) day(Acc) 'This fish has travelled (for) a day.' c. Hann ropti so at tad kundi hoyrast he(NOM) yelled so that it could be-heard langa leid. long(ACc) way(ACc) 'He yelled so that it could be heard (for) a long way.'

Also like Icelandic, Faroese subjects normally occur before the VP (Barnes (1986)), and so I will also posit a preverbal position linker for Faroese specified [ARG] as in Icelandic. 7.2.1 ''Scattered case'' What we want is to capture the following generalization in as explanatory a way as possible: (42)

The idiosyncratic dative is an alternative for nominative (but not accusative) in dative contexts.

276

Case semi-preservation

The present theory derived the priority of cases from their basic distributional features. The way this was carried out was for a constraint, the RC, to refer to the same features which serve to set basic necessary conditions on the use of that case (by the AC). So if a case is to apply both before and after another case, the theory sets strict standards for how much leeway we have to bring this about. Most obviously, we are not free to stipulate the order between two cases, since this would defeat the whole purpose of deriving case priority in the manner of the RC. Furthermore, we should avoid at all costs the possibility of suspending the order between two cases, because this is another very powerful form of stipulation. Ideally, we want the lexical dative with some verbs in Faroese (and similarly for the lexical dative and genitive with most verbs in Greek) to be both more restrictive than the nominative and not more restrictive than the nominative. Since in the present theory case priority ultimately derives from necessary conditions on the appearance of cases, it is necessary to ask how a case could both be more restrictive and not more restrictive than another case. Take the lexical dative with the verb hjdlpa 'help' in Faroese. As demonstrated in §7.1, this dative must appear in object position but is optional in subject position. Compare the dative and the nominative in Icelandic: (43)

Icelandic a. DAT: [ARG|2] (part of the lexical entry of hjdlpa 'help') b. NOM: [ARG]

The minimal difference between Faroese and Icelandic is that the dative alternates with the nominative but not the accusative. This suggests that dative behaves like an alternate nominative (is an "alternative for nominative"), but only for this argument of this verb (a "dative context"). The way to capture this in the present system is for the corresponding dative in Faroese to have two LINK values, one identical to that in Icelandic and an additional one befitting its status as an alternative for the nominative: (44)

Faroese a. DAT: [ARG|2] (part of the lexical entry of hjdlpa 'help') [ARG]

b.

NOM: [ARG]

"Scattered cases" and semi-preservation

277

Note that the second LINK value for the dative does indeed place a necessary condition on its appearance: any NP it attaches to corresponds to an argument. 2 The question is how such double LINK values behave with respect to the three constraints. An important division between the three constraints consists in whether they refer to the notion of restrictiveness. The RC obviously does, as does the LDC, since the LDC looks for the most restrictive linker of a given type on higher arguments than a given argument. But the AC, the constraint concerned with basic distribution, does not refer to restrictiveness. Second LINK values will affect the restrictiveness of a case and so will affect its behavior with respect to the RC and the LDC. The RC-effects will be semi-preservation, and the LDC-effects will be an extra pattern of dissimilation, which I take up in the next section. Multiple LINK values may affect restrictiveness and so behavior with respect to the RC and the LDC, but they do not affect a case's status regarding the AC. Consider again the AC. It states that any LINK value of any linker must unify with that of the NP it links. This means that if a case has two LINK values, they must both unify with that of the NP. The scattered cases I will consider will be ones in which the second value properly subsumes first. Now consider the behavior of such a doubleLiNK-valued case with respect to the AC: (45)

Case: [X|Y] where Y is non-null [X]

The AC will require that if this case links an NP, both values must unify with the LINK value of the NP. Now the first LINK value places a more stringent condition on the case than the second. The first value limits the case not only to the second's contexts but to an even narrower range of contexts. The scattered case passes the AC in just those cases where a case like it but with only the first LINK value would pass the AC. Now note that our hypothesized Faroese dative is exactly of the type of scattered case expected on this theory. And note that it will behave with respect to the AC just like the Icelandic dative which has only the more stringent value. In other words, a scattered case like the Faroese dative is an alternative for nominative in dative contexts, the result we wanted. Unlike the AC, the RC is dependent on restrictiveness. And since double LINK values do affect a linker's restrictiveness, the double LINK value will cause a scattered case to behave differently with respect to

278

Case semi-preservation

the RC than any simple case. The definition of restrictiveness states that a case A is more restrictive than a case B if its LINK value properly contains that of B. If Case A has more than one LINK value, we get two ways to calculate restrictiveness. Case A may be more restrictive than Case B on one of its LINK values but not on the other. The situation we are interested in is where Case A has two LINK values, one of which does properly contain B's and the other is identical with B's: (46)

a. Case A: [X|Y] where Y is non-null [X] b. Case B: [X]

Case A is both more restrictive than B and not more restrictive. In conjunction with the RC, this means that Case A can vary with Case B. This is exactly the situation with the scattered cases like the Faroese dative. The dative has the option of having priority over the nominative or alternating freely with it. The net effect is that the dative and the nominative alternate freely. Note, however that under either value and so under any way of calculating restrictiveness, the dative is more restrictive than the accusative: (47)

Faroese a. DAT: [ARG|2] (part of the lexical entry of hjdlpa 'help') [ARG]

b. NOM: [ARG] c. ACC: [ ] This is exactly the right result: the dative will always have priority over the accusative but will alternate with the nominative in dative contexts, as I now demonstrate. 7.2.2 Semi-preservation It is clear that a lexical case that is semi-preserved is not obligatory but, rather, optional. We have already seen that the type of optionality involved is seemingly difficult to build into a theory of case without excess power, specifically the unneeded power to capture anti-preservation. Recall that both Faroese and Greek show active passive pairs where the object in the active shows only a lexical case (never accusative), but the passive can have either the lexical case or the nominative on the corresponding subject. I called this case semi-preservation. I will illustrate the solution with the sentence in (6), repeated here:

"Scattered cases" and semi-preservation (48)

a. Teir hjalpa they(NOM) help They help him.' b. Hann vard he(NOM) became 'He was helped.' c. ?Honum vard he(DAT) became 'He was helped.'

279

honum. him(DAT) hjalptur. helped hjalpt. helped

For the argument in question, the active object and the passive subject, the following three cases compete: (49)

a. DAT: [ARG|2] (part of lexical entry of hjalpa 'help,' lofa 'praise,' [ARG] etc.) b. NOM: [ARG] c. ACC: [ ]

The dative of the second argument of hjalpa 'help' has two values, which will cause it to behave additionally as an alternative for nominative in dative contexts. When this dative and the accusative compete (in the active), the dative must apply, but when this dative and the nominative compete (in the passive) either can apply. Consider first the linking for (48a = (6a)) in (50): (50)

hjalpa ( x NOM: [ARG]

y ) 'help' (active, dative object, DAT: [ARG/2] (48a = (6a))) [ARG]

_ VP: [ARG] ACC:

[

]

*NOM: [ARG]

*_VP: [ARG] [ ]

ACC:

The nominative is out for the second argument by the LDC, because it has the same LINK value as the most restrictive linker of its type that passes the AC for the first argument, i.e. nominative. The dative, by virtue of the first LINK value, passes the LDC for the second argument. The dative under either LINK value blocks the accusative. Since the dative is more restrictive than the accusative on either of its LINK values, the dative will always block the accusative in the active, the correct result. This gets the obligatoriness of the lexical case in the active.

280

Case semi-preservation

For the passive of these verbs, however, the system developed here yields two possibilities, dative (51, 52a) and nominative (52b): (51)

hjalpa ( xA

y ) 'help' (passive, dative subject, (48b = (6b))) DAT:

[ARG/2]

NOM: [ARG]

_ VP: [ARG] [ ]

ACC:

(52)

a. hjalpa ( xA y ) 'help' (passive, dative subject, (48b = (6b))) DAT: [ARG] NOM: [ARG]

_VP:

[ARG]

[ ] y) 'help' DAT: [ARG]

ACC:

b. hjalpa ( xA

(passive, nominative subject, (48C = (6c)))

NOM: [ARG]

_ VP: [ARG] [ ]

ACC:

Here all three cases compete. Comparing the cases in (49), it is clear that the accusative is blocked by either dative (as in the last example) or nominative. On one LINK value, the first, the dative is more restrictive than the nominative and blocks it, and we get (51). On the other LINK value, the dative and the nominative are neither one more restrictive than the other, and so either (but not both) may apply. The net result is that a scattered dative allows both the impersonal and the personal passive. This solution provides a fully general account of case semi-preservation. The second LINK value expresses the fact that a lexical case may be an alternative for nominative. All lexical cases which are scattered in that they have a second LINK value [ARG] will lead to case semi-preservation. All such cases will vary with the nominative but not the accusative (cf. (49)): (53)

a. CASE: [ARG|X] [ARG]

b. c.

NOM: [ARG] ACC:

[]

Any case of this form will be more restrictive than the accusative and so be the only option in the active. Such a case has the option of not being

''Scattered cases'' and semi-preservation 281 more restrictive than the nominative, and so has the option of not being preserved. Most importantly, this approach rules out the type of grammar that would be needed for a language that had "anti-preservation." An "antipreservation" language would have actives with either the lexical case or the accusative but passives with only the lexical case. Such a language would show both the pattern in (54 = (1)) and the pattern in (55): (54)

N P I - N O M V-ACTIVE N P 2 - D A T <^> N P 2 - D A T V-PASSIVE

(55)

NPpNOM V-ACTIVE NP2-ACC <£> NP2-DAT

(by-NPi)

V-PASSIVE ( b y - N P ^

As I showed in §7.1, if one allows the use of extrinsic rule ordering or global transderivational constraints, one can write a grammar for such a language. Such grammar fragments were given in §7.1.5. In the theory developed here, the grammar for an anti-preservation language is unwritable because of the forms of the rules: the only choices for decreased restrictiveness are cases that fit (53a), repeated here, or possibly (57): (56)

CASE: [ARG|X] [ARG]

(57)

CASE': [ARG|X]

Given a case with two features in its LINK value, there are only two ways for a second LINK value to give that case the option of being less restrictive. For the cases we have been considering, the only two logical possibilities are as in (56) and (57). The case in (57) will indeed be an alternative for the default accusative, but this actually implies that it will alternate with the nominative as well. On the first LINK value it blocks the nominative, but in the second nominative blocks it. It is not possible to formulate a lexical case such that it will vary with accusative but not nominative. The theory thus allows us only full preservation or semipreservation. The state of affairs in (54-55), anti-preservation, is unwritable in the theory. That anti-preservation is not an option follows from the form the theory imposes on grammars. Finally, it is worth pointing out that by making the account of semipreservation depend on individual lexical cases, we have a way of accounting for the isolated semi-preservation languages as well. The only difference between Faroese and Classical Greek on the one hand, and French and Old English on the other, is that in the latter only a few

282

Case semi-preservation

idiosyncratic cases (or prepositions) are scattered, whereas in the former, being scattered is the norm for idiosyncratic cases. The seeming exception for goals in Classical Greek (and possibly Faroese) follows, if we analyze the goal dative as non-scattered. The semi-preservation effect provides yet more evidence that different datives, while morphologically the same, are syntactically distinct.

7.3

Further implications

This analysis makes predictions from the LDC as well. The existence of scattered cases with more than one LINK value leads to semi-preservation, because such cases have two behaviors with respect to the restrictiveness relation and so to the RC. Restrictiveness is equally important to another major constraint on building case frames, the Linking Dissimilation Constraint (LDC). If scattered cases behave differently for blocking through their two-way restrictiveness, then this two-way restrictiveness should have implications for the case's behavior for the LDC. Specifically, the LDC in a type 1 language like Faroese questions the linkers on lower arguments with respect to the most restrictive linker of that type for higher arguments. If the most restrictive linker of that type for the higher argument is a scattered case, i.e. it has two LINK values, we should get two patterns of dissimilation, each depending on one of the LINK values of the scattered case. The lexical cases that participate in semi-preservation are assigned to the logical object. Since this was the lowest argument in the sentences above, the impact of the second LINK value on the LDC was not apparent, because there was no lower argument. If, however, a scattered case is assigned to the logical subject, the LDC, in questioning the linking of a lower argument, might have to consider the scattered case when looking for the most restrictive case applicable to the higher argument. Two patterns of dissimilation should result, one from the first LINK value, and one where the lexical case behaves like a covert nominative: (58) verb ( x y) CASE: [ARG|X]

NOM: [ARG]

[ARG] NOM: [ARG]

_VP: [ARGJ

_VP:

ACC:

ACC:

[ARG]

[ ]

[ ]

Further implications

283

The case designated CASE in (58) is scattered. The question is how a scattered case interacts with case on the lower argument. The lower argument is eligible by the AC for two linkers of the type case, nominative and accusative. By the LDC, the one actually linking the second argument may not have the same LINK value as the most restrictive linker applicable to the higher argument. The scattered case will behave two ways with respect to restrictiveness. On the first value it is the most restrictive. Then the nominative on the second argument does not have the same LINK value as the most restrictive linker applicable to the higher argument (the dative). On the dative's second LINK value, it is neither more nor less restrictive than the nominative. This means that both cases, dative and nominative, are the most restrictive case applicable to the highest argument. The nominative on the lower argument then does have the same LINK value as the most restrictive linker applicable to the higher argument, and so cannot be used under this way of checking the LDC. The net effect is that we should be able to get nominative or accusative on the second argument. This expectation is borne out by DAT-ACC verbs in Faroese and Old English. The accusative is expected with these verbs if the dative behaves like a nominative with respect to the LDC, as I now show. Barnes (1986: 34ff.) documents verbs in Faroese that once took a DAT-NOM frame as in Icelandic and which now tend to take DAT-ACC. Compare Icelandic (59) and Faroese (60) (Barnes (1986: 34) quoting Henriksen): (59)

Henni hefur alltaf frott Olafur leidinlegur. her(DAT) has always thought Olaf(NOM) boring(NOM) 'She has always considered Olaf boring.'

(60)

a. Honum tokti him(DAT) thought 'He considered the b. Honum tokti him(DAT) thought 'He considered the

skatturin tax-the(NOM) tax too small.' skattin tax-the(ACc) tax too small.'

ov litil. too small(NOM) ov litlan. too small(ACc)

Icelandic pykja 'think' in (59) is cognate to Faroese tykja 'think' in (60). This verb has historically taken DAT-NOM (Old Norse pykkja 'think' did as well). In Modern Faroese, however, the DAT-NOM pattern (60a) is considered old-fashioned sounding according to Barnes, with some of his informants even labeling it ungrammatical. The DAT-ACC variant is an

284

Case semi-preservation

innovation. In the present theory this can be captured by the following. For some speakers the experiencer dative has become scattered: (61)

a. DAT: [ARG|EXP] [ARG]

b. c.

NOM: [ARG] ACC:

[]

In the following, assuming for the moment that the dative blocks the nominative for the highest argument, there are two possibilities for the second argument, illustrated in (62): (62)

a. tykja ( X|EXP

y ) 'think'

DAT: [ARG/EXP]

NOM: [ARG]

[ARG] NOM: [ARG]

*_VP:

_VP:

ACC:[

[ARG]

[ ] b. tykja ( X|EXP

[ARG]

]

ACC:

DAT: [ARG/EXP]

y ) 'think' *NOM: [ARG]

[ARG] NOM: [ARG]

_ VP: [ARG] [ ]

*_VP: ACC: [

[ARG]

]

ACC:

If the LDC, when questioning a nominative linking for the second argument, takes the first value of the dative as in (62a), then the nominative will be fine, because it is not the most restrictive case applicable to the higher argument. The result is (60a) with the DAT-NOM frame. If, on the other hand, the LDC uses the second value of the dative (62b), then the nominative (along with the dative) is the most restrictive case applicable to the higher argument. The nominative is ruled out on the second argument, and we get the DAT-ACC frame of (60b). The extra LINK value in the present theory gets another difference between Faroese and Icelandic, the additional DAT-ACC possibility in Faroese. If the experiencer dative is scattered in Faroese, this makes the further prediction that it should alternate with the nominative, i.e. that the verbs which show DAT-ACC frames should also show Nominative Substitution. This prediction is fulfilled: Barnes (1986: 40) notes that the tendency to replace preverbal dative with nominative is a strong trend in Faroese. There is, however, a great deal of variation from speaker to speaker.3

Further implications

285

One class of verbs illustrates the interaction of the two constraints, RC and LDC. Faroese has a class of two-place verbs which evidently allow either argument to be the subject and so should be analyzed as allowing both argument orders. This class is very similar to the class of verbs in Icelandic discussed in the Chapter 5 which show the same double-order behavior (e.g. standa til boda 'be on offer' and hugnast 'like'). One such verb is klceda 'suit,' where there is evidently a great deal of variation among speakers, but the following is the most common pattern: (63)

a. Sjalvandi naturally 'Naturally b. Sjalvandi naturally 'Naturally

klaedir hasin kjolin taer. suits that dress(NOM) you(DAT) that dress suits you.' klaedir taer hasin kjolan. suits you(DAT) that dress(ACc) that dress suits you.'

The sentences in (63) both show inversion, and either NP can immediately follow the verb like a subject. The two most popular frames are NOM-DAT and DAT-ACC (DAT-NOM presumably is marginally possible but archaic sounding, as with other verbs). This verb will have to allow both orders of the arguments: (64)

klaeda ( X|EXP y ) 'suit' —> klaeda ( y X|EXP ) 'suit'

The linking for the order on the left in (64), which is behind (63b), is just like that of the last example in (60,62). What is new is the linking for the second order. Under the order with the experiencer second, we get the linking for the first sentence (63a): (65)

klaeda ( y

X|EXP ) 'suit'

NOM: [ARG]

DAT:

[ARG/EXPJ [ARG]

_ VP: [ARG] ACC:

[ ]

*NOM: [ARG]

*_VP: [ARG] [ ]

ACC:

With the experiencer second, the nominative is clearly disallowed on the experiencer by the LDC, because it is the most restrictive linker applicable to the higher argument, here the non-experiencer. This leaves the dative competing with the accusative for the experiencer. Under either LINK value, the dative is more restrictive than the default accusative and so takes priority. Just as the idiosyncratic cases that showed semi-

286

Case semi-preservation

preservation always blocked the accusative in the active, so the scattered dative of experiencer should block the accusative when they compete in the second (righthand) order in (64). The accusative should be out, and this is what the discussion in Barnes (1986: 37) implies.4 Finally, we expect that the scattered dative should allow nominative by the RC. Continuing with the verb klceda 'suit,' we find that the nominative is possible for most speakers: (66)

Sjalvandi klaedir tii handa kjolan. naturally suits you(NOM) that(ACc) dress(ACc) 'Naturally that dress suits you.'

This possibility follows from the second (67)

klaeda ( X|EXP DAT: [ARG|EXP]

LINK

value of the dative:

y ) 'suit' *NOM: [ARG]

[ARG] NOM: [ARG]

_ VP: [ARG] [ ]

*_VP: ACC: [

[ARG]

]

ACC:

The nominative on the experiencer does not violate the RC on the second LINK value of the dative, and so the nominative-accusative frame is also generated.5 Interestingly, Middle English shows an identical phenomenon. Allen (1986: 398f.) cites the following, where ACC/DAT represents the single case that is the coalescence of the older dative and accusative: (68)

(69)

for dat him ereowe ow for that him pitied you(ACc/DAT) ' . . . because he pitied you.'

(AW, p. 36.11)

swetest him dunched ham sweetest him thinks them 'He thinks them the sweetest/They seem the sweetest to him.' (AW, p. 101.7)

As in Faroese, the dative/accusative arises on the second argument where the nominative is expected. Allen describes this development as the shift to syntactic accusative. On the account here, the experiencer dative is scattered, leading, by restrictiveness and the LDC, to the additional ACC/DAT-ACC/DAT frame, as in Faroese. This can be seen as an extension

Diachronic predictions

287

of the very limited semi-preservation seen in late Old English (see §7.1.4 above). Thus two traditional puzzles in the history of English, semipreservation and ACC/DAT-ACC/DAT sentences, receive a unified account.

7.4

Diachronic predictions

This account explains the path and directionality of the rise of case semipreservation as well. Recall that the general diachronic prediction that follows from the restrictiveness-direct linking view of case is that the restrictiveness of the linker applying to a given argument type should decrease over time. The shift from full preservation to semi-preservation is restrictiveness decreasing, because when a case becomes scattered, the argument it applies to is linked on average by less restrictive linkers. Consider the change for hjdlpa 'help.' The change is from full preservation to semi-preservation, and so on the present theory the change is from a normal idiosyncratic dative to a scattered dative. The change is from (70a) to (70b): (70)

a. b.

DAT: [ARG|2]

c.

NOM: [ARG]

DAT: [ARG|2]

(part of the lexical entry of hjdlpa 'help') (part of the lexical entry of hjdlpa 'help')

[ARG]

Thus the hypothesis is that the Faroese case in (70a) is the descendant of a case like that in (70a). This is confirmed by the judgment mentioned above that the preservation variants like the one in (6) with hjdlpa 'help' are more formal and literary. Often forms that are more formal reflect older norms. Further, as we saw in the last chapter, Old Norse, the common ancestor of Icelandic and Faroese, was like Modern Icelandic in having neither semi-preservation nor dative-accusative sentences. Because the scattered cases are themselves on average less restrictive and because they allow less restrictive cases to apply, all the changes resulting from a case's becoming scattered follow on the present theory. The rise of semi-preservation and the DAT-ACC frames follows from the trend towards decreased restrictiveness of linker applying to a given argument.

288 7.5

Case semi-preservation Conclusion

In this chapter we have seen that the present theory can handle a wide variety of new facts about case. By positing scattered cases with more than one LINK value, the theory automatically captures semi-preservation. Most of all, it precludes capturing the non-occurring anti-preservation. The theory predicts additional LDC effects with such cases. The predicted dissimilations seem to occur and interact with case blocking in the expected way. As was generally true so far for linking changes, the restrictiveness-based theory captures the direction and path of the changes leading to the rise of scattered case effects. As before, the same restrictiveness-based theory that provides a constrained synchronic account also accounts for the rise of the pattern as well.

8

Conclusions

In order to put the preceding proposals into perspective, some concluding remarks on the nature of linking theory are in order. What I have proposed is a theory both of how certain surface properties are related to each other and of how these relate to the wide array of properties associated with grammatical functions, the results of linking on our system. The theory derives grammatical functions largely on a basis that is closely related to surface form. The theory must be tested language by language, and cumulatively, to see whether it handles surface and grammatical function facts well and, most importantly, simultaneously. First, it is important to keep in mind what the consequences of distinguishing morphological and syntactic case are. Most theories make this distinction one way or another. Something in the theory will be distinct both from morphological case and argument structure. As we have argued, distinguishing grammatical functions from morphological case is one such tool but one which does not put much of a constraint on the relation of morphology to argument structure. Direct linking opposes linkers to morphological case and to argument structure. Different morphological cases may not be instantiations of the same linker, but we do allow the same case to reflect different linkers, different syntactic cases. Once we have allowed this, though, one might object that the analysis of a given language is by no means unique. How can we tell if two instances of dative reflect different linkers - different syntactic datives, two rules - or not? As a preliminary matter, it should be remembered that the same problem arises on conventional grammatical function- or structural position-based approaches. The reason is that on such approaches we are doing two analyses with respect to NPs. On the one hand, we have to decide which grammatical relation they bear and on the other hand we have to decide which rules should assign cases. The latter may not be so controversial once the former analysis is done, but the former analysis is 289

290

Conclusions

not predetermined either. For example, in a transitive (two-place) clause, if one argument NP is nominative and the other is dative, should we say that we have a subject and an indirect object with a dative rule for indirect objects? Or should we say we have a subject and an object with dative rules for both indirect objects and the direct objects of certain verbs? It is one of the virtues of the present system that such questions receive at least a partial answer. The reason for this is that which linker analysis is done makes a great difference in several areas at once. Most importantly, LINK values, restrictiveness, and the three main constraints mean that a given analysis has many consequences at once. For example, if a case has been analyzed as non-default, it should not be used multiple times per clause. At this point one might object that this prediction is very weak since we can always posit multiple linkers with the same morphology. But, this is not actually the case. The reason is that the analysis of a language that produces the inventory of linkers is, like other analyses, subject to considerations of economy on the one hand and the interconnectedness of the various predictions on the other. Extra linkers need some justification in the first place. If, taking a set of facts such as double use of what seems a non-default case, we set up two linkers with the same morphology, this has a variety of implications. First, the case that is the default should not show dissimilation patterns. If it does, the analysis can be rejected. In addition, as we have seen, certain facts are particularly good indicators that a certain case or other linker is the default. First, as Gaedicke recognized over one hundred years ago, if a wide range of disparate "functions" is associated with one particular item of morphosyntax, then this is a strong candidate for default linker. Second, if two cases alternate they are strong candidates for linkers standing in the relation less to more restrictive. Third, combining the first two points, if one case or other linker alternates with a wide and disparate set of linkers this is very strong evidence that this is the default. Fourth, historical evidence associated with such alternations will indicate which linker is the linker the trend is toward. This is expected to be the less restrictive linker. Once the default is chosen for each linker type, the other linkers of that type should in non-split systems be non-default linkers. Split systems introduce a new element of the analysis. One question necessarily left open here is exactly what range of splits is possible.

Conclusions 291

As we can see, the best that can be hoped for in accounting for patterns of morphosyntax is an account of relative productivity and markedness. Some cases, the less restrictive ones, are more productive, and this leads us to expect certain kinds of data. Correspondingly, some linkers "cost" more. Thus while it is feasible to write a grammar with only lexical cases, including idiosyncratic nominatives for every verb and idiosyncratic accusative (limited to non-highest arguments) for every transitive verb, such a system clearly misses something. New verbs are lexicalized with nominative and accusative, but such a system misses this. The same principle applies but less obviously to the setting up of a default, etc. The analysis in the present system in terms of default case, etc., is supposed to capture facts like this about productivity. The fact that new verbs are lexicalized with nominative subjects and the fact that a large number of seemingly unrelated cases alternate with the accusative are facts about productivity that the present theory is meant to capture. In addition, the present theory relates these facts about productivity to one another. Some facts will be so salient like these as to be testable in individual languages. Some predictions are more subtle, like the predictions of correlation between non-dissimilation and default behavior (e.g. multiple use per clause should correlate with use on a wide range of adverbials and/or use as an alternative for optional linkers). The present study can only have begun to test such predictions. In these respects, theories about the patterns of case marking and surface form are not so different from the rest of syntax. While it is common to think of Wh-Movement and other pheneomena as being exceptionless, most theories do provide for lexical stipulation and exception. Parameterization, especially at the level of lexical item, is as a possibility no different from the possibility in case theory of stipulating case frames. Once the possibilty of stipulation is admitted, care must be taken to ensure that the predictions that remain interlock enough to ensure that they are indeed strong predictions and not mere ad hocery. The reader is invited to judge whether the predictions, synchronic and diachronic, in the preceding pages meet this test. Finally, it should be remembered that the analysis of the inventory of linkers is connected to a vast set of facts that are implicated in assignment of grammatical relations. Once we have set up an inventory of linkers, the present theory makes strong claims that we can predict the grammatical function status of an NP and all the grammatical behavior that this entails. If the inventory of linkers, which are specified for obliqueness,

292

Conclusions

cannot handle these facts or handles them less well than another theory, then so much the worse for the theory. That there are patterns in morphosyntactic systems, differences in productivity in morphosyntactic devices, and some relation between morphosyntax and grammatical relations seems to be implicit in many theories. Direct linking, if successful, can make some of these notions more precise and the predictions more powerful. I hope that the present study has made a case for some synchronic and diachronic generalizations and has raised some new questions about case theory.

Notes

I

Introduction 1 The scheme in (4) is, however, deceptively simple, since a theory based on grammatical functions is likely to have more elaborate grammatical functions and simpler morphosyntactic rules, but theories of direct linking will have more elaborate morphosyntactic rules and simple definitions of grammatical functions. See the discussion at the end of §1.2.2 for further explanation of the differences between the two approaches and a discussion of why direct linking is "direct." 2 It will be sugested in the next chapter that syntactic obliqueness of NPs should not be stipulated at argument structure but rather should be a feature of linkers (like dative) that is passed on to the argument it links. After I developed this, Kiparsky (personal communication) informed me that he had modified his view of obliqueness. In the latest version of KLT, grammatical linkers may not link demoted arguments, but obliqueness comes from lack of grammatical linking. For the different consequences of the position taken in this work and Kiparsky's more recent ideas, see Chapter 5. 3 Use of "argument" varies. In Kiparsky (1987), arguments are NPs at Lexical Structure which are linked to thematic roles, and on occasion thematic roles are referred to as "arguments." Later I will adopt a syntactic definition of argument along the lines of the familiar one from Categorial Grammar. 4 For Kiparsky "highest" is defined in terms of the thematic hierarchy, so "highest" is "thematically most prominent." In this book "highest" will be defined at a syntactic level of argument structure and will mean "highest in the order of arguments." See §1.2.3 for some discussion of argument ordering without thematic roles. 5 Note that "highest 9-role" is hardly "semantic," since no mention of 0-role label is used. Rather what is important here is that by is syntactically oblique. There are many possible ways of handling the by-phrase of a passive (this has been a topic of great interest in GB, see Baker (1988) for extended discussion and references). One very traditional way would be to say that by is part of the passive rule. Another, more like current GB's treatment (and, incidentally, much like Panini's approach) would be to say that the passive morphology itself expresses the agent. Then the by could be considered an adjunct marker without semantic content of its own. 293

294 Notes to pages 12-22 6 GB handles case and word order facts in a slightly different way. If word order is the "linker," GB handles this by having an abstract Case with a name reminiscent of a morphological case be assigned to a given position. The NP receiving the Case must occupy that position to receive the abstract Case, thus ensuring that the word order facts are captured. 7 Kiparsky also claims as a consequence of the anaphoric status of agreement that languages that have object agreement have subject agreement, i.e. no roles on the hierarchy are skipped by the inventory of agreement markers. In other words, a language can mark a non-highest non-demoted role with agreement only if all other higher non-demoted roles have been linked by agreement. While seemingly simple, this prediction is very difficult to evaluate empirically for two reasons. One is that distinguishing between object agreement and obligatory object clitic pronouns in a given language is a subtle business. The other, more serious difficulty stems from the notion of "availability." To save the prediction one could always claim that higher roles are unavailable for linking by agreement, leading to vicious circularity (agreement links the highest available role; a role is available just in case it shows agreement). For example, absolutive agreement in some Caucasian languages including Kabardian (see Chapter 3 and references there) seems to be oriented towards the lowest available role. It might be possible to claim that in such a language, all the higher roles are demoted or otherwise unavailable for linking by absolutive agreement. The notion of availability needs careful attention, but rather than explore this possibility, I will not assume a special highest-available-role orientation for agreement in this work. 8 One NP can be linked by more than one type of linker, e.g. by both case and word order. Thus an accusative and a dative NP might both be linked by the same word-order position (e.g. within the VP) but must be linked by different cases. NPs differing in case marking must be linked by different cases. 9 The implication in the other direction - that if two NPs show the same surface form, they are linked by the same linker - need not be true. If it were, there would be no need for a distinction between morphology/surface form and morphosyntax. 10 Case assignment to structural position in GB is very similar to case assignment to grammatical function. Thus assignment of NOM Case to Spec-IP (i.e. subject position) is like assignment of nominative to subject in LFG or RG. The structural positions in GB are not all surface positions and so mediate between argument structure and surface form just as grammatical functions do in other theories. 2

Argument case and case alternations 1 The Modern Icelandic data in the following are primarily from Svavarsdottir (1982) and from informants (primarily Thora Arnadottir). Any errors in the following are of course the author's fault. Also, as far as synchronic data goes, I have not relied on Halldorsson (1982), since in that study the author asked the informants not for their own intuitions but what they thought they had

Notes to pages 22-37

295

heard. Svavarsdottir (1982), however, presented 11-year-olds with sentences in a story and had the subjects replace a name that does not show its case with pronominal forms that do. For example in (2), Svavarsdottir (1982: 31, 5Iff.) indicates that for the verb langa 'long' the nominative occured 1.5% of the time (in one location 5.3%). Note that vanta 'lack' discussed below shows the same level of nominative in Svavarsdottir's study, and Andrews (1990: 204) likewise finds some of his subjects labelled (99b) below (with vanta and the nominative) grammatical as an alternative to the accusative. If, nevertheless, the nominative with langa and vanta were the result of hypercorrection, another verb like minna 'remember' or dreyma 'dream' with much higher nominative occurrence (and with variation between three cases) could be substituted for langa without affecting any point made in this study. 2 Actually, given that by Dative Sickness the subject of vanta 'lack' can be dative, some of the usual tests, including control, do not contribute to proving the existence of accusative subjects. The covert subject in (3a) might be dative (by DS) or even nominative in some dialects. Also, the word efni 'material' which appears in Andrews' example is morphologically ambiguous between nominative, accusative, and dative. That it really is nominative can be shown by replacing it with another noun like peningar 'money' that has different forms for these cases. 3 A dative non-experiencer subject is exemplified in (i): (i)

4 5 6 7

8

Ovedrinu linir. storm-the(DAT) abates 'The storm is abating.'

The referent of the subject, ovedrinu 'storm,' is neither entailed to be aware of anything nor to be a sentient being and so is not an experiencer. For accusative non-experiencer subjects, see (18), (20), and (21) below. Furthermore, it has proved very difficult to find sociolinguistic correlates of DS and NS (Svavarsdottir (1982)). For definitions and discussion of feature structures, paths, subsumption and unification, see Carpenter (1992: 33ff.). See Carpenter (1992: 41) for a formal definition of subsumption. See Carpenter (1992: 46) for a formal definition of unification. Basically, the unification of A and B is the smallest feature structure that subsumes both A and B. Furthermore, the argument categories or cases based on them may explain generalizations generally thought to be at least indirectly semantic. For example, Bernodusson (1982) and Yip, Maling, and Jackendoff (1987: 225) observe that Icelandic verbs with lexically case-marked subjects do not passivize and suggest that this is because those subjects are non-agentive. Grimshaw (1990: 118) points out that this cannot be the correct explanation because nonagentive subjects do not generally block passivization. An alternative is that lexical case, much of it based on argument category, blocks passivization.

296

Notes to pages 39-41

9 Other types of systems are possible in this theory. As we will see in the next chapter, in Japanese, the accusative is restricted and the nominative is the default (roughly, accusative may be used only once, but multiple nominatives are possible). Further, if conditions on linking are "reverse" from Icelandic, one gets languages like Kabardian with ergative as the default case (an ergative-absolutive pattern with a range of abverbial ergatives rather than absolutives). German will turn out to be of the same type as Icelandic with respect to the nominative and the accusative, so for now I will only be concerned with the special case represented by Icelandic. See Chapter 3. 10 1 leave open the question of whether the linker for languages like Icelandic and German is nominative case or agreement or both. I l l learned of Gaedecke's work after the completion of Smith (1992) and have added the present discussion to the text. 12 Icelandic nominative is not a zero case; for instance, the masculine singular nominative hestur 'horse' consists of the stem hest- and the masculine nominative singular ending -ur. Nevertheless, Icelandic nominative, like nominative in many other languages, has a special status in that it is the only case compatible with verbal agreement. It is also the case for special NPs, such as those in apposition and left dislocation. 13 Zaenen, Maling, and Thrainsson (1985) note, however, that under special circumstances the accusative of path can be passivized. If it is the only postverbal NP, then passivization is possible, as in (ib): (i)

a. Hann keyrdi ]sessa leid. he(NOM) drove this(Acc) route(ACc) 'He drove this route.' b. l>essi leid hefur aldrei verid keyrd. this(NOM) route(NOM) has never been driven 'This route has never been driven.'

I follow Zaenen, Maling, and Thrainsson (1985) in taking the accusative of path to be a non-argument which can be promoted to argument if there is no logical object. Linking will automatically make this logical object the grammatical object in the active and the grammatical subject in the passive. 14 In Smith (1992: 329) I propose that accusative is the default case in Faroese as well, although Faroese presents additional facts that can be explained by a simple extension of the present theory. See Chapter 7. 15 In Icelandic, special adverbials can be dative and genitive marked. This is not a problem on the present approach as long as accusative as default has the following advantages: (i) Having accusative as the default leads to the greatest economy of rules. Accordingly, if there is no unifying characteristic of various uses of the case, this makes it a likely candidate for default, (ii) The default case is expected to be productive in the sense that if an adverbial case or proposition becomes optional or is lost, the default case is expected to take over unless a new linker arsies by grammaticalization. This is true of time adverbials (e.g. the German preposition wdhrend 'during' alternates with bare

Notes to pages 41-64

16

17

18

19

20

accusative NPs). See also Chapter 6. (iii) All else being equal, a default case in a given language should be the one that fits in best with a cross-linguistic theory, e.g. the one presented in the next chapter. NPs in apposition or left dislocation need some special device in order to be linked, or, alternatively, they fall outside the domain of linking. I leave the treatment of these special NPs for further work. Case Stacking (where an NP receives two cases, e.g. an experiencer gets dative and nominative simultaneously) in Korean (Gerdts and Youn (1988)) is incompatible with such a statement in its strongest form. Nonetheless, no doubling of cases is the unmarked situation. Note that the ordering constraint in Korean on the cases is, in our terms, more to less restrictive. The grammar will of course contain a constraint that an NP bearing this feature appearing anywhere other than as the sister of the finite verb (or in the appropriate postverbal position in an inverted sentence, etc.) will be ruled out. I use "_V" to be neutral about the question of VPs in Icelandic. Our "_V" could be replaced by "_V"" (with, more importantly, a constraint based on sisterhood to VP) without affecting any points made here. Step (iv) could have been simpler if we had said "Delete from the list any linker that has already been used." Constructions like Equi (Chapter 5) and Case Semi-Preservation (Chapter 7) call for the formulation in terms of application "in principle" rather than application in fact. Nominative can be used more than once in Predication, as illustrated by the German sentence in (i): (i)

21 22 23

24 25

297

Der Mann ist ein guter Arzt. The(NOM) man is a(NOM) good(NOM) doctor 'The man is a good doctor.'

The theory above is about the linking of NPs that function as semantic arguments of a predicate, but predicative NPs are predicates themselves. The LDC is not expected to operate across predicate boundaries. However, a passive is interpreted as if the highest argument were still there. For other accounts of nominative objects, see Zaenen, Maling, and Thrainsson (1985) and Yip, Maling, and Jackendoff (1987). One alternative would be to have lexical accusative and other idiosyncratic cases as lexically specified and have a further condition that lexically specified case blocks other cases. However, this can be seen as another instance of the specific-over-general principle, and the present approach is an attempt to capture this. Note that, for many verbs in German, NS is an alternation of dative and nominative with no possibility of accusative, e.g. with ahnen 'perceive.' Generally, the data used in constructing the rest of LCSs, not just in their annotations, seem to come in large part from surface expression, i.e. syntax. Van Valin (1993: 160) makes a similar point about the causal chains proposed in Croft (1991).

298

Notes to pages 65-91

26 One might argue that no two verbs are truly synonymous. If this turned out to be true, then the semantic selection approach could conceivably make idiosyncratic case dependent on that part of a verb's meaning which is unique to that verb. The question again reduces to whether there really are semantic facts independent of the syntax upon which a truly non-syntactic explanation could be built.

$

A typology of case systems

1 Kiparsky (1989) uses the feature [—H(ighest)R(ole)], "cannot appear on the highest role," for Japanese accusative. Kiparsky uses the other (i.e. + ) value as a feature for ergative case, i.e. [ + HR] means "must appear on the highest role," even though this is stronger than the negation of the negative value (one would expect [ + HR] to mean "can appear on the highest role"). However, here [-HA] is the true negation of [ + HA]: [-HA] means "cannot appear on the highest argument," and [ + HA] means "can appear on the highest argument." The [-HA] proposed here just refers to not being at the top of the argument list. Thus [—HA] is the only value that may be mentioned by rules or constraints other than the redundancy rule filling the redundant [ + HA] value. Alternatively, [-HA] is a privative feature (the label being a mnemonic), and there is no immediate need for a [ + HA] on linkers. My ergative will also differ from that of Kiparsky (1989). Since Smith (1992), Woolford (1993) proposed an "Accusative Case Blocking" Principle which bears some resemblance to Kiparsky's and my restrictions on accusative. I will not compare these approaches in detail except to note that the restriction here is part of the definition of accusative in type 2 languages of the typology of this chapter. This part of the accusative will affect its behavior with respect to the RC and the LDC as well as the AC (the AC's effect on a case with the restriction barring use on highest arguments is the closest analogue to the effect of Woolford's principle). Furthermore, GB Case is not constrained morphologically to the extent that case as a linker in direct linking is. 2 Of course it is not necessary to alter the name of the feature depending on its interpretation. The Preference Parameter to be introduced below is really about how the one feature [—XA] is interpreted. I will, however, use [—HA] and [—LA] for clarity. 3 I am of course assuming for now that there would be no idiosyncratic nominative. For the relation between predictions like (38) and surface form and for the corresponding nature and strength of the predictions of the theory, see Chapter 8 below. 4 They only seem to be passivizable when the accusative is interpreted as affected, as discussed for Icelandic by Zaenen, Maling, and Thrainsson (1985) (and see §2.4.2.3 above). It seems that the same holds for Greek. As for Icelandic, I assume that these special passives should be treated the same as pseudo-passivization in English and that the adjunct is promoted to argument status.

Notes to pages 94-106

299

5 I leave open the question of how to derive the causatives (see Comrie (1976), Baker (1988), Mohanan (1994), Alsina (1991) among others). I also leave open what the LINK value is of the "dative" (-ni) that appears in the causative and the passive. It may be an optional part of causativization and passivization or some kind of adjunct. Or it may support the idea that what I have been calling dative of GOAL is really dative of the middle argument of a three-place verb. 6 We do have to say something special about the Japanese positional linker. When the option of linking by position is taken, the case is optional. In Icelandic, by contrast, if a case was more restrictive than a position, both apply. In Japanese here, the case and the position are neither more restrictive, and either can apply. One possibility is that linking by case and by adjacency to V are incompatible in Japanese. This would not preclude an -o-marked NP from appearing next to the verb without being linked by its position. 7 It is worth stressing that the domain of linking excludes NPs in left dislocation and apposition and that the LDC is only a constraint on NPs that are arguments of the same predicate. Thus a sentence with two nominatives that are not co-arguments is not evidence one way or another about the linking system. Based primarily on evidence of syntactic correlates of syntactic patterns, Kuno (1973: 37ff.) distinguishes two kinds of -ga, exhaustive listing -ga and neutral description -ga. To translate his insights into the present approach, the facts about semantic interpretation would probably best be treated separately from those of linking. Linking is not a semantic structure but the satisfaction of three constraints on the relationship of (syntactic) argument structure and morphosyntax. Semantic interpretation can be done in a conventional way from surface structure as in GPSG. With this in mind, it does not seem that within the domain of linking and especially the domain of co-arguments, a syntactic distinction between two types of -ga is called for at all. 8 There is one other possibility: that the ABS-DAT verbs exceptionally do not allow ergative. This might make them analogous to the NOM-NOM verbs in Japanese. Exceptions to cases must, however, be treated cautiously, because they risk undermining the empirical content of the three constraints. One additional piece of evidence that the datives in the ACC-DAT frame do not correspond to verbal arguments is that nouns take dative complements as well (Hale (1973: 333), Nash (1980: 207ff.)) but seem never to take ergative or absolutive ones. 9 It is possible that Warlpiri has two ergative cases, one like (92a) and one like (92b). But, in the absence of compelling evidence, the agentive ergative seems superfluous. Furthermore, it is interesting to note that the ergative in many languages is homophonous with the instrumental. Much is made of this fact when semantic ergatives are proposed. This raises again the question of the synchronic versus diachronic status of generalizations such as these. In these cases, the question usually is not between two such explanations but whether we need a synchronic explanation on top of the diachronic one. This is because a case like the ergative and its homophonous "relatives" are clearly reconstructable to the same proto-case. The question is whether a synchronic relation between them adds to the explanation. In many cases it does not.

300

Notes to pages

107-117

10 Another possibility for the active languages is that the "ergative" in those languages is a type 1 nominative that is interpreted agentively. This would make it very analogous to a partitive. Partitives are like accusatives interpreted in a partitive manner. This can be handled by having a partitive case with the same LINK value as accusative but specified for a partitive interpretation. Either can apply (same LINK value so the same for the RC), but they will be interpreted differently. Similarly, ergative case in active languages could be regarded as a specially interpreted nominative. 11 Since finishing Smith (1992), the discussion of Kabardian in Colarusso (1989) has come to my attention. What is termed "ergative" here is called "oblique" in his discussion. The following introduction to his discussion and examples is very telling: "Kabardian is ergative. The function of the absolutive case is straightforward: the subject of intransitives and predicatives and the direct object of transitives. That of the oblique is more complex, covering all other grammatical roles . . . " (Colarusso (1989: 337)). 12 Colarusso (1989: 338) shows that another class of ditransitives, the causatives, also show the ERG-ERG-ABS pattern, i.e. the ergative is used for the agent of the causative (the causee). More generally, his discussion provides further examples for the wide range of uses of the ergative. 13 I do not mean to imply that all split systems will show a mixture of ergativeabsolutive and nominative-accusative systems. In fact, since the present approach provides for four kinds of two-case system, we might expect mixtures of two nominative-accusative patterns (type 1 and type 2) and mixtures of two types of ergative systems (types 3 and 4). An example of a split between two nominative-accusative systems that has already been proposed is Korean, which shows multiple accusative constructions (including adverbials) and multiple nominative constructions (also including adverbials). Which case can appear multiple times seems to depend on verb class, and therefore the analysis of Korean in the present theory might be as a mixed system. One verb class would behave as type 1, the other as type 2. This could be effected by the use of features and multiple cases, e.g. two defaults, one nominative and one accusative, at least one of which is specified for a particular verb class. The boundary of the two classes is still controversial, but it seems to be related to stativity (Maling (1989)) or agentivity (Kim (1990)). Alternatively, some occurrences of nominative or accusative may be syntactically distinct from the two syntactic cases (introduced by other special "rules"). I leave the proper analysis of Korean and the evaluation of whether and to what extent the most natural analysis fits in with the present proposals to further work. 14 Kiparsky (1987) also cites evidence from Mirror Principle effects to support the lexical insertion approach to linking. 15 The system in Georgian is considerably more complex than is reflected in the discussion in the text. Since the main point here is to demonstrate that an analysis of the common phenomenon of split ergativity based on tense is possible in this system, I will not pursue the further implications of the analysis and the refinements necessary to capture additional complexities such as the behavior of Series II verbs in Georgian (Harris 1981: 228ff.).

Notes to pages 118-127

301

16 As Gazdar et al. (1985: 28) note, an FCR, roughly speaking, constrains what categories can be subsumed by (i.e. more fully specify, extend) a given category. They thus partly define the notion of "fully specified syntactic category." Some FCRs are candidates for universal grammar, and some are clearly language-specific. See Gazdar et al. (1985: 27ff., 40) for further discussion of FCRs. 17 Note that FCRs, unlike defaults, are not overridden by pre-specification. Thus, the category in (115) in the text is ruled out by the FCR in (113) but would simply override the following default: (i)

18

19 20

21

22

23

[SUBCAT|REAL|ERG] —> [TENSE |AOR]

See Gazdar et al. (1985: 29ff., 40) for a discussion of feature specification defaults (FSDs) (here simply called "defaults"). I will not discuss the class of "active" case-marking languages (see Mithun (1991) and references there). In active case marking some intransitive subjects (S) are marked the same as the A and other Ss are marked like Os. Such languages are usually analyzed as having a grammaticalized category of agent, with varying definitions of agent depending on the language. If one wanted to adopt this approach in the present framework, one would probably propose that for such languages there is an argument category AGT which must receive some agent-like interpretation. I leave the question open whether a cross-linguistic minimal interpretation (necessary but not sufficient semantic condition on agenthood) along the lines of EXP (Chapter 2) and GOAL (see Chapter 5) can be found. Georgian may be a language with an ergative that is [ARG|-LA] and a nominative that is [ARG] but which do dissimilate. Thus languages like Georgian, if they are not active, as Harris argues, might be analyzed as having the X case and the Y case which will dissimilate under either the new or the old formulations of the LDC. In this way, the LDC need not itself be parameterized in any event. Goddard (1982: 173f.) also shows the advantages of his three-case analysis for the "adverbial modification" construction in Diyari. See also Austin (1981: 106ff.) for further details. Austin glosses the second word here as "3sgnFS-NEAR," but this should read "3sgnFO-NEAR," since nina is not the S but the O form of the pronoun on Austin's analysis. Following Goddard (1982), I label this "ACC." The reader may have noticed the role of syntax in this definition of morphological case. And of course syntactic case (a case linker) is defined partially in terms of morphology. There is no circularity here since the definition of morphological case does not mention linking or linkers at all. The worst that can be said of adopting this definition of morphological case in the present discussion is that we are "mixing levels." Since the advent of generative theory, however, this is no longer a serious obstacle. I leave open the questions of how to ensure a unique morphological analysis for a given language (at least criteria must be provided for the optimal analysis, e.g. that with

302

Notes to pages 127-193

the most cases conforming to Goddard's definition) and how such an analysis would be acquired by first language learners. 24 As he did for Diyari, Goddard (1982: 181) also shows that his three-case analysis simplifies the treatment of the "adverbial modification" construction in Yankunytjatjara. 4

Linker interactions

1 It is possible for there to be two adjunct cases with the same meaning, especially if one is newly grammaticalized. The prediction here is that alternation with the accusative (or other default) should form a pattern. 2 In this case the third-person feminine-common dual PGN suffix and the subordinative suffix coalesce to give rd. 3 Sadock (1991) notes that the low tone on the subordinative suffix is raised to mid for intonational reasons in the question. Also, the -e is the coalescence of the postconsonantal allomorph -i of the third-person masculine singular PGN suffix and the subordinative suffix. 4 Probably X must also contain material other than [-XA], i.e. it must refer to lexical information (argument category or individual argument). NPs with accusative in type 2 languages are probably never behavioral obliques. 5 Recall that adjuncts are always oblique. This might be handled by a default more specific than (71), like the following: 0)

rA .TTT J UxNT^i C T_] ATD

r[ ADJUNCT 1

[ +OBL

J

This default blocks the default in (71) by the Elsewhere Condition (or a generalized RC). 6 I have not presented a full analysis of Korean nominative and accusative, which is not wholly clear at present. See note 13 of the previous chapter for some suggestions and references. 5

Icelandic

1 It is, however, obligatory between a participle and any nominative NP. 2 Icelandic does have verb-initial (VI) sentences in so-called narrative style. 3 Zaenen, Maling, and Thrainsson (1985) conclude from these and related examples that an OBJ may antecede a reflexive but an OBJ2 may not, given that the examples above get a dual assignment of grammatical functions. As long as one has dual assignments of GFs or some analog of the symmetric object parameter it is not necessary to refer to OBJ2. All that is required is to be able to tell which object is higher, e.g. at argument structure (see §5.3 below). 4 I am assuming that the VP has the same argument structure as the V that heads it, but this is a non-essential assumption. 5 McCloskey (1985) argues that NPs in Irish have the option of not raising but receiving "default accusative." This default accusative is different from the one proposed here, since it is meant to be language-specific. If, on the present

Notes to pages 193-213

303

view, Irish, English, Icelandic, etc., are all type 1, and so all have default accusative, then the differences in Irish raising remain to be accounted for. The difference, however, can be attributed to two differences between Irish on the one hand and English and Icelandic on the other: (i) McCloskey argues that Irish shows clear evidence that the complements of raising verbs are Ss unlike in English and Icelandic, where there is no evidence that the complement is an S, and many have in fact proposed VPs, and (ii) in Irish, agreement seems to be the [ARG] linker, unlike in Icelandic where nominative seems to be the linker and agreement secondary. The second difference needs further motivation. 6 As expected, if these accusatives are not lexical, they are not preserved in the passive, unlike lexical case which is preserved. The following is one passive corresponding to a sentence like (61a), where a nominative subject corresponds to an active accusative object: (i)

Eg/*mig var leyndur sannleikanum. I(NOM)/me(Acc) was concealed truth-the(DAT) 'From me the truth was concealed.'

7 There is one verb which has a NOM-ACC-ACC frame, kosta 'cost,' but since this verb does not passivize, no conclusions about the status of the accusatives (default or lexical) can be drawn. 8 Recall that the feature value pair [PLACE [3]] for argument "place" is just an index to keep arguments syntactically distinct. It is not place in the sense that an argument that is [PLACE [2]] must always be second on the list. There is no reason here for [3] to become [2]. 9 One might be able to ensure that the theme was [ + o] and if it is the third argument by ordering the rules classifying the internal arguments. This too would involve a significant increase in theoretical power. 10 In Rognvaldsson (1990: 368) the coindexing is between Eg and the gap, but it is clear from the discussion that the coindexing should be between myndina and the gap. 11 Sridhar (1976) demonstrates in great detail that Kannada, like Icelandic, has dative subjects. In the ditransitive, Kannada exhibits asymmetric object behavior: only the goal can be the antecedent of tanna 'self s' but not the by-NP nor the "theme" NP: (i)

Bill-igej kelesavdavnuk John-inindai tanna*i/j/*k Bill-DAT servant John-by self s maneliye kottelpattitu. hoUSe-LOC give-PASS-PERF-3sgPAST

'Bill was given the servant by John in his (Bill's) house.'

304 Notes to pages 213-228 (ii)

Kelesavdavnuk Bill-igej John-inindai tanna*i/j/*k servant Bill-DAT John-by self s maneliye kottelpattitu. hoUSe-LOC give-PASS-PERF-3sgPAST 'Bill was given the servant by John in his (Bill's) house.'

In both orders, only the dative goal Bill-ige passes the subjecthood test. Thus a lexically case-marked NP in Kannada is a subject without being positionally linked. The dative NP does not have to appear sentence-initially to be the subject. However, the sentence in (ii) where the goal appears second might involve topicalization of the other NP. What makes Kannada interesting too is that it is not clear whether Kannada has a VP and whether positional linking, if there is any, makes reference to verbal projections. Whether there is a VP in languages like Kannada is important, because one possible constraint on positional linking would be that positional linking must always be relative to a verbal projection, e.g. pre-VP, VP-internal, adjacent to V, adjacent to V'. All the positional linkers proposed in the present theory have indeed involved verbal projections. If positional linking does not require verbal projections, it then remains open what constraints there are on positing positional linkers. 12 On the right in (107), the verb is specified as eligible for lexical accusative which is disallowed by the exception feature. The seeming contradiction could be avoided by allowing the accusative specification in the intransitive entry to be erased in deriving the transitive entry: (i)

blasa ( y ) 'blow' —• blasa ( x

y ) 'blow'

ACC

6

Changes in linking 1 Sentences like (2a) are attested without the genitive, and then the nominative, as predicted, appears. However, the genitive is not attested everywhere the present theory predicts it was possible. There is no reason to assume, however, that unattested necessarily means ungrammatical. 2 Interestingly, as in (3), Old English too shows some evidence of DS (see Mitchell (1985: §1027)). 3 Old Saxon may have shown DS, but the evidence is slim: the verb hreuwan 'rue,' whose Old English cognate, hreowan 'rue' is a DS verb (3), takes a dative in three clear instances and in a fourth where one ms (C) has accusative (see Sehrt (1925: 272) for sources and line numbers). 4 Very interestingly, DS has progressed at very different rates in different German dialects. For example, Hodler (1969) lists impersonal verbs that vary between accusative and dative in Bern German and for each notes whether one or the other case is not found in rural or urban areas. It emerges that DS is more advanced in the city of Bern than in the countryside. Bern German also shows DS with verbs not found in Standard German, such as

Notes to pages 228-245

305

lupfen 'be nauseated,' showing that it is not likely that DS is due to outside influence. 5 The other example that Halldorsson found from Old Norse is lysta 'want,' but, as he points out, this is an Old Norwegian example. However, this is interesting as evidence for Dative Sickness in Old Norwegian. Old Swedish on the other hand, shows widespread DS. Checking through Soderwall's (1884— 90) dictionary of Old Swedish reveals that most of the cognates of the Old Icelandic accusative experiencer class show dative variants in Old Swedish, at least twelve verbs in all. 6 See §6.6.1 below. 7 The second NP in (8a) is morphologically ambiguous (NOM/ACC/DAT), but in any case (8a) does not show DS. 8 For Old English, the question of the relative chronology of the beginning of DS and NS is not simple. NS seems to occur before DS for many verbs, but the problem is that Latin influence cannot be ruled out. 9 Interestingly, another ms. gives the accusative experiencer variant. Such variation in mss. does not occur in every instance, however. 10 Although there are no dative examples with these verbs, examples can be found with the verbs huggrjan 'hunger,' and paursjan 'thirst' as a participle, indicating possible NS. The use of participles in Gothic can, however, be taken to reflect the original Greek, rather than native Gothic syntax. 11 The scarcity of dative experiencers in Gothic suggests that perhaps Gothic had already undergone extensive NS before the fourth century, the period of the written sources. This would make Gothic another example of NS preceding DS. One problem is that any use of the nominative might be an imitation of Greek. On the other hand, if one assumes that there were a number of dative verbs that take nominative in imitation of the Greek, this leaves unexplained the inability of accusative to become nominative by imitation. 12 On the other hand, new cases can be introduced to a language by grammaticalization, e.g. from prepositions. Grammaticalization can ensure the presence of relatively specific cases. Overall, we have equilibrium between restrictiveness-decreasing changes and case loss on the one hand and replacement by grammaticalization on the other. 13 Another view of analogy is based on productivity, e.g. Hock (1991: 173): "it [a productive category] is more likely to be generalized by four-part analogy [grammatically annotated proportional analogy] than other, less productive categories." On this view, restrictiveness, rather than being an inverse correlate of simplicity as in (13), would be a way to make the notion of case productivity precise, leading again to (14). 14 Hirt (1934: 92ff.) discusses what he takes to be a general but not exceptionless tendency towards accusative in Indo-European languages. 15 For another approach to the Scandinavian data, see Faarlund (1990). That study is based on quite different assumptions about subjecthood than is this one. There the hypothesis is acquisition of the notion subject by the entire language, whereas on the present scheme there is change in the term versus oblique status of certain NPs and change in the morphosyntactic expression of

306

16 17

18 19

Notes to pages 245-286

NPs. I assume that Old Icelandic had subjects but try below to show the reasonableness of accepting that the earlier analogues of some - but not all - subject NPs in Modern Icelandic are oblique in Old Icelandic. We also need the idiosyncratic accusatives for the DS verbs, which I have already discussed and so leave out of the discussion here. There is some controversy over whether Old Norse had a passive rule (Dyvik (1980) against and Benediktsson (1980) for). What matters in the present context is whether the lexical cases compete with the nominative and, when they do, whether they win, as predicted on the present account. Lexical accusative NPs seem also to be oblique in Old Icelandic. Icelandic and Norwegian did not split off into languages until after 1200. As is usual, I am taking Old Norse-Old Icelandic to represent the situation in Common Scandinavian, which is the parent of all the Scandinavian languages.

7

Case semi-preservation 1 The example in (17c) is constructed. 2 Note too that having a separate dative specified ARG as in (i) would not do, because such a dative would alternate with the nominative in all nominative contexts: (i)

DAT: [ARG]

3 It seems that using the second, [ARG], value for lexical cases is much more acceptable in what I have termed LDC effects than in RC effects. That is, for some verbs DAT-ACC is quite favored, while the NOM-ACC variant (by NS) is not. As before, the present theory defines the space of possibilities. Variable rule type information will have to be added. The relative "strengths" of the two LINK values for scattered cases must be able to be different for the LDC and the RC. 4 The present account also predicts that scattered case of the type considered here should be optional in raising, because, under one LINK value, the scattered case does not refer to any "raised" (shared) material. This seems to be true (Barnes (1986: 23)). 5 The present account does not explain why some speakers do not accept this sentence, but it may be a form of hypercorrection. Even speakers who accept the nominative experiencer do not judge it as acceptable as the dative experiencer versions. The rise of the DAT-ACC frame is favored according to the Restrictiveness Change Prediction, and this would cause the further change of the rise of NOM-ACC frames. However, one possibility is that speakers adopt a prohibition on agreement with these verbs in order to minimize the difference with earlier patterns (DAT-NOM) which are still present. Such adjustments are common in language change. This adjustment exception is now being lost to a greater or lesser extent, depending on the speaker.

References

Allen, Cynthia. 1984. "Another Look at the History of 'Impersonal' Verbs in English." MS, The Australian National University, Canberra. 1986. "Reconsidering the History of like" Journal of Linguistics 22: 375^09. Alsina, Alex. 1991. "Causatives." MS, Stanford University. Anderson, John. 1986. "A Note on Old English Impersonals." Journal of Linguistics 22: 166-77. Anderson, Stephen R. 1969. "West Scandinavian Vowel Systems and the Ordering of Rules." Ph.D. Dissertation, MIT, Cambridge, Mass. 1977. "On Mechanisms by Which Languages Become Ergative." Charles N. Li (ed.), Mechanisms of Syntactic Change, 317-64. Austin: University of Texas Press. 1986. "Disjunctive Ordering in Inflectional Morphology." Natural Language and Linguistic Theory 4: 1-31. Andrews, Avery. 1976. "The VP Complement Analysis in Modern Icelandic." Papers from the Sixth Annual Meeting of the North Eastern Linguistic Society, Recherches Linguistiques a Montreal 6: 1—21. 1982. "The Representation of Case in Modern Icelandic." Joan Bresnan (ed.), The Mental Representation of Grammatical Relations, 427-503. Cambridge, Mass.: MIT Press. 1990. "Case Structures and Control in Modern Icelandic." Joan Maling and Annie Zaenen (eds.), Modern Icelandic Syntax ( = Syntax and Semantics 24), 187-234. San Diego: Academic Press. Austin, Peter. 1981. A Grammar ofDiyari, South Australia. Cambridge University Press. 1982. "Transitivity and Cognate Objects in Australian Languages." Paul Hopper and Sandra Thompson (eds.), Studies in Transitivity ( = Syntax and Semantics 24), 37-47. New York: Academic Press. Baker, Mark. 1988. Incorporation: A Theory of Grammatical Function Changing. University of Chicago Press. Barnes, Michael. 1986. "Subject, Nominative, and Oblique Case in Faroese." Scripta Islandica 37: 13—46. Behaghel, Otto. 1924. Deutsche Syntax: Eine geschichtliche Darstellung, Vol. 1.: Die Wortklassen und Wortformen. Heidelberg: Carl Winter. Benediktsson, Hreinn. 1980. "The Old Norse Passive: Some Observations." Even Hovdhaugen (ed.), The Nordic Languages and Modern Linguistics 4, 108-19. 307

308

References

Bernodusson, Helgi. 1982. "Opersonulegar setningar i islensku." [Impersonal Sentences in Icelandic] Master's thesis, University of Iceland, Reykjavik. Bolinger, Dwight. 1971. The Phrasal Verb in English. Cambridge, Mass.: Harvard University Press. Bresnan, Joan. 1982. "Control and Complementation." Joan Bresnan (ed.), The Mental Representation of Grammatical Relations, 282-390. Cambridge, Mass.: MIT Press, ed. 1982. The Mental Representation of Grammatical Relations. Cambridge, Mass.: MIT Press. Bresnan, Joan, and Jonni M. Kanerva. 1989. "Locative Inversion in Chichewa: A Case of Factorization in Grammar." Linguistic Inquiry 20: 1^0. Bresnan, Joan, and Lioba Moshi. 1990. "Object Asymmetries in Comparative Bantu Syntax." Linguistic Inquiry 21: 147-85. Carpenter, Bob. 1992. The Logic of Typed Feature Structures. Cambridge University Press. Carter, Richard. 1976. "Some Linking Regularities." Beth Levin and Carol Tenny (eds.), On Linking: Papers by Richard Carter [1988], 1-92. Cambridge, Mass.: Center for Cognitive Science Lexicon Project. Catford, J. C. 1976. "Ergativity in Caucasian Languages." Proceedings of the Sixth Annual Meeting of the North Eastern Linguistic Society, Recherches Linguistiques a Montreal 6: 37—48. Chierchia, Gennaro. 1984. "Topics in the Syntax and Semantics of Infinitives and Gerunds." Ph.D. Dissertation, University of Massachusetts, Amherst. 1989. "Structured Meanings, Thematic Roles and Control." Gennaro Chierchia, Barbara H. Partee, and Raymond Turner (eds.), Properties, Types, and Meaning II, 131-66. Dordrecht: Kluwer. Chomsky, Noam. 1981. Lectures on Government and Binding. Dordrecht: Foris. Cleasby, Richard and Gudbrand Vigfusson. 1957. An Icelandic-English Dictionary. Oxford: Clarendon. Clements, George N. 1975. "The Logophoric Pronoun in Ewe: Its Role in Discourse." Journal of West African Linguistics 10.2: 141-77. Colarusso, John. 1989. "East Circassian (Kabardian dialect)." B. George Hewitt (ed.), The Indigenous Languages of the Caucasus. Vol. 2: The North West Caucasian Languages, 261-355. Cole, Peter, Wayne Harbert, Gabriella Hermon, S. N. Sridhar. 1980. "The Acquisition of Subjecthood." Language 56: 719-43. Collberg, Sheila Dooley. 1986. "Oblique Subjects in Icelandic Passive Constructions." Studia Linguistica 40: 67-95. Comrie, Bernard. 1976. "The Syntax of Causative Constructions: CrossLanguage Similarities and Differences." Shibatani, Masayoshi (ed.), The Grammar of Causative Constructions ( = Syntax and Semantics 6), 261-312. New York: Academic Press. 1978. "Ergativity." Winfred P. Lehmann (ed.), Syntactic Typology, 329-94. Austin: University of Texas Press. 1981. The Languages of the Soviet Union. Cambridge University Press.

References

309

Cowper, Elizabeth A. 1988. "What is a Subject? Non-Nominative Subjects in Icelandic." James Blevins and Juli Carter (eds.), Proceedings of the North Eastern Linguistic Society 18: 94-108. Croft, William. 1991. Syntactic Categories and Grammatical Relations: The Cognitive Organization of Information. University of Chicago Press. Curme, George O. 1922. A Grammar of the German Language. New York: Macmillan. Dal, Ingerid. 1966. Kurze deutsche Syntax auf historischer Grundlage, 3rd edn. Tubingen: Max Niemeyer. Davies, William D. 1984. "Antipassive: Choctaw Evidence for a Universal Characterization." David M. Perlmutter and Carol G. Rosen (eds.), Studies in Relational Grammar 2, 331-76. University of Chicago Press. Delbriick, Berthold. 1893. Vergleichende Syntax der indogermanischen Sprachen, Vol. 3. Strassburg: Triibner. Dench, Alan and Nicholas Evans. 1988. "Multiple Case-Marking in Australian Languages." Australian Journal of Linguistics 8: 1^47. Denison, David. 1990a. "Auxiliary + Impersonal in Old English." Folia Linguistica Historica 9: 139-66. 1990b. "The Old English Impersonals Revived." Sylvia Adamson, Vivien A. Law, Nigel Vincent, and Susan Wright (eds.), Papers from the 5th International Conference on English Historical Linguistics, 111-39. Amsterdam: Benjamins. Dixon, R. M. W. 1979. "Ergativity." Language 55: 59-138. Dowty, David. 1985. "On Recent Analyses of the Semantics of Control." Linguistics and Philosophy 8: 1—4-1. 1989. "On the Semantic Content of the Notion Thematic Role."' Gennaro Chierchia, Barbara H. Partee, and Raymond Turner (eds.), Properties, Types, and Meaning II, 69-129. Dordrecht: Kluwer. 1991. "Thematic Proto-Roles and Argument Selection." Language 67: 547619. Drosdowski, Giinter, et al. 1976. Duden. Mannheim: Duden. Dryer, Matthew S. 1982. "Passive and Inversion in Kannada." Berkeley Linguistics Society 8: 311-21. Dyvik, Helge. 1980. "Har Gammelnorsk Passiv?" [Does Old Norse Have a Passive?] Even Hovdhaugen (ed.), The Nordic Languages and Modern Linguistics 4: 81-107. Einarsson, Stefan. 1949. Icelandic. Baltimore: The Johns Hopkins Press. Elmer, Willy. 1981. Diachronic Grammar: The History of Old and Middle English Subjectless Constructions. Tubingen: Max Niemeyer. Faarlund, Jan Terje. 1990. Syntactic Change: Toward a Theory of Historical Syntax. Berlin: Mouton de Gruyter. Fillmore, Charles. 1968. "The Case for Case." Emmon Bach and Robert T. Harms (eds.), Universals in Linguistic Theory, 1-90. New York: Holt, Rinehart, and Winston. Fischer, Olga C. M. and Frederike C. van der Leek. 1983. "The Demise of the Old English Impersonal Construction." Journal of Linguistics 19: 337-68.

310

References

Fischer, Olga C. M. and Frederike C. van der Leek. 1987. "A 'case' for the Old English Impersonal." Willem Koopman et al. (eds.), Explanation and Linguistic Change, 79-120. Amsterdam: John Benjamins. Gaaf, W. van der. 1904. The Transition from the Impersonal to the Personal Construction in Middle English (= Anglistische Forschungen 14). Heidelberg: Carl Winter. Gaedicke, Carl. 1880. Der Accusativ im Veda. Breslau: Koebner. Gazdar, Gerald, Ewan Klein, Geoffrey Pullum, and Ivan Sag. 1985. Generalized Phrase Structure Grammar. Cambridge, Mass: Harvard University Press. Gerdts, Donna, and Cheong Youn. 1988. "Korean Psych Constructions: Advancement or Retreat?" Papers from the Regional Meeting Chicago Linguistic Society 24: 155-75. Gildersleeve, Basil. 1894. Gildersleeve's Latin Grammar. Boston: Heath. Goddard, Cliff. 1982. "Case Systems and Case Marking in Australian Languages: A New Interpretation." Australian Journal of Linguistics 2: 167-96. Grimshaw, Jane. 1990. Argument Structure. Cambridge, Mass.: MIT Press. Gruber, J. 1965. "Studies in Lexical Relations." Ph.D. Dissertation, MIT, Cambridge, Mass. Hagman, Roy S. 1977. Nama Hottentot Grammar. Bloomington, Ind.: Research Center for Language and Semiotic Studies, Indiana University. Hale, Ken. 1973. "Person Marking in Walbiri." Stephen R. Anderson and Paul Kiparsky (eds.), A Festschrift for Morris Halle, 308^4. New York: Holt, Rinehart & Winston. 1982. "The Essential Features of Warlpiri Main Clauses." S. Swartz (ed.), Papers in Warlpiri Grammar: In Memory of Lothar Jagst. Work-Papers of SIL-AAB, Series A, Vol. 6. Berrimah, N.T.: Summer Institute of Linguistics. 1983. "Warlpiri and the Grammar of Non-Configurational Languages." Natural Language and Linguistic Theory 1: 5—47. Halldorsson, Halldor. 1982. "Urn meranir: Drog ad samtimalegri og sogulegri athugun." [On Using mer: An Attempt at a Synchronic and Diachronic Study] Islenskt Mdl og Almenn Mdlfrced/4: 159-89. Harada, S. I. 1973. "Counter-equi NP Deletion." Annual Bulletin 1. Tokyo: Research Institute of Logopedics and Phoniatrics, University of Tokyo. Harbert, W. and A. J. Toribio. Forthcoming. "Nominative Objects." Cornell Working Papers in Linguistics 9. Harris, Alice C. 1981. Georgian Syntax: A Study in Relational Grammar. Cambridge University Press. 1984. "Case Marking, Verb Agreement, and Inversion in Udi." David M. Perlmutter and Carol G. Rosen (eds.), Studies in Relational Grammar 2, 243-58. University of Chicago Press. Harris, Martin. 1978. "The Inter-Relationship Between Phonological and Grammatical Change." Jacek Fisiak (ed.), Historical Phonology, 159-72. The Hague: Mouton. Henniss, Kathryn. 1989. "'Covert' Subjects and Determinate Case: Evidence from Malayalam." E. Jane Fee and Katherine Hunt (eds.), Proceedings of

References

311

the West Coast Conference on Formal Linguistics 8: 167-75. Stanford, Calif: Center for the Study of Language and Information. Hirt, Hermann. 1934. Indogermansiche Grammatik. Part IV: Syntax I: Syntaktische Verwendung der Kasus und der Verbalformen. Heidelberg: Winter. Hock, Hans H. 1991. Principles of Historical Linguistics, 2nd edn. Berlin: Mouton de Gruyter. Hodler, Werner. 1969. Berndeutsche Syntax. Bern: Francke. Hoekstra, Teun. 1984. Transitivity: Grammatical Relations in Government-Binding Theory. Dordrecht: Foris. Holmberg, Anders. 1986. "Word Order and Syntactic Features in the Scandinavian Languages and English." Doctoral Dissertation, University of Stockholm. Jackendoff, Ray. 1972. Semantic Interpretation in Generative Grammar. Cambridge, Mass.: MIT Press. 1975. "Morphological and Semantic Regularities in the Lexicon." Language 51: 639-71. 1987. "The Status of Thematic Relations in Linguistic Theory." Linguistic Inquiry 18: 369-411. 1990. Semantic Structures. Cambridge, Mass.: MIT Press. 1993. "On the Role of Conceptual Structure in Argument Selection: A Reply to Emonds." Natural Language and Linguistic Theory 11: 279-312. Jakobson, Roman. 1936. "Beitrag zur allgemeinen Kasuslehre: Gesamtbedeutung der russischen Kasus." Travaux du Cercle Linguistique de Prague 6: 240-88. Jelinek, Eloise. 1984. "Empty Categories, Case, and Configurationality." Natural Language and Linguistic Theory 2: 39-76. Jespersen, Otto. 1927. A Modern English Grammar on Historical Principles, Vol. 3. London: Allen and Unwin. Joshi, Smita. 1987. "The Notion 'Logical Subject' in the Grammar of Marathi." MS, Stanford University. Keenan, Edward L. 1976. "Towards a Universal Definition of'Subject'." Charles H. Lee (ed.), Subject and Topic, 303-33. New York: Academic Press. Kemenade, Ans van. 1987. Syntactic Case and Morphological Case in the History of English. Dordrecht: Foris. Kim, Young-Joo. 1990. "The Syntax and Semantics of Korean Case: The Interaction Between Lexical and Syntactic Levels." Ph.D. Dissertation, Harvard University, Cambridge, Mass. Kiparsky, Paul. 1973. "'Elsewhere' in Phonology." Stephen R. Anderson and Paul Kiparsky (eds.), A Festschrift for Morris Halle, 93-106. New York: Holt, Rinehart, and Winston. 1982. Explanation in Phonology. Dordrecht: Foris. 1987. "Morphology and Grammatical Relations." MS, Stanford University. 1989. "Agreement and Linking Theory." MS, Stanford University. Klokeid, Terry J. 1978. "Nominal Inflection in PamaNyungan: A Case Study in Relational Grammar." Werner Abraham (ed.), Valence, Semantic Case, and Grammatical Relations, 577-615. Amsterdam: Benjamins.

312

References

Koutsoudas, Andreas, Gerald Sanders, and Craig Noll. 1974. "The Application of Phonological Rules." Language 50: 1-28. Kress, Bruno. 1982. Isldndische Grammatik. Leipzig: VEB Verlag Enzyklopadie. Kroeger, Paul. 1993. Phrase Structure and Grammatical Relations in Tagalog. Stanford, Calif.: CSLI. Kiihner, Raphael. 1898. Ausfiihrliche Grammatik der griechischen Sprache. Rev. by Bernard Gerth. Part 2. Vol 1. Hannover: Hahn. 1912-14. Ausfiihrliche Grammatik der lateinischen Sprache. Hannover: Hahn. Kuipers, Aert H. 1955. "The North-West Caucasian Languages." Analecta Slavica: 193-206. Amsterdam: De Bezige Bij. 1960. Phoneme and Morpheme in Kabardian. The Hague: Mouton. 1962. "The Circassian Nominal Paradigm: A Contribution to Case-Theory." Lingua 11: 231-48. Kulonen, Ulla-Maija. 1989. The Passive in Ob-Ugrian. Helsinki: SuomalaisUgrilainen Seura. Kuno, Susumo. 1973. The Structure of the Japanese Language. Cambridge, Mass.: MIT Press. Kurylowicz, Jerzy. 1964. The Inflectional Categories of Indo-European. Heidelberg: Carl Winter. 1965. "The Evolution of Grammatical Categories." Diogenes 51: 55-71. Ladusaw, William and David Dowty. 1987. "Towards a Non-Grammatical Account of Thematic Roles." W. Wilkins (ed.), On the Nature of Thematic Roles. New York: Academic Press. Lavotha, Odon. 1960. Das Passiv in der wogulischen Sprache. Journal de la Societe Finno-Ougrienne 62 A. Levin, Beth. 1983. "On the Nature of Ergativity." Ph.D. Dissertation, MIT, Cambridge, Mass. Lehmann, Christian. 1985. "Grammaticalization: Synchronic Variation and Diachronic Change." Lingua e Stile 20: 303-18. Lightfoot, David. 1979. Principles of Diachronic Syntax. Cambridge University Press. 1981. "A History of NP Movement." C. L. Baker and J. McCarthy (eds.), The Logical Problem of Language Acquisition, 86-119. Cambridge, Mass.: MIT Press. 1988. "Syntactic Change." Frederick J. Newmeyer (ed.), Linguistics: The Cambridge Survey, Vol. 1: Linguistic Theory: Foundations, 303-23. Cambridge University Press. 1991. How to Set Parameters: Arguments from Language Change. Cambridge, Mass.: MIT Press. Lockwood, W. B. 1955. An Introduction to Modern Faroese. Copenhagen: Munksgaard. McCloskey, James. 1985. "Case, Movement and Raising in Modern Irish." Jeffrey Goldberg, Susannah MacKaye, and Michael T. Wescoat (eds.), Proceedings of the West Coast Conference on Formal Linguistics 4: 190— 205. Stanford, Calif.: Stanford Linguistics Association.

References

313

Maling, Joan. 1984. "Non-Clause-Bounded Reflexives in Icelandic." Linguistics and Philosophy 7: 211—41. 1989. "Adverbials and Structural Case in Korean." Harvard Studies in Korean Linguistics 3: 297-308. 1990. "Clause-Bounded Reflexives in Modern Icelandic." Joan Maling and Annie Zaenen (eds.) Modern Icelandic Syntax ( = Syntax and Semantics 24): 277-87. San Diego: Academic Press. Marantz, Alec. 1984. On the Nature of Grammatical Relations. Cambridge, Mass.: MIT Press. Meillet, Antoine. 1912. Linguistique historique et linguistique generate, Vol. 1. Paris: Klincksieck. Mitchell, Bruce. 1979. "F. Th. Visser, An Historical Syntax of the English Language: Some Caveats Concerning Old English." English Studies 60: 537^2. 1985. Old English Syntax, Vol. 1. Oxford: Clarendon Press. Mithun, Marianne. 1991. "Active/Agentive Case Marking and its Motivations." Language 67: 510—46 Miyagawa, Shigeru. 1989. Structure and Case Marking in Japanese ( = Syntax and Semantics 22). San Diego: Academic Press. Mohanan, Tara. 1994. Argument Structure in Hindi. Stanford, Calif.: CSLI. Nash, David. 1980. 'Topics in Warlpiri Grammar." Ph.D. Dissertation, MIT, Cambridge, Mass. Noreen, Adolf. 1903-7. Vdrt sprdk. [Our Language] Lund: Gleerup. Nygaard, Marius. 1905. Norron Syntax. [Old Norse Syntax] Kristiania: H. Aschehoug. Ostler, Nicholas. 1979. "Case-Linking: A Theory of Case and Verb Diathesis Applied to Classical Sanskrit." Ph.D. Dissertation, MIT, Cambridge, Mass. Paul, Hermann, Hugo Moser, Ingeborg Schrobler, and Siegfried Grosse. 1975. Mittelhochdeutsche Grammatik. 22nd edn. Tubingen: Max Niemeyer. Perlmutter, David. 1982. "Syntactic Representation, Syntactic Levels and the Notion of Subject." Pauline Jacobson and Geoffrey K. Pullum (eds.), The Nature of Syntactic Representation, 283-340. Dordrecht: Reidel. ed. 1983. Studies in Relational Grammar. University of Chicago Press. Pollard, Carl and Ivan A. Sag. 1987. Information-Based Syntax and Semantics, Vol. 1: Fundamentals. Stanford, Calif.: CSLI. Poser, William. 1991. "What is the 'Double -o Constraint' a Constraint on?" MS, Stanford University. Postal, Paul. 1977. "Antipassive in French." Proceedings of the Seventh Annual Meeting of the North East Linguistic Society. Cambridge, Mass.: MIT Press. Reprinted in Lingvisticae Investigationes 1: 333-74. 1982. "Some Arc Pair Grammar Descriptions." Pauline Jacobson and Geoffrey K. Pullum (eds.), The Nature of Syntactic Representation, 341^25. Dordrecht: Reidel. 1986. Studies of Passive Clauses. Albany: SUNY Press. Reinhammer, Maj. 1973. Om Dativ i Svenska Och Norska Dialekter, Vol 1: Dativ vid Verb. Uppsala: A. B. Lundqvistska Bokhandeln.

314

References

Reis, Marga. 1982. "Zum Subjektbegriff im Deutschen." Werner Abraham (ed.), Satzglieder im Deutschen: 171-211. Tubingen: Giinter Narr. Riemsdijk, Henk van and Edwin Williams. 1981. "NP-Stmcture." Linguistic Review 1: 171-217. Robinson, Orrin W. 1976. "A 'Scattered' Rule in Swiss German." Language 52: 148-62. Rogava, G. V. and Z. I. Keraseva. 1966. Grammatika adygejskogo jazyka. [Grammar of the Adyghe Language] Krasnodar/Majkop: Krasnodarskoe kniznoe Izd. Rognvaldsson, Eirikur. 1990. "Null Objects in Icelandic." In Joan Maling and Annie Zaenen (eds.), Modern Icelandic Syntax ( = Syntax and Semantics 24), 367-79. San Diego: Academic Press. 1991. "Quirky Subjects in Old Icelandic." In Halldor Arman Sigurdsson (ed.), Papers from the Twelfth Scandinavian Conference of Linguistics, 369-78. Reykjavik: Linguistic Institute, University of Iceland. Rognvaldsson, Eirikur and Hoskuldur Thrainsson. 1990. "On Icelandic Word Order Once More." Joan Maling and Annie Zaenen (eds.), Modern Icelandic Syntax ( = Syntax and Semantics 24), 3-40. San Diego: Academic Press. Sadock, Jerrold M. 1991. Autolexical Syntax: A Theory of Parallel Grammatical Representations. University of Chicago Press. Saito, Mamoru. 1983. "Case and Government in Japanese." Proceedings of the West Coast Conference on Formal Linguistics 2: 247-59. Sapir, Edward. 1921. Language. New York: Harcourt, Brace & World. Schwyzer, Eduard. 1966. Griechische Grammatik, ed. by Albert Debrunner, Vol. 2: Syntax und Syntaktische Stilistik. Munich: Beck. Seefrantz-Montag, Ariane von. 1984. "'Subjectless' Constructions and Syntactic Change." Jacek Fisiak (ed.), Historical Syntax, 521-53. Berlin: Mouton. Sehrt, Edward Henry. 1925. Vollstdndiges Worterbuch zum Heliand und zur altsdchsischen Genesis ( = Hesperia 14). Baltimore: Johns Hopkins Press. Shevelov, George Y. 1963. The Syntax of Modern Literary Ukrainian. The Hague: Mouton. Sigurdsson, Halldor Arman. 1988. "From OV to VO: Evidence from Old Icelandic." Working Papers in Scandinavian Syntax 34. 1989. "Verbal Syntax and Case in Icelandic." Doctoral Dissertation, University of Lund. Silverstein, Michael. 1976. "Hierarchy of Features and Ergativity." R. M. W. Dixon (ed.), Grammatical Categories in Australian Languages, 112-71. Canberra: Australian Institute of Aboriginal Studies. Simpson, Jane. 1991. Warlpiri Morpho-Syntax: A Lexicalist Approach. Dordrecht: Kluwer. Smith, Henry: 1992. "Restrictiveness in Case Theory." Ph.D. Dissertation, Stanford University. 1993. "Linking Changes in Icelandic." Henk Aertsen and Robert J. Jeffers (ed.), Historical Linguistics 1989. Amsterdam: John Benjamins. 1994. "'Dative Sickness' in Icelandic." Natural Language and Linguistic Theory 12: 675-736.

References

315

forthcoming. "Discourse Representation in Icelandic." Proceedings of the Third Symposium on Germanic Linguistics. Amsterdam: John Benjamins. Smyth, Herbert Weir. 1956. Greek Grammar, rev. by Gordon Messing. Cambridge, Mass.: Harvard University Press. Sobin, Nicholas. 1985. "Case Assignment in Ukrainian Morphological Passive Constructions." Linguistic Inquiry 16: 649-62. Soderwall, K. F. 1884-90. Ordbok ofver Svenska Medeltidssprdket. [Dictionary of Medieval Swedish] Lund: Berling. Sridhar, S. N. 1976. "Dative Subjects, Rule Government, and Relational Grammar." Studies in the Linguistic Sciences 6.1: 130-51. Steele, Susan et al. 1981. An Encyclopedia of AUX: A Study in Cross-Linguistic Equivalence. Cambridge, Mass.: MIT Press. Svavarsdottir, Asta. 1982. "t>agufallssyki." [Dative Sickness] Islenskt Mdl og Almenn Mdlfmdi 4: 19-62. Thrainsson, Hoskuldur. 1976. "Reflexives and Subjunctives in Icelandic." Papers of the Sixth Annual Meeting of the North Eastern Linguistic Society, Recherches Linguistiques a Montreal 6: 225^0. 1979. On Complementation in Icelandic. New York: Garland. 1986. "On Auxiliaries, AUX and VPs in Icelandic." L. Hellan and K. Koch Christensen (eds.), Topics in Scandinavian Syntax, 235-65. Dordrecht: Reidel. Traugott, Elizabeth Closs. 1992. "Old English Syntax." Richard Hogg (ed.), The Cambridge History of the English Language, Vol. 1: Old English, 168-289. Cambridge University Press. Traugott, Elizabeth Closs and Bernd Heine. 1991. "Introduction." Elizabeth Closs Traugott and Bernd Heine (eds.), Approaches to Grammaticalization, 1-14. Amsterdam: Benjamins. Uszkoreit, Hans. 1987. "Linear Precedence in Discontinuous Constituents: Complex Fronting in German." Geoffrey J. Huck and Almerindo E. Ojeda (eds.), Discontinuous Constituency ( = Syntax and Semantics 20), 405-25. Orlando: Academic Press. Van Valin, Robert D., Jr. 1991. "Another Look at Icelandic Case Marking and Grammatical Relations." Natural Language and Linguistic Theory 9: 145-94. 1993. Review of Syntactic Categories and Grammatical Relations, William Croft. Language 69: 155-62. Verhagen, Arie. 1986. Linguistic Theory and the Function of Word Order in Dutch. Dordrecht: Foris. Visser, F. Th. 1963. An Historical Syntax of the English Language, Part One. Leiden: E. J. Brill. Wasow, Thomas. 1977. "Transformations and the Lexicon." Peter Culicover, Thomas Wasow, and Adrian Akmajian (eds.), Formal Syntax, 327-60. New York: Academic Press. 1981. "Comments on the Paper by Baker." C. L. Baker and J. McCarthy (eds.), The Logical Problem of Language Acquisition, 324-9. Cambridge, Mass.: MIT Press.

316

References

Wechsler, Stephen. 1995. The Semantic Basis of Argument Structure. Stanford, Calif.: CSLI. Whitney, William Dwight. 1892. "On Delbriick's Vedic Syntax." American Journal of Philology 13.3, no. 51: 271-306. Woolford, Ellen. 1993. "Symmetric and Asymmetric Passives." Natural Language and Linguistic Theory 11: 679-728. Yip, Moira, Joan Maling, and Ray Jackendoff. 1987. "Case in Tiers." Language 63: 217-50. Zaenen, Annie and Joan Maling. 1984. "Unaccusative, Passive and Quirky Case." West Coast Conference on Formal Linguistics 3: 317-29. Zaenen, Annie, Joan Maling, and Hoskuldur Thrainsson. 1985. "Case and Grammatical Functions: The Icelandic Passive." Natural Language and Linguistic Theory 3: 441-83.

Index

absolutive 70, 77, 79, 110, 150 agreement 110 case label 102, 109 cognate objects 107 default, as 103, 149f. from nominative 243 type 3 101, 131 type 4 108ff. abstractness of linkers 16f. AC, see Applicability Constraint accusative 25, 29, 39ff., 40f., 44, 58f., 71, 88, 90f., 99, 132, 153, 216ff., 275 adverbials, see adverbial accusatives default, as 39, 40, 125, 132f. lexical 24, 27, 51f., 57ff., 213ff., 216ff., 234, 244, 303f. multiple accusatives 88, 125, 300, see also Default Linker Non-Dissimilation Prediction non-experiencer subjects 59, 295 subject psych verbs 52, 62 type 1 90, 131 type 2 70ff., 9Iff. Acting 1, 2, etc. (RG) 15, 144, 145, 146 active case-marking 107, 301 adjacency-to-V, see verb-adjacent position ADJUNCT 68

adjunct(s) 68, 84, 103f., 112, 131, 142, 302, see also non-arguments accusative 133, 153 cases 113, 131, 302 obligatory and optional 66 obliqueness default 302 promotion to argument 266, 296, 298 adpositions 167 adverbial(s) 41, 84f., 91, 108, 126, 134, 157, 296, 300ff. accusatives 40f., 90f., 99, 114, 132, 145, 153, 275 Adyghe, see Kabardian affectedness 298 agent 6, 39, 62, 105, 107, 301 of causation 105

agentive ergative 299 agentive subjects 15 agentivity 300 agreement 2f., 19, 104, 110, 135, 154f., 163, 167f., 173ff., 182, 192, 266, 294, 296, 303 anaphoric status of (KLT) 155, 294 default 174f. markers/pronouns 104 nominative and 296 subject versus object 294 Allen, C. 226, 233, 251, 286, 307 Alsina, A. 299, 307 alternate realization hypothesis 96 analogy 27, 226, 232, 234f., 254, 257f. exceptions to 23 5f., 243 productivity and 305 reverse 235 rule-based 258 anaphors, see reflective pronouns, reflexivization Anderson, J. 226, 256, 307 Anderson, S. 44, 65, 242, 270, 307 Andrews, A. 23, 25, 57, 192, 260, 295, 307 Anti-DS 134 antipassive/antipassivization 130, 146f., 149ff. any-argument case/linker 62, 74ff., 83f., 87, 101, 108, 114 Applicability Constraint 42, 48, 173, 290 case semi-preservation and 276ff. as constraint on lexical insertion 116 application "in principle" 297 apposition 296f., 299 ARG(ument) feature 31, 38, 42, 75f., 82, 84, 119, 135, 190, 193 argument(s) 5, 31, 39ff., 62, 293 "place" 303 adjunct distinction 64ff., 67 cases from predicators 256 categories 17, 32, 34f., 37ff., 61, 63, 67, 69, 173, 190, 257f., 301 category 32, 50, 59, 82, 253

317

318

Index

covert 184, 188 obligatory and optional 66 ordering 18f., 195 permutation 196, 285 promotion to 296 structure 3, 16, 18, 31, 34, 60f., 63, 116, 151, 188, 289, 299 syntactic 193 types 17, 35 versus adjunct linkers 255 Argument Linker Dissimilation Prediction 83, 87,92, 102, 109, 111 Corollary 83 Arnadottir, T. 294 asterisk notation 47 atomic feature structure 31 Attribute-Value Matrix (AVM) 31, 41 Austin, P. 121, 123, 125f., 301, 307 Australian languages 107 AUX(iliary) 18, 37, 63, 257f. availability (KLT) 294 Baker, M. 293, 299, 307 Barnes, M. 27, 275, 283ff., 306f. Behaghel, O. 228, 307 behavioral object-, subjecthood, see objecthood, subjecthood, termhood Benediktsson, H. 261, 306f. benefactive 112 Bern German 304 Bernodusson, H. 207, 295, 308 bidirectionality of traditional thematic roles 37 binding 18, see also reflexive pronouns, reflexivization Biuniqueness Condition (GB) 11 blocking 62, 68, 73 Bolinger, D. 308 Bresnan, J. 9, 48, 186, 194, 308 Burzio's Generalization 147ff. reverse, type 3 151 reverse, type 4 152 by-phrase 10f., 147, 150, 243, 293 and new ergative marker 243 caret sign (demotion) 8, 10 Carpenter, B. 295, 308 Carter, R. 308 Case/case 3, 19, 40, see also active casemarking, split NP ergativity, threecase systems absorption (GB) 153 alternations 22ff., 58f., 61, 134, 216, 234ff., 290 anti-preservation 270ff., 278, 281 argument-category 50, 59, 82, 180, 253

attraction 264 Case and (GB) 271, 294 change 3, 22Iff., see also changes in linking clash 166 configuration, see word order and case control and 188 cross-linguistic definition of 8If. default, see default case dissimilation 83f., 275, 282ff. distribution 1, 44, 277, see also Applicability Constraint forms versus case systems 122 frames 31, 42 idiosyncratic, see lexical case; see also idiosyncratic case labels 82f., 120 lexical, see idiosyncratic case, lexical case' lexical versus grammatical 180 morphological If., 122, 127, 282 preservation 50, 180ff., 185, 188, 193, 200, 216ff., 220, 246, 259ff., 265, 268ff., 303 repeatability of 84, see also Linker Dissimilation Constraint, Linker NonDissimilation Prediction rules 2, 44, 234 scattered 274ff., 280, 282ff., 306 semi-preservation 259ff., 265ff., 270ff., 274f., 278ff., 287f. stacking 169, 297 syntactic If., 8, 20, 50, 58, 104, see also argument category uniform treatment of 69 word order and, see word order and case Case Theory (GB) 7, lOff., 113, 147, 168, 271, 294, 298 Case Tier Hypothesis 76ff., 184 Case Grammar 5 Case Marker Drop (Japanese) 94, 96 CAT feature 32, 34, 190 Categorial Grammar 19, 293 Catford, J. 108, 151f., 308 Caucasian languages 108, 294, see also Georgian, Kabardian causative(s) 79, 92f., 96, 299f. change(s) in linking 221ff., 225ff., 231, 233, 241ff., 251, 256, 290 analogy and 257 causation 225 cluster view 223f., 226 directionality 237 equilibrium 225, 305 grammaticalization and 257

Index lexicalism and 227 locational approaches 256 tendency view 223ff. unidirectionality of 254 unit of 225 Chierchia, G. 35, 308 chomage (RG) 144 Chomsky, N. 7, 40, 48, 146, 271, 308 Circassian 108, see also Kabardian clause structure, basic 48 Cleasby, R. 228, 308 Clements, G. 250, 308 clitic pronouns 141 ff., 294 cluster view of linking change 223ff. coding subjecthood properties, changes in 236ff., see also subjecthood coding termhood properties, changes in 236, 242, see also termhood cognate absolutives (Warlpiri) 102 cognate accusative 41, 90, 145 cognate objects 102f., 107, 125 coindexation 186f., 189 Colarusso, J. 300, 308 Cole, P. 25, 237f., 241, 243, 248, 308 Collberg, S. 233, 308 comitative 112 commercial transaction verbs 36 Common Scandinavian 245, 306 compatibility 45, 104, 157, 159, 166f., 177f., 186, 192 case and word order position 167 of linkers, 45, 104, 157 of oblique word order position and lexical oblique case 164 complements 20 Comrie, B. 100, 114, 117, 299, 308 configuration 3, 45, 95, 155f., 158f., 162, 167, 242, 299, see also word order case and 167, 213, 176ff., 242 obliqueness and 161ff., 166, 223 constraints 20, 31, 41f., 185 as diagnostic for arguments and adjuncts 61 on lexical insertion 116 ordering 48 control 3, 23, 25f., 88f., 176, 183f., 186ff., 198, 236ff., see also Equi Controllee Constraint 187 coordination 208ff., 213 cost 233, 291 Cowper, E. 272, 309 Croft, W. 258, 297, 309 curly-bracket notation 53 Curme, G. 27, 228, 309 D-Structure (GB) 7, 224, 27Iff.

319

Dal, I. 27, 41, 236, 309 Danish 252, 254ff. dative 2, 8, 19f., 22ff., 29, 50f., 88, 97, 102f., 178, 181f., 200, 230f., 245ff., 280, 283ff., 288, 290, 303f., see also experiencer, goal Dative Sickness (pdgufallssyki) 22ff., 51f., 55ff., 59, 214, 218, 228f., 236, 304f., 306 diachronic aspect 226 evidence about class of linking features, as 62 "Generalized" 61 idiosyncratic accusatives and 306 limitations on 59 property of verbs, as 233 sociolinguistic correlates 295 trend, as 233ff. Dative Substitution, see Dative Sickness dative/agentive (Japanese) 94 Davies, W. 151f., 309 declarative particle (Nama) 116, 135, 139ff. deep grammatical relations 15 default(s) 70, 83, 97, 101, 241, 301 argument category 34f., 42, 190 argument ordering 18 behavior and non-dissimilation 291 case 40, 44ff., 70f., 77, 82ff., 97ff., 101,. 107, 114, 125f., 150, 290f., 296f. lexical case 28 linker 85, 130f., 137, 143, 157, 290f. semantic basis for subject selection 18 Default Linker Non-Dissimilation Prediction 83f., 88, 97, 102, 111, 291 Corollary 84 Delbriick, B. 39, 244, 309 demotion 8, 10f., 13, 15, 18, 19 Dench, A. 168, 309 Denison, D. 226, 309 derivational complexity 224 diachronic trend 220, see also change(s) in linking diglossia 228, 23If. DIR(ect) feature 32, 34f., 38, 42 direct arguments 32, 42 direct linking 2, 3, 5, 16, 144, 209f., 253, 289, 292f., 298 "directional" procedural interpretation of case-marking 76f. directionality of linking changes 221ff., 287 disjunction 53, 167, 214, see also case alternations, Dative Sickness, Nominative Substitution dissimilation 119, 290, see also Linking Dissimilation Constraint, Default Linker Non-Dissimilation Prediction

320

Index

Distinctness 119f. distribution 20, 38, 42 ditransitive(s) 2, 4, 13, 19, 88f., I l l , 125, 137, 142, 144f., 155ff., 159, 162f., 165f., 182, 184, 187, 194ff., 197ff., 209ff. passives, see passive ditransitive Dixon, R. 115,309 Diyari 121, 123ff., 301f. Double -o Constraint (Japanese) 92f., 114 generalized 114 double use of cases 88f., 144, see also Default Linker Non-Dissimilation Prediction double linking 104 Dowty, D. 4, 6, 17f., 34ff., 309, 312 Drosdowski, G. 309 Dryer, M. 30, 309 DS, see Dative Sickness dummies 153 Dutch 161ff., 180 Dyvik, H. 309 Eastern Adyghe 108 see Kabardian economy 40, 290, 296 Einarsson, S. 41, 309 Elmer, W. 226, 309 Elsewhere 39, 40f., 71, 85, 138, 112, 114, 206 Condition 44, 65, 274, 302 Pattern Prediction 84, 89, 99, 102f., 112, 130, 137 Elsewhere-like behavior 65, 82 English 3, 11, 37, 91, 113, 130, 133f., 155ff., 162, 167, 178, 191, 197, 222, 226ff., 231, 235f., 239, 244, 25Iff., 255ff., 303 adverbs 157 agreement 155 ditransitives 155ff., 159, 179, 197 history of 224 impersonal to personal shift as analogy 235 linkers 156 loss of impersonals 226ff. passivization 147, 197, 261 positional linking 133, 155, 158 word order freedom 155ff., 158f. entailments 35f. Equi 176, 186, 189, 247, 297, see also control ergative 70, 79, 101f., 104ff., 11 Iff., 113, 121, 131, 243, 296, 298 adverbial(s) 108, 112,296 case-marking 79 default, as 107, 111, 113 label 109, 121

marker from by-phrase 243 type 3 lOlff., 298 type 4 108ff., 11 Iff., 131 ergative-absolutive case-marking 101, 107, 110, 115 ergative-absolutive languages 76ff., 80, 82 ergativity, 115, 243 ergativity, split, see split NP ergativity evaluation metric 214, 220, 233 Evans, N. 168, 309 exception features 37, 219f., 227, 232, 236 exceptions 34, 291, 299 EXP(eriencer) feature 18, 20, 32, 34f., 37, 42, 53, 54, 59, 61, 63, 172 experiencer 18ff., 23f., 28, 32, 34f., 41, 51, 53, 59, 61, 172, 193, 230, 243, 246f., see also psych verbs argument category (syntactic), see EXP(eriencer) feature explanatory value 17 extended categorial framework 3 external argument 16, 106, 114, 147 extraction (Icelandic) 176 extrinsic ordering 270 Faarlund, J. 305, 309 Faroese 253, 259ff., 263, 265, 274ff., 278, 282f., 285f., 287, 296 case dissimilation 282ff. case semi-preservation 270ff. optional preservation of lexical case 260, 265 passive 261f., 280 permutation of arguments 285 scattered case 276, 284ff. feature cooccurrence restriction (FCR) 118, 241f., 301 feature loss 236 feature specification defaults (FSDs) 301, see also defaults feature structures 31, 217, 295 Fillmore, C. 5, 309 final 1 (RG) 15 final 3 (RG) 30 Finnish 100 Fischer, O. 226, 232f, 309f. French 256, 266f, 274 isolated case semi-preservation 266ff., 282 passive 267f. functional information 19, 62, see also grammatical function(s), grammatical relation(s) Gaaf, W. van der 226, 310 Gaedicke, C. 39f, 85, 290, 296, 310

Index 321 gaps 21 Iff. Gazdar, G. 31, 37, 301, 310 Generalized Phrase Structure Grammar 31, 140, 299 genitive 28, 245ff., 256f. Georgian 100, 117f., 237, 300f. Gerdts, D. 169, 297, 310 German 6, 19, 23, 25f., 34, 40f., 43, 47, 52, 54f., 63, 133, 144, 157, 159, 165, 168, 181, 197f., 228, 231ff., 235, 239, 248, 250, 255f., 258, 296f., 304, see also Bern German, Grisons German acquisition of behavioral and coding subject properties 238 dative 164, 248 ditransitives 159 Dative Sickeness 26f., 51ff., 229, 304 nominative and agreement 296 Nominative Substitution 228, 231, 297 passive, (im)personal 181, 262 prepositional linking 255f. quasi-asymmetric object language, as 197f. word order freedom 159 Germanic languages 30, 39, 60f., 231, 237, 244, 250, 253 Gildersleeve, B. 145, 310 goal(s) 8, 13, 18, 34ff., 63, 88, 110, 155ff., 158ff., 163ff., 172, 199f., 205f., 246f., 257, 282 GOAL feature 34ff., 88, 137, 145, 199f., 299 Goddard, C. 121ff., 125ff., 301f., 310 Goethe 228 Gothic 230, 233, 245, 305 Government and Binding Theory (GB) 7, 9ff., 15, 17, 106, 143, 146, 153, 166, 168, 174, 182f., 186, 270ff., 293f., see also Principles and Parameters case semi-preservation, and 271f. lexical stipulation, and 291 Case Theory, see Case Theory (GB) ordering of levels 273 GPSG, see Generalized Phrase Structure Grammar grammatical cases 44 grammatical function(s) 3, 5, 9, 15, 17, 20, 23, 39, 184, 208f., 289, 293f., see also grammatical relation(s) grammatical linking (KLT) 8ff., 11, 210, 238 and obliqueness (KLT) 293 grammatical relation(s) 1, 11, 15f., 143, 155, 291, see also grammatical functions

grammatical subjecthood, see subjecthood, grammatical grammaticalization 62, 167, 222, 225f., 241, 244, 254, 257f., 296, 302, 305 word order 222 Greek (Classical) 86ff., 90f., 102, 131ff., 139, 144f., 157f., 168, 172, 199, 239, 259, 262ff., 269, 275f., 278, 282, 298 case semi-preservation 262f., 270 (non)preservation of lexical case 260, 265 passive, (im)personal 262f. Gricean imperatives 65, 68 Grimshaw, J. 63, 295, 310 Grisons German 164ff., 180, 253 Grosse, S. 313 Gruber, J. 310 - H A specification 71, 73ff., 79f., 91, 119, 273 generalized to -XA 80 Hagman, R. 135, 137f., 310 Hale, K. 7, 101, 104ff., 299, 310 Halldorsson, H., 29, 228ff., 294, 305, 310 Harada, S. 92, 310 Harbert, W. 182f., 308, 310 Harris, A. 30, 300f., 310 Harris, M. 310 Head-Driven Phrase Structure Grammar 140, 186, 189f., 193 Heine, B. 222, 315 Henniss, K. 186, 310 Hermon, G. 308 hierarchical argument structure 203 hierarchy 18, 115 High German 245, see also German "highest available" 185f. "highest" (KLT) 293 highest in order of arguments 293 highest 0-role (KLT) 293, 298 Hindi 100, 168 Hirt, H. 305, 311 historical triggers 223, 225 Hock, H. 305, 311 Hodler, W. 304, 311 Hoekstra, T. 163, 311 Holmberg, A. 176, 252, 311 HPSG, see Head-Driven Phrase Structure Grammar Hungarian 100 hypercorrection 295, 306 Icelandic 2, 19, 22ff., 26ff., 33ff., 38, 40f., 44f., 47ff., 55, 57ff., 62f., 67, 69, 71, 74ff., 78f., 86ff., 91, 102, 133, 139, 154, 157ff., 167f., 171ff., 176ff., 186ff.,

322

Index

191, 194, 198ff., 205ff., 21 Iff., 216, 220ff., 228, 23Iff., 236, 244f., 247, 249ff., 256f., 259f., 271, 275, 283ff., 288, 295f., 302f., 306 acquisition of behavioral and coding subject properties 238 agreement 173ff., 296 case and word order 176ff. case preservation 216ff., 259f., see also Icelandic passive, Icelandic Raising ditransitives 178ff., 182, 187, 200, 194ff., 209ff. goal 8, 34ff., 63, 172, 199, 205f. idiosyncratic case 25, 38, 41, 50, 59, 62, 173, 214ff., 219 impersonal to personal shift as analogy 235 lexical case 23ff., 171ff., 180, 213, 232, 246, 295 mixed symmetric-asymmetric object language, as 194, 200ff. nominative 3, 19f., 23ff., 29, 38ff., 42ff., 47, 173ff., 185, 276, 296 objects 177, 210ff., 302 passivization 67, 148, 181, 184, 187, 194, 198, 200ff., 216, 261f., 295 word order 171, 176ff., 299 idioms 7 idiosyncratic case 25, 38, 41, 50, 59, 62, 180f., 191, 199, 214ff., 219, 243, 256, 282, 306, see also lexical case partial semantic predictability of 33f. rise of 243 Symmetric Object Operation and 206 idiosyncratic Case (GB) 272 impersonal verbs 231, 226ff., 231, 244 loss of 226ff., 231 incompatibility 143, 161 index 303 indirect object If., 63, 162, 290 Indo-European languages 40, 239, 244, 305 case syncretism 122 infinitival constructions 188 INFL 114, 147, 174, 182f. information content 20 inherent features of NPs 116 inheritance 162, 241 initial 1, 2, etc. (RG) 15f., 30, 144 instrumental 299 interchangeability 122, 127 interconnectedness of predictions 290 internal argument 16 interpretation 34f., 37, 39, 190f., 256, 258, 298ff. adverbs 90 and diachronic explanation of 256

argument 18, 35, 54, 172 ditransitives 158 interpretational dependencies 6 inversion 45, 176f., 201, 207, 247f. Irish 302f. italic notation 47 Jackendoff, R. 6, 18, 28, 36, 50, 58, 63f., 76, 182, 184, 199,295,297,311,316 Jakobson, R. 39f., 99, 311 Japanese 6, 71f., 74f., 79f., 91ff., 101, 114, 148f., 151, 154, 158, 195ff., 205, 240, 272, 296, 299 as symmetric object language 195f. causative 79, 92f., 96 Double -o Constraint 92, 94 double nominative clauses 98 passivization 72, 74, 148f., 195ff. positional linking and case 95ff., 100, 299 Jelinek, E. 104, 311 Jespersen, 0.223,226, 233, 311 Joshi, S. 168,311 Kabardian 79, 108, HOff., 113f., 151f., 156, 167, 294, 296, 300 antipassivization 151 f. Kala Lagaw Langgus 149f. Kanerva, J. 9, 48, 308 Kannada 30, 303f. asymmetric object behavior 303 dative subjects 303f. karaka's (Panini) 5 Keenan, E. 3^ 236 Kemenade, A. van 244, 311 Keraseva 108, 314 Kim, Y.-J. 300, 311 Kiparsky, P. 5ff., lOff., 19, 44, 48, 65, 96, 116, 140f., 144, 155ff., 163, 166, 168, 173, 179, 185, 210, 213, 233f., 238, 252, 258, 265, 293f., 298, 300, 311 Kiparsky's Linking Theory (KLT) 5ff., 11, 14, 16, 18, 210, 293 see also Kiparsky, P. Klein, E. 310 Klokeid, T. 149f., 239f., 311 Klopstock 228 KLT, see Kiparsky's Linking Theory Korean 30, 169, 300, 302 adverbial case 300 Case Stacking 297 type 1-type 2 split 300 Koutsoudas, A. 44, 65, 312 Kress, B. 41, 312 Kroeger, P. 187, 312 Kuhner, R. 145, 265, 312

Index Kuipers, A. 108ff., 312 Kulonen, U.-M. 266, 312 Kuno, S. 73f., 98, 100, 148, 272, 299, 312 Kurylowicz, J. 222, 225, 256, 312 -LA specification 79, 102, 119 and antipassivization 149 generalized to -XA 80 Ladusaw, W. 6, 35, 312 language acquisition 18, 220, 227, 236, 302 Lardil 239f., 258 Latin 4, 8f;, 122, 144f., 158, 239, 250, 256 Lavotha, 6 . 265, 312 LDC, see Linking Dissimilation Constraint Leek, F. van der 226, 232f., 309f. left dislocation 296f., 299 Lehmann, C. 222, 312 Lessing 228 levels 15 Levin, B. 103f., 106, 312 lexical case 23ff., 171ff., 180f., 185, 197f., 213ff., 231f., 245ff., 250, 259ff., 268, see also idiosyncratic case; accusative, lexical, etc. Lexical Conceptual Structure (LCS) 63f., 297 lexical decomposition 34 lexical diffusion 239 lexical entries 2, 214ff., 219 lexical exceptions 232 Lexical Functional Grammar (LFG) 9, 31, 143, 184ff, 188ff., 204, 208ff., 217f., 294 see also Lexical Mapping Theory lexical insertion 116, 151, 300 Lexical Mapping Theory (LFG) 9, 210, 217f., 303 lexical semantic classifications 5 lexical semantics 18, 63 lexical stipulation 291 Lexical Structure 7f, 293 lexical word order position 161 lexicalism 186 lexicalizations, new 291 lexicon 7 LFG, see Lexical Functional Grammar Lightfoot 223ff., 231ff., 244, 312 Limitation Parameter 74f., 80f, 86, 92, 100, 107, 172 preliminary 75 revised 80 LINK 31f., 38, 42, 44, 46, 50, 61, 68, 72f., 75, 81f, 86, 95, 155, 173, 190, 223, 241, 275ff., 280ff., 286, 290 linker(s) 2, 16f., 82, 118, 165, 173f, 176, 200, 245, 252, 254, 289 argument versus adjunct 255

323

distinguishing 289 loss of 258 morphologically related 255 reformulation of 258 syntactic 16 types of 45, 167ff. linking, 5, 116 change, see change(s) in linking theory, in general 289 versus semantics 299 word order 244, see also positional linking, word order linking Linking Condition 42, 44, 95, 174 Linking Dissimilation Constraint 37, 47f., 67f., 74, 76, 83, 87, 94ff., 116, 119, 172f., 200, 282, 290, 297, 301 argument structure and 68 case semi-preservation and 277ff., 297 control and 187 cross-linguistic effects 114 parameterized 78 predication and 193, 299 Revised 120 scattered case and 282ff. local ordering 270 local simplifications 233 localism 1 locational approaches 1, 256 Lockwood, W. 262, 312 logical object 7, 15 logical subject 7, 9, 11, 15f., 19, 23, 30, 32 logophoricity 249f. long distance anaphora 249f. Ludwig, A. 164 McCloskey, J. 302, 312 Mainland Scandinavian 256, see also Scandinavian languages, Danish, Norwegian, Swedish Malayalam 186 Maling, J. 23, 28, 34, 40, 46, 50, 58, 76, 143, 174, 176, 179, 182ff., 199ff., 204f., 207, 209, 214ff., 219, 249, 251, 260, 295ff., 300, 302, 313, 316 Marantz, A. 7, 73f., 272f, 313 Marathi 168 marked behavior 34 markedness 199, 291 mediated linking 5, 177, 294 Meillet, A. 222, 313 Middle English 227, 235, 286f. Middle High German 228, 239 Mirror Principle 300 Mitchell, B. 268f., 304, 313 Mithun, M. 301, 313 mixing levels 301

324

Index

Miyagawa, S. 99, 313 Modern Icelandic, see Icelandic Mohanan, T. 168, 299, 313 morphological case 25, 122, 127, 129, 167, 227 defined 122 morphological spell-out (GB) 272 morphological versus syntactic case 16, 31, 96, 289, 298, 301 case semi-preservation, and 282 morphological zero or default case 40 morphologically unmarked case 101, 296 morphology 17 morphosyntax 3, 5, 20, 32, 45, 292ff., 299 see linking Moser, H. 313 Moshi, L. 194, 308 multiple linking 45, 166ff., 169ff., 176ff. multiple morphological case 168 Nama (Hottentot) 116, 135ff., 139ff., 158 Nash, D. 102, 105ff., 299, 313 Noll, C. 44, 65, 312 nominative 3, 19f., 24f., 29, 39f., 42ff., 47, 87, 97, 173ff., 185f., 276 adverbials type 2 99 agreement and 296 change to absolutive 243 experiencer class 54 objects 50f., 180ff., 185ff., 192f., 297 subject coding property, as 237 type 1 87, 91 type 2 97, 131 Nominative Sickness/Substitution 23ff., 51, 53ff., 57, 60, 228, 230ff., 236, 295, 305 acquisition of coding properties, as 238 diachronic aspect 226, 228, 233ff., 238, 243 reverse (Icelandic) 236 property of verbs, as 233 scattered case and (Faroese) 285 sociolinguistic correlates 295 nominative-accusative case marking/ systems 71, 74, 77f., 80, 82, 87, 172, 242 non-arguments 90f., 102, see also adjuncts non-clause-bounded reflexivization, see long distance anaphora non-preservation, optional 260, see case semi-preservation non-promotional passives 153f. Noreen, A. 253, 313 North-West Caucasian languages 108, 110, 114, 151 Norwegian 159, 252ff., 306 notation 17

NP-Movement 7, 188 NP-Structure 7 null subject 244 Nygaard, M. 175, 228, 251, 313 Ob-Ugrian languages 265f. OBJ function (LFG) 209f., 217f. object(s) 3, 9, 113, 135, 156, 163, 177, 210ff., 238, 251, 302 (a)symmetric, see symmetric object(s) nominative, see nominative objects secondary (2OBJ) (LFG) 209f. objecthood 8, 238, see also termhood acquisition of 238ff., 241, 250 behavioral, acquisition of 241 Obligatoriness Condition 13 Obligatoriness Constraint (KLT) 11 oblique(ness) 3, 9f., 15, 19, 26, 55, 60, 102, 130, 159, 161, 163, 167, 171, 179, 181, 187, 197, 203, 210, 213, 229, 237f., 241, 243f., 250f., 291, 293, 302, 305f. agreement and 210 behavioral 159, 164f., 239, 248 coding 164 feature of linkers, as 241 inheritance 19, 241, see also Obliqueness Inheritance lexical case and 242 morphological and syntactic 19 nonobliqueness default 241 syntactic 8 word order position and 159f., 161 ff., 166, 180, 210, 223, 252 Obliqueness Constraint 160, 241f. Obliqueness Inheritance 161f., 166ff., 170 obsolescence 227 Old English 52, 157f., 227, 232, 245, 253, 256f., 268f., 274, 283, 287, 304f. ACC/DAT-ACC/DAT sentences (Old and Middle English) 287 case alternations 304f. case dissimilation 283 case preservation 268f. impersonal verbs 232 passive, (im)personal 268f. semi-preservation isolated 266, 282 Old Icelandic 52, 175, 228ff., 233, 245ff., 288, 305f. Dative Sickness 228f., 305 lexical case obliqueness of 245ff. Nominative Substitution 230 obliqueness 229, 250, 306 passive 246, 306 Old Norse see Old Icelandic, Old Norwegian Old Norwegian 305, see also Old Icelandic

Index Old Saxon 304 Old Swedish 253, 305 One-to-One Constraint (KLT) 10, 14 opacity 243 operation 32 Optimality Hypothesis 214, 220f., 234 optional argument 205 optionality 51ff., 56, 58ff., 131ff., 138, 216, 228, 232, 234f., 268, 296 ordered argument theories 19, 31, 203 ordering of levels (GB) 273 ordering, partial 270 Ostler, N. 5, 8, 313 Ostyak 265f. PamaNyungan 239 Panyjima 168 parameter changes 225 parameter setting 227, 242f. parameterization 291 parameters 74f., 185 parentheses notation 52 partitive 300 passive 9, 10, 15, 18, 48f., 72, 74, 130, 144ff., 147ff., 162f., 181, 201, 216, 246, 259, 261ff., 267ff., 280, 293, 295, 297 see also passivization adjectival 26If. ditransitives 165f., 184, 187, 194ff., 200ff. dynamic type 261 f. impersonal 181, 263, 267f. idiosyncratic case and (Icelandic) 202ff. non-promotional 153f. Paninian approach 293 reanalysis of 242 stative type 261 f. suppression view of 48, 130, 147 symmetric vs. asymmetric 163 verbal 26If. passivization 91, 146ff., 153, 188, 218, see also passive and antipassivization compared 15Iff. ditransitives 194ff. in two-case systems 147 suppression of the highest argument 146 paths (Attribute-Value Matrix notation) 32, 295 patient/theme 62 Paul, H. 313 Panini 5, 44, 293 percolation 161 perfect, reanalysis of 242 Perlmutter, D. 15, 144, 313 permutation of arguments 204ff., 208, 285

325

Persian 100 Peskovsky 40 phonological case change 224, 252 phrase structure 12, 17, 156, 158, 176, 188 PLACE feature 32 Pollard, C. 186, 313 Polynesian 237 Poser, W. 92f. 97f., 114, 313 positional linking 95, 137, 143, 154, 158, 179, 244, 299 see also linking, positional; linking, word order position; word order; configuration Postal, P. 149, 267, 313 pre-specification 301 predicators/predication 167, 190, 256, 297, 299 Preference Parameter 74, 76, 79ff., 83, 86, 92, 100, 107, 109, 152, 172 antipassivization and 149 feature names and 298 interaction with Limitation Parameter 81 "pre-linking" of lexical case 50, 182 prepositional linking 157f., 255ff. preservation of case, see case preservation, lexical case preverbal position 3, 45, 176, 251f., 275 primary object behavior 194, see also objects primitive functional information 144 primitives 15 Principles and Parameters 40, 224, 226, see also Government and Binding Theory Priority Clause (KLT) 11, 13, 20 priority in linking 20 priority of cases 44, 270, 276 priority of grammatical linking (KLT) 20 priority of idiosyncratic case 50, 271 privative features 298 PRO 183, 186, 188 productivity 34, 62, 291f., 296, 305 and restrictiveness 305 promotion to argument 296 proper subsumption 32 Proto-Germanic 231, 245 pseudo-passivization 298 psych verbs 22ff., 30, 182, 229f., 243, 247, 249f., see also experiencer Pullum, G. 310 Punjabi 100 quirky case 24Iff., see also idiosyncratic case, lexical case raising 176, 182, 188ff., 191ff., 236f., 247, 302f.

326

Index

case preservation and 192f. nominative objects and 192 scattered case and 306 versus control 189 RC, see Restrictiveness Constraint RCP, see Restrictiveness Change Prediction REAL 32, 42, 116, 193 realization 32, 155 reanalysis 223ff., 231, 234, 242, 243f. reconstruction, historical 254 redundancy 214 redundancy rule(s) 18, 216 redundancy-decreasing interpretation 37 reflexive pronouns 3, 12 and logophoricity 249 reflexivization 176, 237, 247, 249, 251, 302f. Reinhammer, M. 253, 314 Reis, M. 25, 314 Relational Grammar (RG) 1, 9, 15, 29, 63, 143f., 169, 240, 294 relativization 93 repeatability 85, 88, 97f., 108, 11 If., 137, see also dissimilation replacement rules 30 restrictiveness 20, 38, 44, 290 case preservation and 186 case semi-preservation and 277f., 282 case stacking and 297 linking change and 233ff. optimality and 220 productivity and 305 Restrictiveness Change Prediction 234f., 238ff., 242f., 253ff., 258, 306 acquisition of objecthood and 250 loss of obliqueness and 251 reconstruction and 255 shift to semi-preservation and 287f. word order and 245 Restrictiveness Constraint (RC) 44f., 48, 51ff., 59, 68, 84, 143, 167, 172f., 180, 184, 247, 286, 290, 302 argument structure, at 60 case preservation and 182 case semi-preservation and 276ff. case stacking and 169 constraint on lexical insertion, as 116 scattered dative and 286 semantic selection alternative 63, 65 RG, see Relational Grammar Riemsdijk, H. van 7, 314 Robinson, O. 274, 314 Rogava, G. 108, 314 Rognvaldsson, E. 176, 179, 211, 247, 303, 314

Romance 239, 256 rule ordering 44, 260, 281 rule ordering effect 44 rules 234 Russian 39, 239 S-curves of innovation 225 S-Structure (GB) 7, 271ff. Sadock, J. 135, 138f., 141, 302, 314 Sag, I. 186, 310, 313 Saito, M. 95, 99, 100, 314 Sanders, G. 44, 65, 312 Sapir, E. 314 Scandinavian languages 244f., 253f., 305f. linking summary of 254 Mainland 255, see also Danish, Norwegian, Swedish scattered rules 274, see also case, scattered Schrobler, I. 313 Schwyzer, E. 263, 314 scrambling 95 Seefrantz-Montag, A. v.on 226, 314 Sehrt, E. 304, 314 semantic blocking principle 63 semantic cases 102, 112, see also semantic generalization(s) semantic experiencer(s) 35, 172 "semantic" generalization(s) 41, 256, 295, 299 semantic goal 36, 88 semantic linking (KLT) 9, 11, 210, 238 semantic selection 65, 190 semantic structure 63 semantically (un)restricted grammatical linking (KLT) 9 semi-preservation, see case semipreservation Shevelov, G. 133, 153, 314 Sigurdsson, H. 182f., 244, 314 Silverstein, M. 115, 314 simplicity 220, 233f., 305 Simpson, J. 101, 104, 106f., 314 sister-of-V position 13f., see also verbadjacent position sister-of-V' position 13 Smith, H. 179, 249, 296, 298, 300, 314f. Smyth, H. 89, 91, 132f., 264, 315 Sobin, N. 153, 315 sociolinguistic variation 30 Soderwall, K. 305, 315 Songhai 100 South Asian languages 110, 168 South Caucasian (Kartvelian) 117 specific-over-general principle 297, see also Elsewhere Condition, Restrictiveness Constraint

Index split NP ergativity 100, 115, 125f., 129, 300 factors conditioning 115 three-case analysis of 125f., 129 split systems 290, 300 Sridhar, S. 30, 303, 308, 315 stability of linkers and restrictiveness 253 stativity 98, 300 Steele, S. 37, 315 structural case (GB) 157, 272 structural position-based approaches to linking 289 structure sharing 189ff., 193 structured feature structure 31 stylistic rules 98 SUBCAT feature 31, 190, 193 subcategories 37 subcategorization 40, 193 subject 3, 9, 10, 19, 102, 141ff., 176ff., 197f., 247, 251, 306, see also subjecthood logical 23 selection 18, 64 subjecthood 2f., 304f., see also subject acquisition of 237f., 241, 243, 250 coding 3 behavioral/grammatical 3, 8, 19, 23, 174, 176, 182, 187, 201, 207, 236ff., 241, 250, 295, 304 substitutability 127 substitution classes 122f. subsumption 20, 31f., 44, 295, 301 suppression 146 of lowest argument in antipassivization 149 view of passive, see passive, suppression view of surface form 16, 289, 291, 294 surface morphology 30 surface subject 15 Svavarsdottir, A. 27, 57, 236, 294f., 315 SVO word order 224, 233 rise of 244 Swedish 252, 254ff. Swiss German 164 symmetric object(s) 194ff. languages 201, 194ff., 303 operation 195ff., 203f., 206 parameter 196, 203, 208 versus asymmetric object languages 194ff. syncretism 122, 126, 129, 233 synonymy 298 syntactic "case" (linker) frames 20 syntactic argument categories, see argument categories syntactic case, see case, syntactic

327

syntactic category, fully specified 301 syntactic experiencer, see experiencer, syntactic syntactic linkers, see linking, syntactic syntactic obliqueness, see obliqueness, syntactic syntactic selection 190f., 193 syntactic versus semantic case 104, see also case, syntactic syntactic versus morphological case 20, see also case, syntactic syntax and linking 291 Tagalog 100 tense 118 term(s) 9, 171, 247, 260, 197, 305, see also termhood termhood acquisition of 238, 245, 251 behavioral acquisition of 240ff. coding and Obliqueness Constraint 242 "thematic" generalizations 37 thematic hierarchy 6f. Thematic Hierarchy (KLT) 155, 293 thematic relations 4, 38, 61, 172, 210, see also thematic roles, theta roles thematic role(s) 17, 35, 37, 63f., 69, 105, 215 goal 68 labels 17, 19 KLT 155, 157, 293 LFG 217 thematic structure 3, 60f. thematic uniqueness 36, 173 theme 18, 163, 218 KLT 157 theme/patient 39 theta roles 8, 11, 18, 64 GB 271f. theta-structure 9 Theta Theory 7, 10 Thrainsson, H. 23, 34, 40, 46, 144, 174, 176, 179, 182ff., 198, 200ff., 204f., 207, 209, 249, 251, 260, 296f., 302, 314ff. three-case systems 119f., 125 time adverbials 84, 296 ergative (Kabardian) 113 (Greek) 133 (Nama) 135, 137 TO function in LCS 63 tone (Nama) 302 "top-down" linking 46f. topicalization 23, 93, 140, 136, 176, 304 Toribio, A. 182f. Trace Erasure Principle (GB) 224 transitive clauses 48, 92

328 Index Transparency Principle 224 Traugott, E. 222, 226, 315 triggers 231, 242 truth conditions 36 Turkish 100 two-case systems 70ff. type 1 case system 71, 79, 81, 83, 86ff. type 2 case system 71ff., 81, 91ff., 100 type 3 case system 79, 81, lOOff., 107 type 4 case system 78f., 81, 107ff. types of linkers 19, 96, 154, 167, 294 Udi30 Ukrainian 133, 147, 153f. unaccusativity 15f. unidirectional changes 225, see also linking change, unidirectionality of unification 31f., 42, 140, 161, 277, 295 unification-based framework 20, 32 "unmarked" word order 6 Uszkoreit, H. 195, 315 valence changing operations 130, see also passive, causative, etc. Van Valin, R. 28, 174f., 297, 315 variation 235, 261, 285, 306 Vedic Sanskrit 39, 85 verb meaning 24 verb movement 179 verb phrase 176f., 180, 304 verb-adjacent position llff., 45ff., 95ff., 154ff. verb-first, -second (Icelandic) 176f, 302 verbal complex (Kabardian) 110, 112 verbal prefixes (Caucasian) 108 Verhagen, A. 162, 315 Vigfusson, G. 228, 308 Visser, F. 226, 268f., 315 Vogul 265f. VP-external position (_V") 12f., 156f. VP-internal position (V") 12, 156ff. Warlpiri 79f, lOlff, 106f ergative 102, 104ff., 299 history of verb classes 105f.

Wasow, T. 261, 316 weakly governed accusatives 40, 99 Wechsler, S. 157, 163, 253, 316 well-formedness conditions 31 Western Desert Language 127 Wh-Movement 7, 140f., 291 Whitney, W. 40, 316 Williams, E. 7, 314 Woolford, E. 298, 316 word order (position) 2f, 6, llf, 17, 19, 130, 137, 154, 164, 167, 171, 244, 252f., see also verb-adjacent position, VP-external position, VP-internal position Case and (GB) 294 case and 95, 159ff, 164ff, 176ff., 242, 253, 299, 304 change 222, 244f., 254, see also changes in linking freedom and obliqueness 159f., 161ff., 223 lexical 161 oblique 163ff. sister-of-V', introduced 12 "working 1" (RG) 15 -XA specification 80, 82, see also - H A specification, -LA specification Yankunytjatjara 127f, 302 Yip, M. 28, 50, 58, 76, 182, 184, 199, 295, 297, 316 Youn, C. 169, 297, 310 Yukulta 239f. Zaenen, A. 23, 34, 40, 46, 143, 174, 176, 179, 182ff, 200ff., 204f, 207, 209, 214ff., 219, 260, 296f, 302, 316 zero case 101, 118,296