Sentence Processing: A Crosslinguistic Perspective

SYNTAX AND SEMANTICS VOLUME 31 EDITORIAL BOARD Series Editors BRIAN D. JOSEPH AND CARL POLLARD Department of Linguis...

Author: Dieter Hillert

210 downloads 1862 Views 26MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Report copyright / DMCA form

DOWNLOAD PDF

SYNTAX AND SEMANTICS VOLUME 31

EDITORIAL BOARD

Series Editors BRIAN D. JOSEPH AND CARL POLLARD Department of Linguistics The Ohio State University Columbus, Ohio

Editorial Advisory Board JUDITH AISSEN University of California, Santa Cruz PETER CULICOVER The Ohio State University

PAULINE JACOBSON Brown University MANFRED KRIFKA University of Texas

ELISABET ENGDAHL University of Gothenburg

WILLIAM A. LADUSAW University of California, Santa Cruz

JANET FODOR City University of New York

BARBARA H. PARTEE University of Massachusetts

ERHARD HINRICHS University of Tubingen

PAUL M. POSTAL Scarsdale, New York

A list of titles in this series appears at the end of this book.

SYNTAX and SEMANTICS VOLUME 31 Sentence Processing: A Crosslinguistic Perspective Edited by

Dieter Hillert Department of Psychology University of California, San Diego LaJolla, California

ACADEMIC PRESS San Diego London Boston New York Sydney Tokyo Toronto

This book is printed on acid-free paper. Copyright © 1998 by ACADEMIC PRESS All Rights Reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the Publisher. The appearance of the code at the bottom of the first page of a chapter in this book indicates the Publisher's consent that copies of the chapter may be made for personal or internal use of specific clients. This consent is given on the condition, however, that the copier pay the stated per copy fee through the Copyright Clearance Center, Inc. (222 Rosewood Drive, Danvers, Massachusetts 01923), for copying beyond that permitted by Sections 107 or 108 of the U.S. Copyright Law. This consent does not extend to other kinds of copying, such as copying for general distribution, for advertising or promotional purposes, for creating new collective works, or for resale. Copy fees for pre-1998 chapters are as shown on the title pages. If no fee code appears on the title page, the copy fee is the same as for current chapters. 0092-4563/98 $25.00

Academic Press a division of Harcourt Brace & Company 525 B Street, Suite 1900, San Diego, California 92101-4495, USA http://www.apnet.com Academic Press Limited 24-28 Oval Road, London NW1 7DX, UK http://www.hbuk.co.uk/ap/ International Standard Book Number: 0-12-613531-2 PRINTED IN THE UNITED STATES OF AMERICA 98 99 00 01 02 03 QW 9 8 7 6

5

4

3 2 1

CONTENTS

Contributors Acknowledgments

xiii xvii

From Alexander to Wilhelm von Humboldt: A Crosslinguistic Perspective DIETER HILLERT 1. 2. 3. 4. 5. 6. 7.

Introduction Word Recognition Fixed Expressions Grammatical Constraints Empty Categories Parsing Conclusion References

1 3 5 7 8 9 9 10

Lexical Ambiguity Resolution: Languages, Tasks, and Timing KATHLEEN V. AHRENS 1. 2. 3. 4. 5. 6.

Introduction Context-Dependent and Context-Independent Models Natural Speech Comprehension Language-Driven Differences Experiment: Lexical Ambiguity Resolution in Mandarin Chinese Conclusion References V

11 12 14 17 19 27 29

vi

Contents

Crosslinguistic Variation and Sentence Processing: The Case of Chinese PING LI 1. 2. 3. 4. 5. 6.

Introduction Experiment 1: Sentence Interpretation Experiment 2: Sentence Gating Experiments: Cross-Modal Naming Experiment 4: Word Gating General Discussion References

33 36 40 44 46 49 51

Comprehension Repair in the Processing of a Short Oral Discourse Involving a Lexically Ambiguous Word GIYOO HATANO AND KEIKO KUHARA-KOJIMA 1. Strategies for Reducing Lexical Ambiguity 2. Repairing Comprehension: Experiment 1 3. Recognizing Incongruity and Revising the Initial Constiial: Experiment 2 4. Economy of Listeners'Repair Strategies References

56 57 64 67 68

Ambiguity of Reanalysis in Parsing Complex Sentences in Japanese YUKI HlROSE AND ATSU INOUE

1. 2. 3. 4. 5. 6.

Introduction Easy versus Difficult Reanalysis Previous Proposals Effect of Thematic Ambiguity of the Head Noun Further Experimental Evidence Discussion References

71 72 74 76 80 83 90

The Processing of Empty Subjects in English and Japanese TSUTOMO SAKAMOTO AND MATTHEW WALENSKI 1. 2. 3. 4. 5.

Introduction Processing of Empty Subjects in English Processing of Empty Subjects in Japanese Comparison of English and Japanese Thematic Hierarchy

95 96 98 100 102

Contents

6. Theta-Checking Strategy 7. Models of Human Parser 8. Final Remarks References

Vii

105 107 108 109

The Production of Agreement in English and Japanese: Animacy Effects (or Lack Thereof) JANET L. NICOL 1. Introduction 2. Studies of Agreement Errors 3. General Discussion References

113 116 126 127

Context-Independent Sentence Processing SUSAN BORSKY AND LEWIS SHAPIRO 1. 2. 3. 4.

Introduction Investigating Context Effects The Importance of Experimental Tasks for Interpreting Data Conclusions References

131 132 145 146 147

The Influence of Canonical Word Order on Structural Processing TRACY E. LOVE AND DAVID A. SWINNEY 1. 2. 3. 4. 5.

Introduction Basic Issues Processing of Long-Distance Dependencies in English A Crosslinguistic Perspective Conclusions and Discussion References

153 154 160 162 163 164

Lexical Access and Coreference Processing in Bulgarian MAXIM I. STAMENOV AND ELENA ANDONOVA 1. 2. 3. 4. 5.

Introduction Background Some Specific Features of the Bulgarian Language The Problem of Gap Filling and Cross-Modal Lexical Priming Discussion References

167 168 169 169 178 180

viii

Contents

The Contribution of Word Form and Meaning to Language Processing in Spanish: Some Evidence from Monolingual and Bilingual Studies JOSE E. GARCIA-ALBEA, ROSA M. SANCHEZ-CASAS, AND JOSE M. IGOA 1. Introduction 2. Are Morphological Relations Reducible to Form and Meaning Relationships ? 3. How Do Form and Meaning Contribute to the Cognate Effects in Bilinguals? 4. Is the Cognate Effect a Special Kind of Morphological Effect? 5. Concluding Remarks References

183 186 194 199 201 203

Accounting for Crosslinguistic Variation: A Constraint-Based Perspective ROBERT THORNTON, MARIELA GIL, AND MARYELLEN C. MACDONALD 1. 2. 3. 4. 5.

Introduction Psycholinguistic Approaches to Crosslinguistic Variation Pragmatic Constraint on Crosslinguistic Variation The Case of Word Order and Genitives Conclusions References

211 212 215 218 222 223

The Time Course of Attachment Decisions: Evidence from French JOEL PYNTE 1. 2. 3. 4.

Relative Clause Attachment Adjectival Phrase Attachment Prepositional Phrase Attachment The Dynamics of Attachments References

228 232 236 240 244

Verb Processing in German and English: Ambiguity, Discontinuous Forms, and Thematic Complexity DIETER HILLERT 1. Introduction 2. Verb Ambiguities 3. Discontinuous Verb Particle Combinations

247 248 250

iX

Contents

4. Thematic Complexity 5. Conclusions References

257 259 261

The Dimensional Conception of Space and the Use of Dimensional Prepositions in Different Languages THEO HERRMANN AND JOACHIM GRABOWSKI 1. 2. 3. 4.

Introduction Different Conceptions of the First Horizontal Axis and Its Poles Some Determinants of Origo Setting Experiments on the Interpretation of "in Front of" and "Behind" and Their Counterparts in German, Dutch, Italian, and French 5. A Psycholinguistic Explanation of Crosslinguistic Differences References

265 267 275 277 284 287

Syntactic Ambiguity Resolution in German BARBARA HEMFORTH, LARS KONIECZNY, CHRISTOPH SCHEEPERS, AND GERHARD STRUBE 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

Introduction The Differential Perspective Parallel Constructions Constraint Ordering Constituent Ordering and the Time Course of Phrase Structure Building Constituent Order and Thematic Prominence Attachment Ambiguities An Intermediate Summary Limits of Parametrized Head Attachment Conclusions References

293 294 296 297 299 300 302 305 306 308 309

Challenges to Recent Theories of Language Differences in Parsing: Evidence from Dutch DON C. MITCHELL AND MARC BRYSBAERT 1. 2. 3. 4.

Introduction Background A Closer Look at the Crosslinguistic Differences Theoretical Interpretations

313 314 315 318

X

Contents

5. 6. 7. 8.

A Preliminary Evaluation of These Proposals Further Evidence from Dutch Conclusions from Dutch Overall Conclusions References

324 326 330 331 332

Syntactically Based Parsing Strategies: Evidence from Typologically Different Languages MARICA DE VINCENZI 1. Introduction 2. The Minimal Chain Principle 3. The Minimal Chain Principle and the Processing of Declarative Sentences 4. Conclusions References

337 338 339 342 343

The Development of Sentence Comprehension in Italian and Serbo-Croatian: Local versus Distributed Cues ANTONELLA DEVESCOVI, SIMONETTA D'AMico, STAN SMITH, Ivo MIMIC A, AND ELIZABETH BATES 1. 2. 3. 4. 5. 6.

Background Overview of Italian and Serbo-Croatian Grammar Experiment 1 Experiment 2 Experiments General Discussion References

346 349 353 359 366 370 374

Role of Context in the Comprehension of Ambiguous Italian Idioms LUCIA COLOMBO 1. Introduction 2. The Comprehension of Figurative Language 3. The Comprehension of Ambiguous Idioms in an Off-Line Task: Experiment 1 4. The Comprehension of Ambiguous Idioms in On-Line Tasks: Experiments 2 and 3 5. Implications from the Present Data and Conclusions References

379 381 384 387 400 402

Contents

Xi

Compactness and Conceptual Complexity of Conventionalized and Creative Metaphors in Italian CRISTINA CACCIARI 1. 2. 3. 4. 5.

Introduction Metaphor Compactness Innovative Metaphorical Descriptions Conventional and Innovative Metaphors Conclusions References

Index

405 407 416 419 421 422 427

This page intentionally left blank

CONTRIBUTORS

Kathleen V. Ahrens (11), Department of Foreign Languages and Literatures, National Taiwan University, Taipei, Taiwan Elena Andonova (167), Cognitive Science Department, New Bulgarian University, Sofia, Bulgaria Elizabeth Bates (345), Center for Research on Language, University of California, San Diego, La Jolla, California 92093 Susan Borsky (131), Department of Communicative Disorders, San Diego State University, San Diego, California 92182, and Department of Psychology and Center for Complex Systems, Florida Atlantic University, Boca Raton, Florida 33432 Marc Brysbaert (313), Department of Psychology, University of Ghent, Ghent, Belgium Cristina Cacciari (405), Department of Psychology, University of Bologna, 40127 Bologna, Italy Lucia Colombo (379), Department of General Psychology, University of Padua, Padua, 35139 Italy Simonetta D'Amico (345), Dipartimento di Psicologia dei Processi di Sviluppo e Socializzazione, Universita La Sapienza, 00185 Roma, Italy Antonella Devescovi (345), Dipartimento di Psicilogia dei Processi di Sviluppo e Socializzazione, Universita La Sapienza, 00185 Roma, Italy Marica De Vincenzi (337), Institute of Psychology of the National Research Council, 00137 Rome, Italy Jose E. Garcia-Albea (183), Department of Psychology, Universitat Roviro i Vigili, 43007 Tarragona, Spain Mariela Gil (211), Department of Psychology, University of Southern California, Los Angeles, California 90089 xiii

xiv

Contributors

Joachim Grabowski (265), Department of Psychology, University of Mannheim, Mannheim, Germany Giyoo Hatano (55), Department of Human Relations, Keio University, Tokyo, Japan 113-0021 Barbara Hemforth (293), Center for Cognitive Science, Institute of Computer Science and Social Research, Freiburg University, D-79098 Freiburg, Germany Theo Herrmann (265), Department of Psychology, University of Mannheim, Mannheim, Germany Dieter Hillert (1, 247), Department of Psychology, University of California, San Diego, La Jolla, California 92093 Yuki Hirose (71), Program in Linguistics, Graduate Center, The City University of New York, New York, New York, 10036 Jose M. Igoa (183), Department of Basic Psychology, Univeridad Autonoma de Madrid, Madrid, Spain Atsu Inoue (71), Faculty of Economics, Kanto Gakuin University, Yokohama, Japan Lars Konieczny (293), Computational Linguistics, University of Saarland, D66041 Saarbriicken, Germany Keiko Kuhara-Kojima (55), Department of Psychology, Tokyo Woman's Christian University, Tokyo, Japan 168-0082 Ping Li (33), Department of Psychology, University of Richmond, Richmond, Virginia 23173 Tracy E. Love (153), Department of Psychology, University of California, San Diego, La Jolla, California 92093 Maryellen C. MacDonald (211), Department of Psychology, University of Southern California, Los Angeles, California 90089 Ivo Mimica (345), Wheeler Cognitive Neuropsychology Laboratory, Good Samaritan Hospital and Medical Center, Portland, Oregon Don C. Mitchell (313), Department of Psychology, University of Exeter, Exeter, Devon EX4 4QG, England Janet L. Nicol (113), Departments of Psychology and Linguistics, University of Arizona, Tucson, Arizona 85721 Joel Pynte (227), Centre National de la Recherche Scientifique, and University of Provence, Aix-en-Provence, France Tsutomo Sakamoto (95), Faculty of Letters, Kyushu University, Fukuoka, Japan Rosa M. Sanchez-Casas (183), Department of Psychology, Universitat Roviro i Virgili, 43007 Tarragona, Spain Christoph Scheepers (293), Department of Psychology, University of Glasgow, Glasgow G12 9YR, Scotland Lewis P. Shapiro (131), Department of Communicative Disorders, San Diego State University, San Diego, California 92182

Contributors

XV

Stan Smith (345), Wheeler Cognitive Neuropsychology Laboratory, Good Samaritan Hospital and Medical Center, Portland, Oregon Maxim I. Stamenov (167), Institute of the Bulgarian Language, Bulgarian Academy of Sciences, Sofia, Bulgaria Gerhard Strube (293), Center for Cognitive Science, Institute of Computer Science and Social Research, Center for Cognitive Science, University of Freiburg, D-79098 Freiburg, Germany David A. Swinney (153), Department of Psychology, University of California, San Diego, La Jolla, California 92093 Robert Thornton (211), Department of Psychology, University of Southern California, Los Angeles, California 90089 Matthew Walenski (95), Department of Linguistics, University of California, San Diego, La Jolla, California 92093

This page intentionally left blank

ACKNOWLEDGMENTS I greatly appreciate the support from the Alexander von Humboldt Foundation while I was a Feodor-Lynen Fellow in the Department of Psychology at the University of California, San Diego in 1994-1996. This volume is a direct outcome of this support. I am grateful to many people in the Department of Psychology for assistance during the preparation of this book, especially David Swinney. I thank all the authors; their contributions reflect both fundamentally different theoretical positions and a relatively broad crosslinguistic perspective. In considering the diversity of languages and cultures in our world, however, there is still a long way to go.

xvii

This page intentionally left blank

FROM ALEXANDER TO WILHELM VON HUMBOLDT: A CROSSLINGUISTIC PERSPECTIVE DIETER HILLERT Department of Psychology University of California, San Diego La Jolla, California

1. INTRODUCTION The present volume emphasizes the investigation of sentence processing in different languages. This crosslinguistic approach involves, to a certain extent, a cross-cultural perspective and provides us with a new look at sentence processing in general. Most psycholinguistic studies focus on a single language (English in particular) and provide predictions for a general model of human sentence processing. In comparing the results found for different languages one might be in a position to challenge the universal characteristics of the "language processor." What is enriched is the database that provides a fruitful exchange between theoretical modeling and empirical findings. This research field is related to the Lebenswerke (life-works) of both Freiherr Wilhelm von Humboldt (1767-1835) and Alexander von Humboldt (1769-1859). For Wilhelm, as well as for his Cartesian and romantic precursors, the use of language implies a creative act, and it is the "character" rather than the "form" of a language that implies value as novelty. He remains in the Cartesian framework insofar as he regards "language" as a means of thought and self-expression rather than as a functional communication system as used by animals: Die Sprache ist das bildende Organ des Gedanken. Die intellectuelle Thatigkeit, durchaus geistig, durchaus innerlich und gewissermassen spurlos vortiberSyntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

1

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

2

Dieter Hillert gehend, wird durch den Lout in der Rede dusserlich und wahrnehmbar fur die Sinne. Sie und die Sprache sind daher Eins und unzertrennbar von einander. (W. von Humboldt, 1830-35, p. 53) [The language is the organ that forms the thought. The intellectual work, quite geistig, quite inner and in a way without leaving a trace, appears superficial and perceivable for the senses by the sound in speech. They and the language are therefore at one and non-separable of each other.]

Generative grammar, in the modern sense, is a development of the Humboldtian notion of "form of language," if the latter is understood as form in the sense of "possession of knowledge" (cf. Chomsky, 1966, p. 28). With respect to language acquisition, Humboldt's Platonism is quite clear (W. von Humboldt, 1930-35, p. 58): Dass bei den Kindern nicht ein mechanisches Lernen der Sprache, sondern eine Entwicklung der Sprachkraft vorgeht, beweist auch, dass, da den hauptsdch listen menschlichen Kraften ein gewisser Zeitpunkt im Lebensalter zu ihrer Entwicklung angewiesen ist, alle Kinder unter den verschiedenartigsten Umstanden ungefahr in demselben, nur innerhalb eines kurzen Zeitraums schwankenden Alter, sprechen und verstehen. [That no mechanical learning happens in children, but a development of the "Sprachkraft" (language force), proves that, because the primary human forces rely on a certain stage of age of their development, under the various kinds of conditions all children speak and understand in about the same period of time varying little in age.]

In other words, language acquisition is a matter of cognitive capacities (competence) that grow and mature under appropriate external conditions. The "form of language" is largely programmed by the internal disposition of a genetic program (cf. Hillert, 1982). By contrast, Alexander—a naturalist, discoverer, and adventurer—lived and traveled throughout the world to find new, undiscovered cultures and languages, as documented in Kosmos. Moreover, Alexander was polyglot (Biermann and Schwarz, 1997); he had native language competence not only in German but also in French, as his friends note (Bruhns, 1872, p. 383): En effet, jamais esprit ne Jut plus francais, plus mele de sentiments generaux, exquis et de fine satire. [Indeed, no "esprit" was more French, was more mixed with generous and exquisite sentiments, and with fine satire.]

He also spoke Spanish and English fluently, studied Persian for a decade, learned Dutch, Danish, Swedish, Italian, and Arabic, had excellent knowledge of Greek and Latin, wrote in Hebrew, and knew "Plattdeutsch" (Lower German: general term for dialects spoken in northern parts of Germany). In addition, the following was reported (Stapfer, 1804, pp. 29-30):

Introduction

3

Er hat uber dreihundert Sprachen, die von einander eben so sehr als die englische vom deutschen, abweichen sollen, Bemerkungen gesammelt und glaubt berechtigt zu sein, die Anzahl der amerikanischen Sprachen von drei bis auf viertausend zu schdtzen. uber die mexikanischen Hieroglyphen undAlterttimer, uber den Ursprung (hindostanisch, wie er glaubt) und die Geschichte der Peruaner getraut er sich viel Licht zu verbreiten. [He has collected notes of more than three hundred languages, of which each should differ as much as the English differs from German, he believes to be entitled, to estimate the number of American languages from three to four thousand. He is confident to shed much light on the Mexican hieroglyphs and antiquities, on the origin (Hindustani, as he believes) and the history of the Peruvians.]

The present approach unifies fundamental aspects of the Lebenswerke of the von Humboldt brothers: On the one hand, a linguistic theory that characterizes universal cognitive features of the human language processor or the human mind and its biological source, independent of a single language structure (Wilhelm); on the other hand, a language theory that considers the diversity of linguistic structures and provides a more powerful theory of language processing (Alexander). The latter point has been less considered in designing a language processor since the existence of modern linguistics. Although not all of the present authors and I share this opinion, I believe that a methodological approach that combines a deductive and also an inductive method is essential for innovative research. Accordingly, the present volume includes research on Bulgarian, Catalan, Cantonese Chinese, Mandarin Chinese, Dutch, English, French, German, Italian, Japanese, Serbo-Croatian, and Spanish, which are contrasted in most cases with one of these languages. The range of topics refers to (a) word recognition, (b) fixed expressions, (c) grammatical constraints, (d) empty categories, and (e) parsing.

2. WORD RECOGNITION What is the format of the basic lexical unit when we comprehend a word? Morphologically related words appear to share a common representation in the mental lexicon, as shown by within- and between-language (cognates) priming effects among semantically and morphophonologically related words. Morphophonologically related words (inflections or derivations) show facilitation in a masked priming task (which seems to exclude strategic factors), while solely orthographically related or solely semantically related words do not. Thus, semantic representations as part of the mental lexicon are organized by morphological structures (see chapter by Garcia-Albea et al., this volume).

4

Dieter Hillert

A different question is how listeners access and integrate different meanings associated with identical morphophonological structures (lexical ambiguities) during on-line auditory sentence comprehension. Different processing models were suggested to account for divergent experimental findings. In particular, one might believe that structural differences among languages do not result in different findings on how words are accessed and integrated within a sentential context. Specifically, this prediction may be made for accessing noun meanings, because in most cases their isolated meanings are not modified or changed by contextual features. Pragmatic (discourse) and semantic context provides cues for the relevant meaning but no complete new meanings will be generated by the pragmaticsemantic context given. This excludes, of course, the character of language (see above), which allows the speaker to create a new meaning for an otherwise wellconventionalized word. Thus, the prediction is that no crosslinguistic effect will be measured, and that access and integration of words in a sentence works identically in all languages. This is exactly what most research on this topic seems to tell us (see chapters by Ahrens and by Borsky & Shapiro, this volume; but see Li, this volume). Opposite findings, such as context-relevant versus context-free access, seem to be due to methodological differences rather than to a crosslinguistic diversity. These differences refer to the interval used between experimental item (homophone) and probes; different recognition points of the probes; the kind of probes used (words versus pseudowords, or context-related versus homophone-related); the location of probes (as part of the experimental sentence itself versus as part of a secondary task); frequency of the homophone's meanings; salience of contextual features; the type of lexical decision task used (yes-no-response versus yesnonresponse); or speed of sentence presentation. All of these different factors can contribute to the variance obtained in on-line experiments on noun ambiguity resolution. But what seems to be sure is that despite a biasing context a multiple access to both homophone meanings can be found. If both meanings are equally frequent, a cross-modal priming paradigm has been applied using factors such as "yes-no-response," a zero-interval between experimental word and probe (0-interstimulus interval), and probes that are directly related to the homophone's meanings. This factor constellation seems to make it possible to measure the first "parse," which is exhaustive. This is confirmed by experiments testing access to verb homophones in English and access to ambiguous verb stems (discontinuous words) in German (see my chapter, this volume). Although particle information presented in cleft sentences constrains access to a single-verb meaning, an exhaustive access to both verb meanings is found (see experiment 1). Again, support for this modular model is found with discontinuous words; that is, meanings of German verb particle combinations are exhaustively accessed at the offset of the verb stem in a biasing context, although the particle occurs at the final position of

Introduction

5

the sentence (experiment 2). Thus, the listener takes into account possible morphologically complex verb forms at the verb stem, although the meaning of the stem and of the verb particle combinations are completely different. The influence of other factors appears to prevent a first parse measurement, and responses seem to reflect processes due to the contextual integration of the relevant meaning. Even if a selective effect is measured when context strongly biases the most frequent meaning of the homophone by salient features, it is questionable whether the less frequent meaning has not been activated at all. Such a weak activation may not be detectable by current experimental methods. Another influencing factor may be related to the listener himself, without challenging the modular access theory. Do good listeners use less context than bad listeners? A modular model of lexical access during sentence processing certainly goes beyond the examination of novel (newly composed) word meanings. These computations may be quite comparable to those involved in accessing words in a second language: they require the use of explicit context information. However, there is no doubt that further crosslinguistic investigations on lexical ambiguity resolutions are necessary to reconsider the picture sketched here. The investigation of discourse integration of ambiguous words goes beyond automatic processing of lexical access during on-line sentence comprehension. The question is whether listeners or readers apply universal strategies at a postlexical stage during off-line sentence or discourse understanding when modular access is followed by context-relevant integration processes. As found in Japanese, discourse understanding (macroprocesses) appears to operate quite independently of lexical comprehension (microprocesses), even during metalinguistic (off-line) operations (see Hatano & Kuhara-Kojiima, this volume). Also, the crosslinguistic comparison of the off-line comprehension of spatial terms such as dimensional prepositions reveals systematic response patterns at the conceptual-pragmatic level. Obviously, they depend on the features of the reference object itself, the kind of discourse context, and the relationships between concepts and word forms (see Herrmann & Grabowski, this volume).

3. FIXED EXPRESSIONS A topic that is closely related to lexical access of literal meaning representations is the investigation of figurative speech, such as idioms or metaphors during sentence processing (see Colombo, this volume, and Cacciari, this volume). Although the amount of on-line data increases continuously, and more and more studies are taking into account crosslinguistic comparisons, this research field remains a poor relative to the investigation of literal speech. As Jackendoff (1997, p. 156) points

6

Dieter Hillert

out, fixed expressions and idioms are hardly a marginal part of our use of language (cf. his "Wheel of Fortune" example): A speaker of English must be carrying around thousands of compounds, idioms, names, meaningful names, and cliches, and at least hundreds of titles and quotations that could be potential Wheel of Fortune puzzles. As for names, think of all the names of people you know that could not be Wheel of Fortune puzzles because they're not famous enough—Noam Chomsky, Otto Jespersen, all your family, colleagues, neighbors, and odd classmates—thousands more. As for titles and quotations, think of all the poetry you know, including lyrics of popular songs, folk songs, and nursery rhymes, plus advertising slogans. There are vast numbers of such memorized fixed expressions; these extremely crude estimates suggest that their number is about the same order of magnitude as the single words of the vocabulary.

From a psycholinguistic viewpoint, the difference between an idiom and its literal counterpart relies on the strength of association between the words. Thus, the idiom "to kick the bucket" is treated as lexical verb with internal structure (la) in contrast to the literal analysis (Ib). (1) a. [V [VP [V kick] [NP [Det the] [N bucket]]]] b. [VP [V kick]] [NP [Det the] [N bucket]] Similar to the examination of (literal) word recognition is the question of whether sentential context constrains access to the figurative or the literal meaning. Methodological difficulties, such as the question of to what extent a self-paced reading task can be regarded as an on-line task, lead often to controversial issues. Other factors appear to be highly relevant for the interpretation of how idioms are accessed during on-line sentence processing; for example, the (de)compositional degree of an idiom (semantically or/and syntactically), the idiom key (e.g., sentence context or idiom recognition point of a lexical sequence: the point at which an idiomatic interpretation is very likely), or the frequency of the idiomatic and literal meaning. Overall, the data seem to point to a processing model in which literal activation is activated first, if no idiom key is given. Thus, figurative meanings seem to have another status during sentence processing, although presumably organized identically, such as nonfixed (literal) meanings. However, access to idioms is a process that can be compared at best with access to literal meanings when the lexical form consists of a comparable sequence of segments. Thus, the examination of single words having either an idiomatic-literal or a literal-literal meaning (and having equivalent frequencies) would be the appropriate test for deciding whether idiomatic lexical meanings are exhaustively accessed as literal lexical meanings. Because of space limitations, other chapters on figurative speech were

Introduction

7

not included in the present volume, but a crosslinguistic approach on this topic may be as challenging as it is for modeling a universal language processor of literal speech. Taking into account the overwhelming role that fixed expressions play in our everyday linguistic processing, a model of the human language processor that does not consider figurative language must remain inadequate.

4. GRAMMATICAL CONSTRAINTS To what extent the recognition and production of words is facilitated by grammatical features (inherent or overt) during sentence processing is the general topic of a large number of studies. In English only a weak or nonexisting grammatical effect in terms of priming or inhibitions was measured. In contrast, the investigation of a morphologically richer language system revealed effects in accessing a content word in sentence context. However, one reason for the dispute whether or to what extent (congruent) grammatical properties facilitate access to lexical recognition/retrieval in a sentential context appears to be the diversity of methodological approaches used. From the present point of view, tasks such as unimodal sentence continuation, gating, and self-paced reading, for example, cannot be regarded as being truly online compared to the cross-modal paradigm. The former experimental methods appear to be sensitive to processes involved in lexical integration and not to the lexical activation patterns initiated immediately on lexical access; in contrast, the latter experimental methods appear to be sensitive to activation patterns (if they exist) independent of sentence and pragmatic contexts (see Nicol, this volume). Thus, the dispute about whether congruent context facilitates or whether incongruent context inhibits access to a word (in production or comprehension) appears to be a pseudocontroversy. Lexical integration tasks examine the validity of a grammatical cue for the selection of certain lexical entries in a given context. The crosslinguistic comparison allows one to evaluate the role of a different grammatical property in a certain language-specific and sentence-specific context during the selection process (see Devescovi et al., this volume, and Li, this volume). This approach is highly relevant for the investigation of global and local context effects on lexical processing. A short-term goal may be to receive a complete picture of lexical activation patterns during sentence processing in different languages; the long-term goal may be, however, to describe the interactions (modular or nonmodular) between different levels of linguistic representations that are universal in character. This requires, however, a theoretical framework that allows the prediction of the temporal course

8

Dieter Hillert

of activation patterns for all languages on the relevant levels of representations within the human sentence processor.

5. EMPTY CATEGORIES The importance of linguistic theories for the experimental investigation of sentence processing has been demonstrated by coreferential assignment during online processing: noncanonical sentence structures containing moved constituents that leave behind a trace of that movement (gap). To understand sentences with moved constituents the listeners must fill the gap with the relevant antecedent. That such trace-chains are simply not linguistic constructions but computed during on-line processing has been shown in a series of studies. For example, in using a cross-modal lexical priming (CMLP) task, objectrelative sentences strongly biased toward a single meaning of an ambiguous object: "The professor insisted that the exam be completed in ink, so Jimmy used the new pen #1 that his mother-in-law recently #2 purchased #3 i because the multiple colors allowed for more creativity" (see Love & Swinney, this volume). In examining the on-line activation of the ambiguous meanings by presenting visually at all three test points (#) related and control words for both noun object meanings, it was found that priming occurred only at test point 1 for both object meanings. In contrast, at testpoint 2, no priming was found, and at test point 3, only the context-relevant meaning was facilitated. The question arises where and how does the listener fill the gap, although at the pregap control test point (#2) no semantic activation of the antecedent has been measured? A series of studies revealed that coreference assignment (gap filling) during on-line sentence processing occurs by the reactivation of the antecedent. This speaks for a deep structural or conceptual format that constrains access to the context-relevant meaning of the antecedent. A slightly different response pattern was found in Bulgarian (see chapter by Stamenov & Andonova this volume). Whether this effect is a matter of free word order or of morphological markers (or a combination of both), remains an issue for further investigation with nonconfigurational languages (in contrast to English). Another type of empty category is PRO, at which the verb assigns an argument (subject or object). In contrast to Frazier, Clifton, and Randall's (1983) conclusion that subject-control sentences are easier to process than object-control sentences, the comparison between English and Japanese appears to argue for the inverse pattern (see Sakamoto & Walenski, this volume). Moreover, case markers seem to influence the Japanese parser's initial decisions (but see below). Again, a third type of empty category is pro (null pronominal in case position). The parser seems to work after the "minimal chain principle," that is, in the first step an element

Introduction

9

directly receives a thematic role in the deep-structure position (see De Vincenzi, this volume).

6. PARSING The human sentence-processing parser may use different kinds of linguistic information to take into account different kinds of cues available at a certain point during language comprehension. Parsing structural ambiguities can be influenced by lexical factors, such as verb subcategorization preferences (see Borsky and Shapiro, this volume), and syntactic factors such as minimal attachment or minimal chain principle (see De Vincenzi, this volume, and Hemforth et al., this volume). Again, a multiple or delayed parsing model predicts that a number of different information sources (e.g., lexical, syntactic, or pragmatic preferences) are consulted before the processor commits to a certain structural analysis (see Hirose & Inoue, this volume, and Dillert, this volume: experiment 3 and 4). Alternatively, it has been proposed that the parser analyzes sentential input solely on the basis of frequency distributions (lexical or structural) and with pragmatic constraints being top-down driven (see Thornton et al., this volume). Accordingly, research strategies are quite different in the attempt to model how sentences are parsed (see Mitchell & Brysbaert, this volume). In the attempt to design a universal human parser, the focus on general linguistic principles taking into account crosslinguistic variations is highly relevant. Otherwise, having as its goal describing the variation caused by differences among individuals, subject groups, or languages, parsing might be the product of a number of completely different factors. Overall, the examination of parsing has focused mainly on different reading techniques (e.g., self-paced or eye tracking). Thus, it is important to run more studies on auditory sentence parsing, which includes the full range of linguistic information available to a native speaker.

7. CONCLUSION This volume provides an overview of central psycholinguistic topics on sentence processing from a crosslinguistic perspective. My hope is that it motivates further studies on an even broader variety of languages. However, as pointed out before, it is important from a methodological point of view to work with a theoretical framework developed in focusing necessarily on the analysis of a single language. We simply live with different languages (culture organisms), and their analysis is ground for a never-ending controversy among theorists following a

10

Dieter Hillert

variety of different paradigms. Languages and theories both have in common that they express our thoughts; how we do this and how we describe how we do this changes continuously. Both changes are hopefully quite different: Languages change quantitatively, but theories change (hopefully) qualitatively. Therefore, the symbiosis between linguistic variety and theory, which reflects the Lebenswerke of Alexander and Wilhelm von Humboldt, may provide the best approach to investigate universal principles of the human sentence processor. I conclude with the words of Wilhelm von Humboldt (1830-35, p. 59): Dass der Mensch iiberall Eins mit dem Menschen ist, und die Entwicklung des Sprachvermogens daher mit Hulfe jedes gegebenen Individuum vor sich gehen kann. [That everywhere the human is at one with the human, and the development of the language ability can therefore take place with the support of each individual that exists.]

ACKNOWLEDGMENTS Supported by a Feodor-Lynen fellowship of the Alexander von Humboldt Foundation (Hi V.3-FLF). I am grateful to Vicky Bouck and Eileen Favorite for editorial comments.

REFERENCES Bruhns, K. (1872). Alexander von Humboldt. Eine wissenschaftliche Biographie. Vol. 2. Leipzig. Bierman, K-R., and Schwarz, I. (1997). Der polyglotte Alexander von Humboldt. Alexander von Humboldt Stiftung, Mitteilungen. AvH-Magazin Nr. 69, 39-44. Chomsky, N. (1966). Cartesian linguistics. A chapter in the history of rationalist thought. New York and London: Harper and Row Frazier, L., Clifton, C., and Randall, J. (1983). Filling gaps: Decision principles and structure in sentence comprehension. Cognition 13, 187-222. Hillert, D. (1982). Zur mentalistischen Sprachtheorie Noam Chomkys und ihre Bewertung durch psycho- und neurolinguistische Evidenz. Unpublished master's thesis, Johann Wolfgang Goethe-Universitat, Frankfurt am Main. Humboldt, Wilhelm von (1830-35). Uber die Verschiedenheit des Menschlichen Sprachbaues. In A. Leitzmann (Ed.), Wilhelm von Humboldts Werke. 7th Vol., 1907. Berlin: B. Behr's Verlag. (facsimile edition 1968: Berlin, Walter de Gruyter and Co). Jackendoff, R. (1997). The architecture of the language faculty. Cambridge, MA: MIT Press. Stapfer, P. A. (1804). Brief an Paul Usteri. In H. Beck (1959), Gesprdche Alexander von Humboldts. Berlin.

LEXICAL AMBIGUITY RESOLUTION: LANGUAGES, TASKS, AND TIMING KATHLEEN V. AHRENS Department of Foreign Languages and Literatures National Taiwan University Taipei, Taiwan

1. INTRODUCTION The current work explores language, task, and timing issues in lexical ambiguity resolution. Three hypotheses are presented concerning the different results previously found for ambiguity resolution in English and Italian. The first hypothesis (language-driven hypothesis) is that the difference is language based. A strong version of this hypothesis (i.e., that context-prominent languages allow context to select the appropriate meaning of an ambiguous word) is tested with respect to lexical ambiguity resolution in Mandarin Chinese, and ruled out. A second hypothesis (speeded-selection hypothesis) is that context-prominent languages use context faster once lexical access has occurred. The third hypothesis concerns methodological issues, especially as pertains to the length of time that the visual target appears. I conclude by postulating that the key reason for the discrepancies in lexical ambiguity resolution to date are not because of crosslinguistic differences, but instead have to do with timing differences, and I suggest that experiments that adequately measure immediate, automatic lexical retrieval will demonstrate that the modularity hypothesis holds. Language thrives on ambiguity, at both the word and sentence level. Lexical ambiguity in one way simplifies language because it allows for a language to expand its range of meanings without having to continuously add new phonological forms to its already long list of lexical items. However, lexical ambiguity at the Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

11

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

12

Kathleen V. Ahrens

same time complicates language, because the processor must, upon encountering a phonological word with multiple meanings, determine which meaning is correct. Because people rarely notice that lexical ambiguities interfere with their comprehension of a speaker's utterance, lexical ambiguity resolution must occur before a listener consciously notices it—that is, it must occur very, very quickly. There are two basic theoretical models of lexical ambiguity resolution: context-dependent models and context-independent models. A context-dependent account postulates that the context that precedes an ambiguous word will aid the processor in accessing only the contextually appropriate meaning, and the contextually inappropriate meaning will not be accessed. A context-independent account says that the preceding context will not be able to influence which meaning is accessed at an ambiguous word. The context-dependent model is compatible with a languageprocessing account that says that information among different language subsystems (such as lexical access and discoursal knowledge) can interact in ongoing language processing at any point in time (e.g., McClelland, 1987). The contextindependent model is compatible with a modular language-processing account (Fodor, 1983). Under a modular view of language processing, processing must be completed in one language subsystem (such as lexical access) before information from another language subsystem (such as discoursal knowledge) can influence the results. Thus, the implications for lexical ambiguity resolution extend not only to the nature of the language-processing system, but also to the makeup of the underlying cognitive architecture that support language. In this chapter I first discuss the evidence that pertains to each model, and then point out that lexical ambiguity resolution must be looked at within the paradigms of natural speech comprehension and reading comprehension separately in order to better understand the nature of lexical access. Next I point out that the findings in lexical ambiguity resolution differ along language lines, and postulate a hypothesis that accounts for this fact. I then test this hypothesis, and at the same time test the context-dependent hypothesis, and find that neither hypothesis holds. I revise the original language-driven hypothesis, and also suggest other methodological reasons for the discrepancies in the English and Italian results.

2. CONTEXT-DEPENDENT AND CONTEXT-INDEPENDENT MODELS I distinguish and discuss two types of context-dependent models: a "strictly" selective account and a "modified" selective account. A strictly selective account says that context alone is enough for the processor to select the appropriate meaning of the ambiguous word (e.g., Glucksberg, Kreuz, and Rho, 1986; Simpson, 1981). This means that if the context is biased towards the dominant meaning of a word, then only the dominant meaning is accessed. Moreover, if the context is biased toward the subordinate meaning of a word, then only the subordinate mean-

Lexical Ambiguity Resolution

13

ing is accessed. In effect, context can block the access of the noncontextually appropriate meaning, and select only the meaning that is contextually appropriate. A modified selective account says that when context and dominance (frequency of meaning of the lexical item) interact, only the contextually appropriate meaning of a word is selected (e.g., Tabossi, Columbo, and Job, 1987; Tabossi, 1988; Tabossi and Zardon, 1993). The prediction of the modified selective account is as follows: When the contextually congruent meaning is also the dominant one, the subordinate meaning ... receives some initial activation from the perceptual input, but this may never become strong enough to be detected by current experimental methodologies. Instead, when context constrains the subordinate meaning, the activation received from ... the dominant meaning is still sufficient to render it detectable, so that both meanings will result [in being] reliably active. (Tabossi, 1993:369)

This account "conceptualizes access as a continuous event which is accomplished when semantic information about a word is sufficiently activated to be integrated, when appropriate into prior discourse" (Tabossi, 1993:369-370). There are two types of context-independent models: one version is known as an exhaustive or a multiple-access model. This model predicts that all meanings will be immediately accessed regardless of the sentence's contextual bias, or the relative frequency ranking of the meanings (e.g., Swinney, 1979; Onifer and Swinney, 1981). This finding holds irrespective of whether the contextual bias is toward the primary or secondary meaning of the word. However, after lexical access has occurred, and by 1500 ms downstream, only the contextually appropriate meaning is left available (e.g., Swinney, 1979; Tanenhaus et al., 1979; Onifer and Swinney, 1981; Swinney and Love, 1996). The interpretation is the following: "Lexical access appears to be an exhaustive and autonomous subroutine of the sentence comprehension process (autonomous in the sense that it does not appear to be driven or guided by previously occurring semantic information)" (Onifer and Swinney, 1981, p. 232). Another version of the context-independent model postulates that the order of access of word meanings is related to their frequency ranking (Hogaboam and Perfetti, 1975). That is, regardless of the contextually appropriate meaning, the most common meaning will first be retrieved and tested for its compatibility with the context. If it is found to be incompatible, then the next meaning is retrieved and tested, until a match is found. An ordered access model predicts that if the context is biased toward the most frequent meaning, that will be the only meaning that is retrieved, but if the context is biased towards a less frequent meaning, then the contextually biased meaning and all more frequent meanings will be retrieved. Note that this model makes the same predictions as the modified selectiveaccess account does for words with two possible meanings, although the models on which they are based are completely different. In both models, if the context

14

Kathleen V. Ahrens

is biased toward the dominant meaning, the dominant meaning is activated, and if the context is biased towards the subordinate meaning, then both the dominant and subordinate meanings are activated. In cases where there are more than two meanings, it is possible that the modified selective access account will again make the same prediction as the ordered access account; namely, if any subordinate meaning is picked, all more dominant meanings will also be accessed. However, the modified selective access account could also predict that only the contextually appropriate subordinate meaning (e.g., M4, the fourth meaning), and the most dominant meaning (e.g., Ml, the first meaning) will be accessed. Given Tabossi and Zardon's definition above, the question revolves around whether or not the more dominant (but not contextually appropriate subordinate meanings (e.g., M2 and M3)) have enough activation to be detectable. This conceptualization of the modified selective access could potentially differentiate it from the ordered access model, because an ordered access model does not concern itself with level of activation (although it could easily incorporate this concept). However, what is crucial for the discussion at hand is that the modified selective access model is (under either conceptualization) still a context-independent model (although Tabossi and colleagues originally postulated otherwise), because context does not select a meaning. Instead, all meanings are activated (i.e., "When the contextually congruent meaning is also the dominant one, the subordinate meaning . . . receives some initial activation from the perceptual input, but this may never be strong enough to be detected by current experimental methodologies"—Tabossi, 1993:369). Whether or not the activation can be detected is a separate, methodological issue. The theoretical issue is whether preceding context does or does not select a single, appropriate meaning. Since the modified selective access account does not, and since it postulates activation of all meanings, it is an account that exists within the context-independent model.

3. NATURAL SPEECH COMPREHENSION In the preceding section, I argued that the modified selective access account should be viewed as being within the context-independent model. Even so, there is still evidence for the context-dependent view. In fact, the question as to why evidence can be found for both models has perplexed researchers working on lexical ambiguity resolution for the past two decades. Simpson (1994) reviews the relevant empirical studies on lexical ambiguity resolution and concludes that there are no clear methodological reasons as to why different studies have supported different models. The methodological variations he considered include (a) method of presentation of the context (either auditory

Lexical Ambiguity Resolution

15

or visual); (b) type of task (such as lexical decision or naming or Stroop); and (c) location of the ambiguity within the context sentence (sentence-medially or sentence-finally). However, it could be the case that the question he is attempting to answer, given in (1) below, is too broad to allow a unified explanation. (1)

"To what extent do higher level semantic representations, such as those arising from the processing of a sentence, constrain the activity of a lower process (in this case the identification of a word)?" (Simpson, 1994:359)

If instead we begin with a different and narrower question, as stated in (2), and look at the studies that pertain to this question, we will find that the reasons for the differences can be explained. (2) To what extent do higher level semantic representations constrain the activity of a lower process (in this case the identification of a word) in natural speech comprehension? Question (1) refers to any type of comprehension process (i.e., reading, listening, or signing comprehension). However, it is not clear that it is appropriate to conflate the different modalities, because (a) there are different areas in the brain that deal with auditory and visual stimuli; (b) speech (or sign language) is learned earlier than reading; and (c) that speech (or sign language) is the basis for learning how to read. Under these assumptions the time course of lexical ambiguity resolution in reading comprehension is a different question from that of lexical ambiguity resolution in speech comprehension and should be examined separately.1,2 When we narrow our focus to lexical ambiguity resolution in speech comprehension, we naturally need to look at cross-modal experiments. Cross-modal priming experiments involve subjects listening to a sentence (spoken by a native speaker at normal speed) over headphones and making either a lexical decision to a visual target that appears on the screen (cross-modal lexical decision task) or naming the visual target (cross-modal naming task). The visual target usually appears at the offset of the ambiguous word, but there is no pause in the sentence that the subject is hearing—she or he hears a complete and natural-sounding sentence. This type of technique is advantageous to studying the time course of language processing (a) because subjects are listening to natural speech as they would in nonexperimental settings; (b) because the presentation of the visual target can be precisely controlled in order to see the time course of ambiguity resolution; and (c) because the task (either naming or lexical decision to the visual target) does not require conscious reflection on the nature of possible relationship between the information they are hearing over the headphones (i.e., the ambiguous word) and the visual target (i.e., a semantic prime or a control word) (Nicol, Fodor, and Swinney, 1994; Swinney and Love, 1996).

16

Kathleen V. Ahrens

When we concentrate on question (2) and review the work on cross-modal lexical ambiguity resolution experiments, we find that there are, indeed, methodological reasons as to why some researchers report findings for the strictly selective access model (Simpson, 1981; Glucksberg et al., 1986), the modified selective access model (Tabossi et al., 1987; Tabossi, 1988; Tabossi and Zardon, 1993), as well as the multiple access model (Swinney, 1979; Tanenhaus et al., 1979; Onifer and Swinney, 1981; Seidenberg et al., 1982; Swinney and Love, 1996). The areas we will look at include (a) position of visual target; (b) experimental task; and (c) language. The position of the visual target in relation to the ambiguity is of paradigmatic importance in understanding the time course of ambiguity resolution. As a number of researchers have pointed out (McClelland, 1987; Onifer and Swinney, 1981; Simpson, 1994), intervals that occur between the offset of the ambiguous word and the presentation of the visual target mean that the activation is not indicative of immediate processing. One cross-modal study that found evidence of strictly selective access (Simpson, 1981) presented the visual target with a delay from the offset of the ambiguity at 120 ms. These findings might, in fact, represent the effects of context once the word has been accessed, and cannot be construed as supporting a model of selective access. The other cross-modal experiment found for strictly selective access was a cross-modal interference task with pseudowords. In Glucksberg et al.'s (1986) experiment, subjects were presented the visual target at the offset (0 ms) of the ambiguity, but the visual targets of the ambiguous words were all pseudowords. The idea was that the subjects would be slower to reject them as nonwords, because the pseudowords would remind subjects of related words. However, it is not clear that this interference task was measuring processing relating to ambiguity resolution, nor is it clear that the interference task allowed a decision that was within the time window of lexical access. (Please see Prather and Swinney (1988) for a discussion of these and related issues.) More importantly for our discussion here, an interference task with pseudowords has not been used in any other cross-modal lexical decision task for ambiguity resolution, and thus it is difficult to ascertain if the finding for selective access has to do with the fact that it was a different type of task (i.e., an interference task with pseudowords). Thus, the cross-modal on-line sentence-processing data in support of a strictly selective account is limited, and involve timing and task issues that make it difficult to ascertain if the findings were artifacts of these issues. Because findings for a strictly selective account would be the clearest indicator that language processing is an interactive (as opposed to modular) process, we will test this hypothesis in Mandarin Chinese. The reason is that, as we will explain in the next section, Mandarin is a language that relies heavily on contextual information in order to arrive at linguistic interpretation, and is most likely to support a strictly selective account.

Lexical Ambiguity Resolution

17

I turn now to the three works that find for a modified selective access account in a cross-modal priming paradigm, which are all in Italian (Tabossi et al., 1987; Tabossi, 1988; Tabossi and Zardon, 1993). When I compare these three experiments, which all find that context and dominance can influence lexical access, with experiments that have found that context and dominance do not influence access (Swinney, 1979; Tanenhaus et al., 1979; Onifer and Swinney, 1981; Seidenberg et al., 1982; Swinney and Love, 1996), I find that they are similar on several methodological counts: they all are cross-modal tasks (either naming or lexical decision); they all have the visual target appear immediately at the offset of the ambiguous word; and they are all priming tasks (as opposed to the interference task with pseudowords that Glucksberg et al., 1986, used). In the Italian experiments, the ambiguities occur sentence-medially, whereas in the English experiments the ambiguities occur either sentence-medially or in the sentence-final position. The main methodological differences between the two groups of experiments are the following: first, the languages in the experiment are different; second, the lexical decision tasks are different (i.e., in the Italian case a go/no-go paradigm is used. This means instead of the usual task of pressing one button if the visual target is a word, and another button if it is a nonword, subjects only press a button if the visual target is a word, and do nothing otherwise.) Third, the presentation of the visual target in the Italian experiments is 1500 ms, whereas in the English experiments, the presentation of the visual target is never more than 1000 ms. In what follows I will first explore the hypothesis that the difference in findings derives from the different linguistic properties of English and Italian. I will present an experiment that tests the strongest form of this hypothesis, and show that it does not hold. This experiment will, at the same time, test for the strictly selective access account. To foreshadow the results, the evidence will demonstrate that neither hypothesis is correct. In the general discussion section I will return to the two methodological possibilities as alternatives for the differences in the findings.

4. LANGUAGE-DRIVEN DIFFERENCES If it is the different linguistic properties of English and Italian that are driving the difference in lexical ambiguity resolution results, this would lead to a nearWhorfian conclusion that native languages shape how we process that language, and that there is no universal language processor. This would imply that the underlying cognitive wiring for language processing in Italian and English is very different. What could be the reason that Italian allows a combination of context and dominance to influence lexical access, while English does not? English has an impoverished agreement system, which means that it relies more heavily on structural information for semantic interpretation (i.e., in English the

18

Kathleen V. Ahrens

subject position is often the agent of the action, and the object position is often the patient of the action). In addition, English does not allow the omission of either the subject or the object. Italian has a richer agreement system, which means that the referents are often determined by the agreement markers on verbs and not necessarily by an overt noun. Italian also allows the dropping of the subject to occur. Thus, Italian is a language that requires more information to be recovered from context than English, because in Italian the referents can be missing, whereas English relies on overt referents that have a reliable syntactic-semantic association. A language-driven hypothesis would suggest that because contextual information is crucial to interpreting the incoming string of information, it is also available and able to influence lexical access. The language-driven hypothesis is formulated as in (3). (3) Language-driven hypothesis: If a language is context-prominent, then context will be able to influence lexical access immediately and automatically. A language is defined as being context-prominent if contextual information plays a prominent role in interpreting a sentence. For example, languages that allow dropping of the subject, or of the object, need context in order to interpret the missing information. One way to test the language-driven hypothesis is to look at ambiguity resolution in a language that relies even more heavily on contextual information than Italian. Mandarin Chinese is one such language.3 Mandarin Chinese is like English in that it has an impoverished agreement system, but more like Italian in that it is a prodrop language. In fact, in Mandarin both the subject and object can be omitted in situations where context allows the information to be reconstructed. For example, if someone asks, "Do you like ice cream?" or "Do you like the ice cream I bought home from the store yesterday?", one can respond to either question with a simple "Like." In Mandarin, the previous sentence is enough to "fill in" for both the speaker and the hearer who is doing the liking, and what is liked. Moreover, even though in Mandarin subjects and objects are structurally encoded as in English, when topicalization occurs, a subject and object can switch their surface position without changing the dominant interpretation, as shown in (4) and (5) (Huang, 1985). (4) The zhong cong, niao chi de hen duo. This type bug, bird eat DE very many 'Birds eat a lot of this type of bug.' (5) Zhe zhong niao, cong chi de hen duo. This type bird, bug eat DE very many 'This type of bird eats a lot of bugs.' In (4) and (5), both participants occur preverbally, even though in reversed order. The dominant interpretation of both sentences is semantically equivalent,

Lexical Ambiguity Resolution

19

with 'cong' (bug) interpreted as the preposed object in each case. As for 'riao' (bird), it occurs in the canonical subject position in (4), and is interpreted as the preposed subject in (5). Thus, in Mandarin structural information alone does not reliably determine the semantics of the proposition.4 From the above discussion we see that Mandarin Chinese relies on contextual information to assign interpretations, instead of structural information that may either be missing or unreliable. In this respect it is even more contextdependent than Italian, which is already more context-dependent than English. If the language-driven hypothesis is correct, the reason that Tabossi and colleagues found context and dominance influenced ambiguity resolution in Italian could be because Italian is more dependent on contextual information for prepositional interpretation to take place. Furthermore, if the language-driven hypothesis is viable then context should influence ambiguity resolution in Mandarin Chinese because Mandarin relies even more heavily than Italian on contextual information for semantic and propositional interpretation. The following experiment will test this hypothesis in its strongest form because we will look at the effect of context alone. I will test for the activation of both meanings at the onset position when the sentential context is biased toward the secondary (subordinate) meaning of the ambiguity. I select the onset position in order to rule out the possibility that a meaning was accessed but then faded. I bias toward the secondary meaning because this is the strongest test for a hypothesis, which postulates that top-down information can influence lexical access (e.g., the strictly selective access account). If I find that only the contextually appropriate meaning is accessed, then the language-driven hypothesis will become more plausible. It will also be the first time that evidence for a strictly selective account is found in an on-line cross-modal priming paradigm. If this is indeed the case, it will then be important to look at the position of the visual probe point in order to make sure that the lexical item was recognized. If I find both meanings accessed, it will mean that the strong language-driven top-down hypothesis did not hold, and that the strictly selective access account once again did not receive empirical support from an on-line sentence comprehension paradigm. It will also mean that the onset position was not too early to see an activation from either the contextually appropriate or contextually inappropriate meaning.

5. EXPERIMENT: LEXICAL AMBIGUITY RESOLUTION IN MANDARIN CHINESE 5.1. Subjects All subjects were native speakers of Mandarin Chinese, had lived in Taiwan since birth, and were undergraduate or graduate (M.A.) students of National Chung Cheng University (NCUU), in Chia-yi, Taiwan. In Taiwan, there is usually

20

Kathleen V. Ahrens

another language spoken in the home in addition to Mandarin, such as Taiwanese, Hakka, or an Austronesian language. Therefore, in order to keep the subject pool as homogeneous as possible in terms of language background, all subjects used in the following experiments were exposed not only to Mandarin but also to Taiwanese in the home before the age of 7. They were also screened for any brain injury, learning disability, or other abnormal mental behavior. All subjects were paid for their participation. The actual number of subjects used in each pretest or experiment are given in each of the appropriate sections below. 5.2. Pretest for Ambiguity Bias Thirty NCCU undergraduates were presented with 90 words (52 nouns, and 38 verbs) that had at least two meanings. Presentation was auditory, because words often have different auditory and visual biases. For each word, subjects were asked to provide the first meaning they could think of, and then they were asked to provide another meaning if they could think of one. Tallies of the numbers of first and second choice for each meaning of the words were made. For the present study 16 nouns were chosen that had a preference for a primary interpretation (i.e., the chosen primary meaning had to have been listed as the first-choice meaning for speakers over 75% of the time); additionally, the secondary meaning had to have been listed as the second-choice meaning for speakers over 60% of the time. 5.3. Pretest—A Priori Equated Reaction Times for Experimental and Control Probes Probes for the Primary and Secondary meaning of the ambiguity were chosen by using one of the three most frequently provided associates, with the restriction that if these associates were not disyllabic, a disyllabic word closely related to the most frequent associate was used. A large number of words equated to the "related" associates on the basis of frequency, length, and form class were included with the experimental/related associate words in an isolated lexical decision task. Twenty subjects participated in this experiment. After the average for each word was calculated, a "matched" control word was chosen for each "related" associate for each meaning of the ambiguity. Overall, the mean reaction times for the primary experimental and control condition was 483 ms and for the secondary experimental and control conditions was 467 ms. In addition, the experimental and control conditions were matched across individual items for syllable length, and the frequency of the control conditions was always higher than for the experimental conditions overall, which works against my hypothesis (since higher frequency words are more likely to be accessed more quickly). The frequency data was based on written norms (CKIP, 1993).

Lexical Ambiguity Resolution

21

5.4. Creation and Pretest of Sentential Materials Each ambiguous word was placed in a sentence biased toward its secondary meaning. Subjects were given the actual experimental sentence up to the point just before the ambiguity and asked to complete the sentence. Following Tabossi's criteria for creating a bias towards a "strong" aspect of one meaning of the ambiguity, at least 75% of 21 subjects agreed on the intended aspect of the contextual bias. An example is given in (6a) below. The experimental and control visual targets for the primary and secondary meanings are given in (6b). (6) a. tonghua gushi li de wangzi ru yao yingqiu bei kun zai chengbao fairy tale story in DE prince if want win BEI confine in castle nei de gongchu shi ta bishu guo chongchong de jiguan cai neng in DE princess when he must pass multiple-level DE trap then can cong ermo shou zhong qiu hui ta de xinshangren, yushi congci from devil hand middle rescue back he DE loved-one then from yihou guo zhe xinfuquaile de rezi. after live particle happy DE days 'In a fairy tale, when a prince wants to rescue a princess that is locked up in a castle, he must overcome serious traps in order to rescue the one he loves from evil, so that the two of them can live happily ever after.' b. ambiguous word: jiguan primary meaning (institution) secondary meaning (trap)

Experimental visual target xingchen 'administration' xianjin 'trap'

Control Visual target xiaoshi 'hour' shumian 'report in writing'

The entire experiment consisted of 42 sentences: 16 experimental sentences and 26 filler sentences. All sentences were recorded by a female speaker to the hard drive of an IBM-compatible Pentium computer with the aid of the Creative Wave sound card using the Creative Wave program. The time from the beginning of the sentence to the onset of the ambiguity was measured using the same program. This information was then entered into a control list that associated the time of presentation along with each respective sentence and visual target. At the time of the onset of the ambiguity, a visual target was flashed on the screen for 300 ms. An internal dedicated CPU in the button box measured the time from the presentation of the visual target until a response was made on the button box or 2000 ms had passed, whichever was earlier. The program controlled for timing problems in video display, such as raster position. The measurements of the dedicated CPU are accurate to the thousandths of a millisecond (.001 ms). The measurements themselves were made to the nearest millisecond. The sentences occurred in random order, and there was a 5-sec delay between sentences.

22

Kathleen V. Ahrens

5.5. Procedure Subjects sat in front of a computer monitor in a sound-attenuated booth and were told that they had two tasks. The first task was to listen and understand the sentences that they heard. They were told that they would be tested on the comprehension of the sentences at the end of the experiment. They were also told that they had a second task—to watch the computer screen and when they saw characters appear on the screen to decide if the characters made up a word or not. They were told to press the right-hand button (marked "word") as fast as possible if they thought it was a word, and to press the left-hand button (marked "nonword") if they did not think it was a word. They were asked to keep their fingers directly on the buttons at all times. The auditory sentence continued on without interruption even when the visual probe was being presented and the subjects were making their decision. Each subject's attention was focused on the middle of the screen by means of a black mask that allowed only a small word-sized area of the screen to be visible. At the end of the cross-modal portion of the experiment, the subjects were given 10 sentences printed on a sheet of paper and asked to mark which sentences they had just heard. 5.6. Results Subjects were screened for errors on an individual basis. Subjects with errors greater than 15% on the lexical decision task (i.e., more than six wrong) and those with errors above 20% on the comprehension test (i.e., more than two wrong) were dropped from the analysis. After screening for errors, there were 48 subjects with reliable data. Table 1 presents the mean reaction times of correct responses in each condition. The overall variance of data was run on individual subjects data employing Lists (4) as a between-subjects factor (materials counterbalancing factors) and Ambiguity Meaning (Primary and Secondary) and Probe Type (Related and Control) as within subjects factors. An overall significant main effect of Meaning (F(l,44) = 7.42, p < .009) and Probe Type (F(l,44) = 14.49, p < .001) was found. TABLE 1

MEAN REACTION TIMES (IN MILLISECONDS) FOR RELATED AND CONTROL PROBE LEXICAL DECISION TIMES FOR BOTH PRIMARY AND SECONDARY MEANINGS OF THE AMBIGUITY Probe type Meaning of ambiguity

Related

Control

Primary Secondary

622* 576*

651 642

* significant difference between related and control contrast. See text for details.

Lexical Ambiguity Resolution

23

These effects were further examined via a priori planned (1-tailed) comparisons of the Probe Type condition (Related vs. Control Probes) for each of the Ambiguity Meaning Conditions (Primary and Secondary). There was significant priming for the related (compared to control) probes for the Primary meaning (F(l,44) = 3.01, p < .045), as well as for the Secondary meaning (F(l,44) = 17.78, p<. 001). 5.7. Discussion When occurring in a context that constrains its subordinate (or secondary) meaning, an ambiguity primes a visual target associated both to the dominant (or primary) meaning, as well as the subordinate meaning. This means that neither the strictly selective account nor the strongest form of the language-driven hypothesis holds. Thus, even though context plays a very important role in propositional interpretation in Mandarin, the language processor does not give context special priority with regards to determining the meaning of an ambiguous word. 5.8. General Discussion Even though Mandarin Chinese is a language in which context plays a critical role in sentential interpretation, context alone is not enough to influence lexical access. Thus, the language-driven hypothesis, as formulated in (3) above, does not hold. It is not the case that if language is context-prominent, context can, in turn, influence lexical access in that language. It remains to be seen if, and how, other linguistic properties of a language necessitate different language-specific processing mechanisms. I also did not find evidence to support a strictly selective access account. When I biased the context towards the secondary meaning of the ambiguous word, both the primary and secondary meaning were facilitated relative to their respective controls. This is evidence against an interactionist model of language processing, which postulates that information in one module can influence information in another module before or while the information in the latter module is being processed. At this point then, there is no evidence for an interactionist view of lexical access in ongoing natural speech comprehension. The only evidence there is to date for a modified selective access account is in Italian. In the introduction I pointed out the modified selective account made the same predictions as the ordered access account, which were that when the sentence is biased towards the dominant meaning of the ambiguity, only the dominant meaning is accessed, but when the sentence is biased towards the subordinate meaning of the ambiguity, both meanings are accessed. This prediction is compatible with a context-independent model because it could be the case that context was ignored, and the most common meaning was retrieved first, and then if that wasn't compatible, the next most common meaning is retrieved.

24

Kathleen V. Ahrens

Thus, the question at this point is no longer whether lexical access is influenced by context, because it is not, as is clearly shown from this experiment. Lexical access occurs independently of context. The question is in a language where context plays an important role (as in Italian and Mandarin), can the combination of context and dominance influence lexical access? There is a recent cross-modal lexical ambiguity resolution study in Cantonese, which addresses this question (Li and Yip, 1996). Li and Yip created contexts for both the dominant and subordinate meaning of the ambiguity and ran a cross-modal lexical naming task where the ambiguity occurred sentence finally. They found that both meanings of an ambiguity were accessed when the visual probe was presented at the onset of the ambiguity, but that only the contextually appropriate meaning was accessed when the visual probe was presented at the offset of the ambiguity. These results held when the sentence was biased toward either the dominant or subordinate meaning of the ambiguity. Their interpretation is as follows: Our results from the Onset. . . condition seem to suggest that there is a very brief moment of lexical access of multiple meanings following the onset of the acoustic signal. However, this moment is short-lived and other information such as frequency could start to play a role rapidly thereafter. (Li and Yip, 1996, p. 231).

Cantonese is mutually unintelligible with Mandarin, although they both share the same writing system. However, Cantonese does share the same characteristics of an impoverished agreement system, and subject and object omission, although it differs from Mandarin in that the subject and object position are more straightforwardly related to the agent and patient of a sentence, as in English. (That is, sentence (5) above is considered strange in Cantonese [Hintat Cheung, personal communication, January 1997]). Thus, Cantonese also relies on context to interpret the meaning of a sentence, although perhaps not to the extent that Mandarin does, but certainly more so than English. Li and Yip's findings are interesting for three reasons: first, they also do not support the language-driven hypothesis; that is, in Cantonese, which is also a context-prominent language, both meanings are always accessed first, and then the contextually appropriate meaning is selected. Second, their findings support a multiple-access account over a strictly selective or modified selective account, because both meanings were accessed at the onset position irrespective of the contextual bias. Third, Li and Yip found that context alone (even in the subordinate case) was enough to influence the result of lexical access at the offset of the ambiguity, which differs from what was found for Italian. The possibility then exists that it is because Cantonese is sensitive to contextual information that drives language processor to decide on the contextually appropriate reading as soon as the offset of the ambiguity. That is, lexical access is a modular and bottom-up

Lexical Ambiguity Resolution

25

process, but the fact that Cantonese relies on context so heavily forces the processor to choose between the ambiguous meanings faster than it does for Italian. Italian, in turn, relies on context more than English, and so the interaction of context and dominance contains enough information to allow the selection between ambiguous senses faster than in English. This, if true, would support a modified language-driven hypothesis, given in (7). (7) Speeded-Selection hypothesis: If a language is context-prominent, then the language processor will select the contextually appropriate meaning faster (once all meanings have been accessed) than it would in languages that rely less heavily on context for interpretation. The speeded-selection hypothesis differs crucially from the language-driven hypothesis in that it is not the structure of the language that affects the interpretive process. Instead, it is the processing strategy that affects the interpretive process. In other words, context-prominent languages may select contextually favored meanings faster because the contextual processing routine is performed much oftener in the language. Familiarity in the processing regime allows for quicker processing time. In addition, this hypothesis does not conflict with a context-independent account, because it assumes a modularly designed language-processing system. What this hypothesis does do is explain why Italian and Cantonese allow context to play a role so much faster than it does in English. But before we jump to this conclusion, we need to examine what else could possibly be different among the Italian, Cantonese, and English experiments. Methodological differences may be what is driving the difference in results among the different experiments. The task used in Italian was a go/no-go lexical decision task. That means that if the subjects decided that the string of letters flashing on the screen was a word, they pressed a button, but if they decided it wasn't a word, they did nothing. This task is different from the usual task of pressing one button if they see a word, and another button if they don't. The go/no-go task might lead to hesitancy on the part of the subject because they must hold back from performing an action when making a nonword decision. Instead of automatically performing an action as soon as they make a decision they must then remember not to perform an action. This could inhibit the automaticity of the decision. This is relevant to the modularity hypothesis because the modularity hypothesis assumes that only tasks that reflect the automaticity and immediacy of lexical access will not allow preceding information to influence lexical access (Fodor, 1983). Anytime that automaticity or immediacy is compromised, there is the possibility that other information could have interfered with access. However, if we follow this line of reasoning, then it is not immediately apparent why Cantonese, which used a naming task, found that context influenced lexical access more quickly than in English (i.e., at the offset of the ambiguous item). Naming tasks have been used in experiments that

26

Kathleen V. Ahrens

have found for the context-independent hypothesis (i.e., Swinney and Love, 1996; Seidenberg et al., 1982; Tanenhaus et al., 1979). Moreover, the naming task used seems to be consistently automatic in that only words are flashed on the screen, and the subjects are asked to immediately pronounce whatever word they see. There is no reason to hesitate in this type of task. (If, however, both words and nonwords flashed on the screen, and the subjects were asked only to name the words, then the task would involve the same hesitancy as the go/no-go lexical decision task.) Thus, it does not appear that the issue of automaticity can straightforwardly account for the differences in results that we find for Cantonese, Italian, and English. This brings us to the other critical methodological point—the issue of immediacy. The timing of the ambiguity resolution is critical to understanding lexical access. We have already pointed out in the introduction (and others have noted previously) that even a slight delay in the presentation of the visual target could mean that context was selecting a meaning after both meanings were already selected. In addition to the position point of the visual probe, there is another timing issue that to date has been overlooked, to the extent that its value is sometimes not clearly given in the methodology section. The timing issue we are referring to is the length of presentation of the visual probe. In Italian the visual probe appeared for 1500 ms (Tabossi, Columbo, and Job, 1987; Tabossi, 1988; Tabossi and Zardon, 1993), and in Cantonese it seems that the probe appeared for 2000 ms ("A visual probe then occurred on the computer screen. . . . Subjects were given a maximum of two seconds to respond, counting from the onset of the visual probe"; Li and Yip, 1996, p. 229). The experiments in English, however, never had the visual probe appear for more than 1000 ms (Onifer and Swinney, 1981), and was as short as 300 ms (Swinney and Love, 1996). It is difficult to ascertain if the length of response time in the Italian and Cantonese experiments was significantly longer than the response time to visual targets in English, but that is not what is at issue here. Obviously, the average length of response times to visual targets could differ across languages. What could be critical is that the subjects in the Italian (Tabossi, Columbo, and Job, 1987; Tabossi, 1988; Tabossi and Zardon, 1993) and Cantonese (Li and Yip, 1996) experiments see the visual target until they make a decision or the outer limit of time (either 1500 ms or 2000 ms) is reached. It is quite possible that when the visual target is presented at the offset of the word, and up to 1500 ms (or 2000 ms) is allowed to pass, context is able to influence the choice after modular lexical access has already occurred. In the cases—as in English (Swinney and Love, 1996) and Mandarin (see previous section)—where the visual target only flashes onscreen for 300 ms, the immediacy of making the decision is more pronounced. Again, since the modularity hypothesis assumes that immediate and automatic access is taking place, one must be sure that one is, in fact, measuring immediate and automatic access. In addition to the concerns of immediacy and automaticity, other methodologi-

Lexical Ambiguity Resolution

27

cal details should also be considered carefully. These include making sure the context is biased in a strong enough manner (Tabossi et al., 1987), and making sure that the experimental and control visual targets are equated in terms of a priori reaction times, and balanced for in terms of frequency and length. Also, to avoid the possibility of wrap-up effects, it is probably prudent to place the ambiguity in a sentence-medial position. The sentence-medial position also is a better test of ongoing processing. When the ambiguity occurs sentence-finally, the decision to the visual target actually occurs after the sentence ends. Only when all relevant factors, such as immediacy and automaticity of access, as well as other methodological issues, have been controlled for, will the theoretical questions concerning the lexical access of ambiguous words be resolved. In sum, the possibility exists then that the reason that selective access is found in Cantonese (Li and Yip, 1996) and Italian (Tabossi, Columbo, and Job, 1987; Tabossi, 1988; Tabossi and Zardon, 1993) at the offset position is (a) either because these languages are more sensitive to context effects than English (and so context influences the result of multiple-lexical access faster), or (b) because the visual probes appear on the screen long enough for context effects to choose between the meanings provided.5 If the first possibility is correct, then the Speededselection hypothesis stands. If the second possibility is correct, then when the visual probes appear for a shorter amount of time, both meanings should be accessed even when biased towards the dominant meaning, which would mean that the Speeded-selection hypothesis would not receive support. In either case, though, multiple access still occurs. In order to rule out multiple access, one has to show that context influenced lexical access at the point where the word is identified (which may be earlier than the actual end of the word itself) and also that it occurred within an appropriate time frame (i.e., the visual target occurred for only 300 ms at the identification point), and all other relevant methodological factors have been controlled for.6 In order to rule out the context-dependent account, one has to show that in a strongly biasing context, both (or all) meanings are still accessed at the identification point of the word (again with all the relevant methodological factors controlled for). This has been demonstrated for Mandarin (when the context is biased towards the subordinate meaning) and has also been demonstrated in English and Cantonese. It remains to be shown in other languages.

6. CONCLUSION This work points out that there is no evidence for a context-dependent, interactionist account of lexical access within the paradigm of natural speech comprehension. I also argue that the findings for a modified selective access account are

28

Kathleen V. Ahrens

(a) actually accountable for within a context independent framework, and (b) limited to Italian. Given the fact that the findings to date for lexical ambiguity resolution differ along language lines (i.e., in English both meanings are always immediately accessed—multiple access—whereas in Italian if the context is biased towards the dominant meaning, only that meaning is accessed—modified selective access), I hypothesize that there is something unique to the language that drives the processing differences (e.g., language-driven hypothesis). I examined this hypothesis by looking at a language, Mandarin Chinese, that is noted for its reliance on context for semantic interpretation. I presented findings from a cross-modal lexical decision experiment and demonstrated that both meanings of an ambiguity are semantically primed, which shows that even in a language that relies heavily on context, context alone cannot influence lexical access. Thus, there is no support for the language-driven hypothesis. The experiment, moreover, clearly demonstrated that there is no evidence within a cross-modal priming paradigm that supports a context-dependent account of lexical access in language processing. Two possibilities were put forth to explain the reason why multiple access occurs in English, regardless of the contextual bias, but not in Italian. One possibility is given in the speeded-selection hypothesis, which postulates that contextprominent languages (such as Italian), will allow context to select a meaning (once all meanings have been accessed) faster than in languages that are not as context prominent (such as English). Another possibility is that the length of presentation of the visual target is another timing factor that is crucial to understanding the time course of lexical ambiguity resolution. If the length of the visual presentation of the target is too long, the immediacy and automaticity of lexical access may be compromised. In addition, methodological details should be controlled for as tightly as possible, including controlling for equality of reaction times of groups of visual targets, similarity in task and sentential position, pretesting that a context is strongly enough biased, and so on. Future research in lexical ambiguity resolution will benefit from limiting the scope of inquiry to one area of language comprehension at a time, ascertaining the relevance of the relative context-prominence of a language, and examining issues relating to the automaticity and immediacy of lexical access.

ACKNOWLEDGMENTS The author gratefully acknowledges the support of a research grant from National Chung Cheng University during 1995-1996, and from National Taiwan University during 19961997. I would like to thank Jean Hsu and Douglas Chiu for their assistance in collect-

Lexical Ambiguity Resolution

29

ing the data presented here. I am grateful to Chu-Ren Huang, Tracy Love, and David Swinney for their comments on the work presented here. All remaining errors are my sole responsibility.

REFERENCES Ahrens, K. (1994). Wh-in-situ and Specificity. Proceedings of the Chicago Linguistics Society. Volume 30. CKIP. (1993). Corpus-Based Frequency Count of Words in Journal Chinese. Taipei: Academia Sinica, Chinese Knowledge Information Processing Group. Fodor, J. A. (1983). Modularity of the Mind. Cambridge: MIT Press. Glucksberg, S., Kreuz, R J., and Rho, S. H. (1986). Context can constrain lexical access: Implications for models of language comprehension. Journal of Experimental Psychology: Learning, Memory and Cognition, 12, 323-335. Kintsch, W., and Mross, E. F. (1985). Context effects in word identification. Journal of Memory and Language, 24, 336-349. Hogaboam, T, and Perfetti, C. (1975). Lexical Ambiguity and Sentence Comprehension. Journal of Verbal Learning and Verbal Behavior, 14, 265-274. Huang, C. R. (1985). Possessive subjects. Cornell Working Papers. Li, P., and Yip, M. C. (1996). Lexical Ambiguity and Context Effects in Spoken Word Recognition: Evidence from Chinese. Proceedings for the 18th Annual Meeting of the Cognitive Science Society, (pp. 228-232). Hillsdale, NJ: Erlbaum. McClelland, J. L. (1987). The case for interactionism in language processing. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 3-36). Hillsdale, NJ: Erlbaum. Nicol, J., Fodor, J. D., and Swinney, D. A. (1994). Using cross-modal lexical decision tasks to investigate sentence processing. Journal of Experimental Psychology: Learning, Memory and Cognition, 20(5), 1229-1238. Onifer, W, and Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 15, 225-236. Paul, S. T, Kellas, G., Martin, M., and Clark, M. B. (1992). The influence of contextual features on the activation of ambiguous word meanings. Journal of Experimental Psychology: Learning, Memory and Cognition, 18, 703-717. Prather, P., and Swinney, D. A. (1988). Lexical Processing and Ambiguity Resolution: An autonomous process in an interactive box. In S. Small, G. Cottrell, and M. Tanenhaus (Eds.), Lexical ambiguity resolution: Perspectives from psycholinguistics, neuropsychology and artificial intelligence. San Mateo, CA: Morgan Kaufmann Publishers. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136. Simpson, G. B. (1994). Context and the Processing of Ambiguous Words. In M. Gerns-

30

Kathleen V. Ahrens

bacher (Ed.), Handbook of Psycholinguistics. (pp. 359-374). San Diego: Academic Press. Simpson, G. B., and Krueger, M. A. (1991). Selective access of homograph meanings in sentence context. Journal of Memory and Language, 30, 627-643. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645 - 660. Swinney, D. A., and Love, T. (1996). Co-reference processing and levels of analysis in object-relative constructions: Demonstration of antecedent reactivation with the cross-modal priming paradigm. Journal of Psycholinguistic Research. 25, (1) 5-24. Tabossi, P., Colombo, L., and Job, T. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential context. Journal of Memory and Language, 27, 324-340. Tabossi, P., and Zardon, F. (1993). Processing ambiguous words in context. Journal of Memory and Language, 32, 359-372. Tanenhaus, M. K., Leiman, J. M., and Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18,427-440. Till, R. E., Mross, E. F., and Kintsch, W. (1988). Time course of priming for associate and inference words in a discourse context. Memory and Cognition, 16, 283-298.

NOTES 1

Even though unimodal (reading) experiments have found for both the selective (Simpson and Krueger, 1991; Paul et al., 1992) and multiple-access models (Kintsch and Mross, 1985; Till et al., 1988), the reasons for the discrepancies in these cases might be related to the method of visual presentation, the time course of the presentation, the reading span and/or reading level of the subjects, or the task. For example, both Kintsch and Mross (1985) and Till et al. (1988) used a Rapid Serial Visual Presentation method in which the words appeared one at a time in the center of the screen. Simpson and Krueger (1991), on the other hand, used a paradigm where the entire sentence was presented, and Paul et al. (1992), used a RSVP method where the sentence appeared to "unfold" across the screen from left to right. Thus, the reasons for the difference in results should be looked at within the methodological constraints of the unimodal experiments, and should not be mixed in at this point in time with cross-modal experiments that are examining lexical ambiguity resolution in ongoing speech comprehension. 2 Signing comprehension is an interesting case because it shares characteristics with both reading comprehension and speech comprehension. It shares with reading the fact that it occurs in the visual modality, but it shares with speech the fact that it is used for communicative purposes with other speakers, is learned earlier than reading, and is the basis for learning how to read. 3 Mandarin Chinese is the official governmental language that is spoken throughout China and Taiwan by the majority of speakers, who may in addition speak another regional

Lexical Ambiguity Resolution

31

Chinese language at home. In this chapter, I focus on the Mandarin Chinese that is spoken in Taiwan. 4 In addition, Ahrens (1994) demonstrated that the syntactic phenomenon of wh-in situ extraction over islands in Mandarin (at the level of Logical Form) depended not on a syntactic argument/adjunct asymmetry, but instead had to do with whether or not there was a possible set of referents accessible to the speaker and hearer. This is an example of semantic information determining syntactic behavior and prepositional interpretation. 5 It could be the case in Italian, however, that the issue of automaticity is interacting with the issue of immediacy. That is, when the probe appears for up to 1500 ms and the task is a go/no-go task, the combination of both could lead to context and dominance influencing access. 6 In addition, the point of lexical access in words will vary—especially in longer syllable words, the word may be accessed before the final syllables have been heard. This issue is critical because the question is whether all meanings are retrieved at the point of access. If the word has already been accessed (at, say, the second syllable), but the visual target doesn't appear until the end of the fourth syllable, 200-250 ms could have already passed by.

This page intentionally left blank

CROSSLINGUISTIC VARIATION AND SENTENCE PROCESSING: THE CASE OF CHINESE PING LI Department of Psychology University of Richmond Richmond, Virginia

1. INTRODUCTION Theories of sentence processing have for the last two decades focused on issues such as parallel versus serial, interactive versus modular, and top-down versus bottom-up processes. One classic example is the study of lexical ambiguity (Small, Cottrell, and Tanenhaus, 1988; see other chapters in this volume), out of which two contrastive hypotheses have emerged. The first is the exhaustive access hypothesis, which argues that all meanings of an ambiguous word will be accessed momentarily following the occurrence of the word, and that semantic context can only help to select the appropriate meaning at a post-access stage. This hypothesis assumes that sentence processing is a modular, bottom-up process in which nonlexical, sentential information does not penetrate lexical access (Onifer and Swinney, 1981; Swinney, 1979). The second hypothesis is the contextdependency hypothesis, which argues that the contextually appropriate meaning of an ambiguous word can be selectively accessed early on if sentence context provides a strong bias to the appropriate meaning. This hypothesis assumes that sentence processing is an interactive process in which information can flow both bottom-up and top-down, and that lexical access and sentential context can mutually influence one another at a very early stage (Simpson, 1981; Simpson and Krueger, 1991; Tabossi, 1988). Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

33

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

34

Ping Li

Most sentence-processing theories have been built on facts from Western languages, in particular, English. They emphasize the extent to which languages resemble each other in sentence processing, for example, on the use of a universal set of parsing principles (such as "late closure" and "minimal attachment," see Frazier, 1987, 1990). In recent years, however, we begin to see growing interests in crosslinguistic research in sentence processing. In particular, the competition model of Bates and MacWhinney (1982, 1987, 1989; see also Bates, McNew, MacWhinney, Devescovi, and Smith, 1982) has attracted attention from researchers with a crosslinguistic perspective on sentence processing. In contrast to other models, the competition model emphasizes the extent to which languages differ, and the extent to which these differences affect sentence processing. Chinese, because of its many unique linguistic properties, plays an important role in crosslinguistic studies of sentence processing. Spoken by one-fifth of the world's population, Chinese differs significantly from most Indo-European languages and offers unique features in its phonological, lexical, and syntactic structures. On a phonological level, Chinese involves a tonal system that distinguishes lexical items, with each item carrying a particular tone. Tonal information can differentiate lexical items, but it does not eliminate lexical ambiguities associated with homophones: on a lexical-morphemic level, Chinese has a massive number of homophones. Moreover, on a syntactic level, first, Chinese does not have devices that indicate differences in tense, number, gender, or case; in other words, grammatical functions and relations for sentence constituents are not linked by morphological associations. Chinese has some grammatical morphemes such as the object-patient marker ba and the passive-agent marker bei, but these are often optional in natural speech. Second, Chinese involves a high degree of ellipsis. A sentence can be subjectless or objectless, sounding telegraphic in a richly inflected language when literally translated. Third, word order is relatively free in Chinese. In addition to the canonical SVO order, Chinese permits several other word orders in daily spoken language: OSV, in which the object is topicalized; SOV, in which the speaker provides information counter to the expectation of the listener (Li and Thompson, 1981); and VOS, in which the subject is usually an after-thought (Lu, 1980). These language-specific properties of Chinese raise interesting questions about mechanisms of sentence processing. For example, the lack of inflections, the high degree of ellipsis, and the variability of word order together offer sentence status to many constructions that would be incomplete and/or ungrammatical in English or other Indo-European languages (e.g., V, OV, VO). Given that many daily sentences consist of fragments with omissions, we may expect partial information and sentence fragments to play a more important role in Chinese processing than they do in the processing of Indo-European languages. One of the goals of this chapter is to examine the differences between Chinese and other languages with respect to the use of different types of information in sentence processing.

Variation and Sentence Processing in Chinese

35

The basic principles of the competition model seem to give rise to a natural explanation of the language-specific properties of Chinese and how these properties influence Chinese sentence processing. The competition model is a cue-based interactive model. It is concerned with how speakers integrate various types of information or cues in a sentence (e.g., word order, subject-verb agreement, and lexical semantics) to determine sentence roles (e.g., who does what to whom). A central tenet of the model is that the same cues vary in their validity across languages, with validity being determined by how often the cue is available, and how reliably the cue leads to the correct identification of linguistic functions. A typical finding in this research is that the strongest cue in one language can be one of the weakest cues in another. For example, in Italian, the strongest cue to the identification of sentence roles is verb morphology, whereas in English, morphological marking is extremely weak as a cue (Bates et al., 1982). In the last few years, my colleagues and I have used the basic constructs of the competition model to study how Chinese speakers assign sentence roles in the absence of overt grammatical devices and in the presence of a relatively free word order (Li, Bates, Liu, and MacWhinney, 1992; Li, Bates, and MacWhinney, 1993; Li, 1996a). This chapter provides an overview of some basic results from these studies. The assignment of sentence roles is concerned with global functional (i.e., syntactic and semantic) relationships among sentence constituents.1 In addition to the understanding of sentence role assignment in Chinese, I have also studied how Chinese speakers resolve local ambiguities associated with homophones within a sentence (Li, 1996b; Li and Yip, 1996; in press). Previous research in lexical ambiguity has focused on English and several Indo-European languages (e.g., Dutch and Italian). In those languages, homophony is a relatively low-frequency event. In Chinese, by contrast, homophony is extensive. According to the Modern Chinese Dictionary (Institute of Linguistics, 1985), 80% of the monosyllables in Chinese are ambiguous, and half of them have five or more homophones. The single syllable yi with the dipping tone has up to 90 homophones (e.g., skill, justice, benefit, discuss, intention, translate, hundred-million, etc.), and this number would increase to 171 if identical syllables with different tones were considered as homophones. Upon hearing yi in a sentence, do Chinese speakers activate all 90 or more meanings of the syllable? The exhaustive access hypothesis should predict they do, because lexical access is an autonomous and capacity-free process. The context-dependency hypothesis should predict that Chinese listeners activate only the contextually appropriate meaning with aid from sentence context. This chapter reviews relevant experimental evidence from my studies of the processing of Chinese homophones. The chapter is organized as follows. First, I present results from experiments on how Chinese speakers determine sentence roles in a sentence-interpretation task (Experiment 1) and a sentence-gating task (Experiment 2). Second, I present results from experiments on Chinese speakers' processing of homophones, in a

36

Ping Li

cross-modal naming task (Experiment 3) and a word-gating task (Experiment 4). Finally, I conclude with a discussion on the importance of crosslinguistic variations in our understanding of the mechanisms of sentence processing.

2. EXPERIMENT 1: SENTENCE INTERPRETATION The sentence-interpretation task, according to Bates and MacWhinney (1982), is a task in which speakers choose one of the two nouns in a sentence as the agent or subject of the sentence. In the experiment, adult participants respond verbally for a choice and children act out the sentence with toys, after listening to the sentence (such as The elephant kicked the dog). A number of previous studies in Chinese have employed this task, including Miao (1981), Chen, Tzeng, and Bates (1990), and Li, Bates, Liu, and MacWhinney (1992). In the study reported here, a reaction time technique was used, in which listeners make their choices by pressing one of the two buttons upon a computerized presentation of a spoken sentence (Li, Bates, and MacWhinney, 1993). 2.1. Method 2.1.1. PARTICIPANTS Eighteen native Mandarin Chinese speakers from mainland China participated in this experiment. 2.1.2. MATERIALS AND DESIGN Three independent variables were tested in this experiment: 1. Noun animacy. A noun in a sentence was either animate or inanimate. Because all test sentences contained two nouns, the first and the second nouns varied in three ways: Animate-Inanimate (AI), Inanimate-Animate (IA), and Animate-Animate (AA). 2. Word order. The sentences varied in three orders: the canonical Noun-VerbNoun (NVN), usually associated with a SVO interpretation; the noncanonical Noun-Noun-Verb (NNV), associated with either a SOV or a OSV interpretation, and the noncanonical Verb-Noun-Noun (VNN), associated with a VOS interpretation. 3. The presence or absence of a grammatical marker. The marker ba, which marks the following noun as the patient of the sentence, or the marker bei, which marks the following noun as the agent of the sentence.

Variation and Sentence Processing in Chinese

37

A total of 162 test sentences were used in this study: 54 each for (a) simple transitive sentences, (b) sentences with the marker ba, and (c) sentences with the marker bei.2 In each set of sentences, there were six instances for each of the following nine types: (a) AVA (animate noun-verb-animate noun), AAV, VAA, AVI (animate noun-verb-inanimate noun), AIV, VAI, IVA (inanimate noun-verbanimate noun), IAV, and VIA; (b) AVbaA, AbAV, VAbaA, AVbaI, AbaIV, VAbal, IbdVA, IfeaAV, and VIbaA; and (c) AVbeiA, AbeiAV, VAbeiA, AVbeiI, AbeiIV, VAbeiI, IbeiVA, IbeiAV, and VlbeiA. The markers ba and bei always occurred before the second noun, to match up with their position of occurrence in natural speech. 2.1.3. PROCEDURE The test sentences were read by a native Mandarin speaker and were digitized into a Macintosh computer with a sampling rate of 22 kHz. During the experiment, participants received each sentence auditorially, and an Experimental Control System (Clynes and MacWhinney, 1990) controlled the presentation of test sentences. At the onset of the presentation of each sentence, participants also received on the computer screen a pair of pictures that were described by the two nouns of the sentence. They were instructed (in Chinese) to determine, as quickly and as accurately as possible, which of the two pictures represented the agent or the actor in the sentence they heard. They indicated this choice by pressing one of the two buttons mounted on a CMU button box (Cohen, MacWhinney, Flatt, and Provost, 1993). 2.1.4. DATA ANALYSIS Although data from both choice responses and reaction times were collected and analyzed, only the choice response data are discussed here (see Li et al., 1993, for analyses of the reaction time data). For the scoring of choice responses, a participant's response was given a score of 1 if it was a choice of the first noun, and 0 if it was a choice of the second noun. The percent of first-noun choice is thus inversely related to that of second-noun choice: a score of 100% represents an exclusive choice of the first noun, while a score of 0% represents an exclusive choice of the second noun. 2.2. Results and Discussion Figures la to 1c present the percent first-noun choice for the simple transitive sentences, the sentences with the marker ba, and the sentences with the marker bei, respectively. A 3 X 3 (animacy by word order) analysis of variance (ANOVA) was conducted separately on the results of each set of sentences, and a 3 X 3 X 2

38

Ping Li

Figure 1. Percent first-noun choices for the simple sentences (la), the ba sentences (Ib), and the bei sentences (1c). (From Li, Bates, and MacWhinney, 1993. Reprinted by permission of Academic Press.)

(animacy by word order by the presence or absence of the marker) was conducted separately on the combined data of the simple and the ba sentences and those of the simple and the bei sentences. The latter procedure was used to understand the effects of the grammatical marker more clearly. For the simple sentences, ANOVA revealed that there were significant main effects of animacy (F(2,34) = 75.54, p < .01) and word order (F(2,34) = 21.61, p < .01), and an interaction between the two (F(4,68) = 11.46, p < .01). An analysis of the magnitude of effect (7 score) shows that animacy accounted for 72% of the experimental variance, whereas word order accounted for 23%. This analysis shows that animacy is a more important cue than word order in Chinese, in contrast to the predominant role of word order in English. The interaction between animacy and word order shows that when the two cues agreed (i.e., when they both pointed to the same noun as agent), the sentence led to consistent responses; when they competed (i.e., when they pointed to different nouns as agent), the sentence led to inconsistent responses. For example, listeners chose 98% first noun for the AVI sentences, in which both animacy and word order favor the first noun. In contrast, listeners chose only 31% first noun for the IVA sentences, in which word order favors the first noun but animacy favors the second. These effects clearly demonstrate how different cues compete and collaborate in determining the assignment of sentence roles. For the sentences with the marker ba, a 3 X 3 ANOVA revealed that there were main effects of animacy (F(2,34) = 66.64, p < .01) and word order (F(2,34) = 10.32, p < .01), and an interaction between the two (F(4,68) = 11.28, p < .01). A 3 X 3 X 2 ANOVA incorporating the simple sentences revealed that there was a significant main effect of the marker ba (F(l,17) = 15.75, p < .01), suggesting that the presence of ba contributed to the identification of sentence roles. How-

Variation and Sentence Processing in Chinese

39

ever, a comparison of Figure Ib with la shows that the basic patterns for the simple sentences and for the ba sentences are similar, except that the AbaAV sentences elicited a higher percentage of first-noun choice than the simple AAV sentences. The presence of ba had its effect most clearly on the NNV word order, since this is the order in which ba naturally occurs in speech. For the sentences with the marker bei, a different pattern of response emerged, in contrast to the patterns for the simple and the ba sentences. A 3 X 3 ANOVA revealed that there was a main effect of animacy (F(2,34) = 12.80, p < .01) and a significant interaction between animacy and word order (F(4,68) = 2.89, p < .05), but there was no main effect of word order (F(2,34) = 1.51, ns). Figure 1c shows that the difference between the three word-order types within the bei sentences was minimal. A 3 X 3 X 2 ANOVA incorporating the simple sentences revealed that there was a significant main effect of the marker bei (F(l,17) = 27.65, p < .01). A comparison of Figure 1c with la indicates that in all cases, the presence of bei suppressed the first-noun choice and promoted the second-noun choice. The 3 X 3 X 2 ANOVA also revealed a significant interaction between animacy and bei (F(2,34) = 9.97, p < .01). This interaction indicates effects of competition and convergence between cues: when bei (which marks the second noun as agent) agreed with animacy (when the second noun is animate, i.e., IA), the sentence led to almost exclusive second-noun choice; when bei competed with animacy (i.e., AI), the sentence had less consistent responses (about 40% first-noun choices). But it was clear that the effect of bei overwhelmed both animacy and word order in all cases of competition. This effect differed from that of ba, which was reflected only in the NNV orders. The above results suggest that the strength of cues in Chinese sentence processing can be rank ordered as follows: the marker bei animacy word order the marker ba. The results show that Chinese speakers use a weighted array of linguistic cues to help solve the task of sentence role assignment. Speakers of Indo-European languages usually rely on one primary type of cue in sentence interpretation: morphological, syntactic, or semantic, since in these languages the primary type of information is both highly available and highly reliable in sentence processing. Chinese lacks such primary types of cues (e.g., bei is highly reliable but often not available), so Chinese speakers have to rely on a multitude of cues, including noun animacy, word order, and the grammatical markers. The difference between the effects of ba and those of bei raises another important issue. Although both ba and bei are considered as grammatical markers (or "semimorphological markers," see discussion in Li et al., 1993), their influence on sentence processing differs. In the above cue-strength hierarchy, ba and bei (grammatical markers) are intercepted by animacy (semantic cues) and word order (syntactic cues), which suggests that cues that fall into the same cluster of linguistic type do not necessarily fall on the same hierarchy of cue importance to sentence processing. Such results are best described by an interactive process in

40

Ping Li

which different types of linguistic cues work together in collaboration and in competition. The outcome of the interactive process depends on the strengths of cues and the directions in which they interact (e.g., word-order effect is significant in sentences with ba but not in sentences with bei). Experiment 1 indicated the relative importance of several cues in Chinese sentence processing, and the ways in which these cues interact to determine the final outcome of sentence role assignment. Although a reaction time study, the experiment was, however, confined to the "one sentence one choice" decision process, and was thus relatively uninformative to the question of when a choice decision is reached (i.e., the time course of sentence processing). To understand how cues interact with one another at different temporal locations of the sentence, in Experiment 2 I adapted the gating method from spoken-word recognition to study sentence role assignment (Li, 1994, 1996a).

3. EXPERIMENT 2: SENTENCE GATING 3.1. Method 3.1.1. PARTICIPANTS Twenty native Mandarin Chinese speakers from mainland China participated in this experiment. None had participated in Experiment 1. 3.1.2. MATERIALS AND DESIGN To match up with Experiment 1, the test sentences included three sets of sentences: simple, ba, and bei. Each set had 36 sentences, 12 each for the three-word orders NVN, NNV, and VNN. The design included three levels of word order (NVN, NNV, and VNN), and either three or four levels of gate size (simple sentences had three levels, i.e., 1, 2, or 3 words; ba and bei sentences had four levels, i.e., 1,2, 3, or 4 words). A third variable, position of the marker, was nested within the ba and bei sentences only, reflecting the position of ba or bei in the sentence (before the first word, before the second word, or before the third word). Thus, the design was 3 X 3 for the simple sentences, and 3 X 4 X 3 for ba and bei sentences. Animacy was not varied in this experiment because the relative importance of animacy versus word order and ba and bei is clear from Experiment 1. 3.1.3. PROCEDURE The procedure was similar to that in Experiment 1, except the following. In Experiment 1, listeners heard the complete sentence and responded for their choice. In this experiment, listeners were presented auditorially with gated fragments of a sentence, one at a time, in increasing length, until they heard the com-

Variation and Sentence Processing in Chinese

41

plete sentence. At each presentation, they were required to decide on the agent of the sentence on the basis of the partial information provided up to that point (the last gate corresponded to the complete sentence). This procedure is an adaptation of the word-gating paradigm developed by Grosjean and his associates (Grosjean, 1980; Cotton and Grosjean, 1984; Tyler and Wessels, 1985), and it bears similarity to the self-paced sentence reading task. 3.1.4. DATA ANALYSIS The gating study produced a large number of data points because each sentence involved at least three choice responses. The following discussion focuses on the simple sentences, the canonical NVN orders with the marker ba, and the canonical NVN orders with the marker bei. As in Experiment 1, only the choice responses are discussed here (see Li, 1996a, for a detailed discussion of the data). 3.2. Results and Discussion Figures 2a to 2c present the percent first-noun choice for the simple sentences, the NVN sentences with the marker ba, and the NVN sentences with the marker bei, respectively. For the simple sentences, a 3 X 3 (word order by gate size) ANOVA indicated a significant main effect of word order (F(2, 38) = 85.75, p < .01), showing that different orders led to different interpretations.3 There was also a significant main effect of gate size (F(2, 38) = 28.78, p < .01), showing that listeners' choice responses tended to change along the course of the sentence. More important, the significant interaction between gate size and word order (F(4, 76) = 26.21, p < .01) showed that listeners built up their interpretations across the sentence in

Figure 2. Percent first-noun choices for the simple sentences (2a), the NVN sentences with ba (2b), and the NVN sentences with bei (2c). (From Li, 1996a. Reprinted with permission of the Psychonomic Society Publications.)

42

Ping Li

different ways for different sentence types. For example, in Figure la, first-noun choices increased for NVN as gate size increased, but decreased for NNV and VNN. Listeners started out with either biased or random default interpretations, and then gradually adjusted (confirming or disconfirming) the interpretations as more information came in. An increase in the amount of sentence information may lead to more certain decisions (as in NVN or VNN), but it can also lead to more ambiguous decisions (as in NNV), depending on the interpretations that the construction is associated with in the language (e.g., NNV is ambiguous between SOV and OSV interpretations). For the NVN sentences with ba, a 3 X 3 (position of marker by gate size) ANOVA indicated a significant main effect of position of the marker (F(2, 38) = 41.92, p < .01), no main effect of gate size (F(3, 57) = .58, ns), and a significant interaction between the two (F(6,114) = 17.51, p < .01). Figure 2b shows that the NfozVN and NVfozN sentences had mostly first-noun choices, similar to the simple NVN sentences without the marker ba (cf. Figure 2a). The similarity in the basic patterns of interpretation across the simple NVN and the Nba VN and NVbaN sentences suggests that the ungrammaticaliy of ba (i.e., occurring before and after a verb) did not disrupt listeners' comprehension; in Chinese, ba is always before a noun and it is ungrammatical before or after a verb. In contrast to N b V N and NVbaN, for the baNVN sentences, listeners quickly decided on the second noun as the agent (by the time they heard baN), and this interpretation was further confirmed when they heard the verb (baNV). Note that listeners' choice decisions were based on an interpretation of possible fragments (baN and baNV) from the complete grammatical NbaNV sentence: the N after ba was marked as the patient so that the alternative N could be assigned the agent role. Now if this interpretation were to hold, the sentence would end with a verb. But the sentence continued as baNVN, with a final N after the verb. This continuation created a competition between two fragment interpretations: that of baN or fozNV, in which the N before the verb should be interpreted as the patient, and that of VN, in which the N after the verb should be interpreted as the patient, leading to a final increase in first-noun choice. For the NVN sentences with bei, a 3 X 3 (position of marker by gate size) ANOVA indicated significant main effects of both position of the marker (F(2, 38) = 66.87, p < .01) and gate size (F(3, 57) = 30.44, p < .01), and a significant interaction between the two (F(6,114) = 42.73, p < .01). Figure 2c shows that listeners had a first-noun bias when they heard only the first N alone in the NbeiVN and NVbeiN sentences (as they did with the simple NVN and NNV sentences; see Figure 2a and Note 3). First, for NVbeiN, the bias was further confirmed when listeners heard NV, leading to a very high 98% first-noun choice with a S V interpretation. But when they heard bei after NV (i.e., NVbei), the firstnoun choice lost weight, since bei indicates the following noun as the agent in the language. The weight went down further to favor a second-noun choice when the final N appeared, due to a competition between NV and beiN (with both nouns competing for agenthood). Second, for Nbei VN, the decision moved quickly from

Variation and Sentence Processing in Chinese

43

a first-noun bias to a second-noun choice when bei occurred, and the second-noun choice continued through to the final gate. Note that the occurrence of bei in NbeiVN, unlike ba in NbaVN, strongly affected the SVO interpretation associated with NVN. This is because the noun following bei can be legitimately omitted from NbeiNV in the language (resulting in NbeiV similar to the English truncated passives), but the noun after ba in NbaNV cannot be omitted. Thus, listeners could readily assign an object role to the first N in NbeiV, using fragment interpretation based on the complete form NbeiNV. Finally, for beiNVN, the choices started at random when only bei was heard, but quickly moved to a first-noun choice. This pattern was the opposite of that for baNVN, because the fragment interpretations of beiN (N as agent) and VN (N as patient) were consistent, unlike the two nouns in baN and VN, which competed for patienthood in baNVN. Experiment 2 points to the usefulness of the gating method in unraveling the process of sentence comprehension as the auditory stimulus unfolds in time. Gating allows one to see the temporal structure involved in this process: for example, it shows how choice decisions move along the course of the sentence. Gating also permits the evaluation of listeners' processing of a sentence on the basis of partial or fragment information for that sentence: for example, a given sentence such as baNVN is associated with multiple fragment interpretations (as shown in parentheses below), including baN(O), NV (SV), NVN (SVO), and VN (VO), and gating provides a means of probing into these interpretations. Experiment 2 suggests that Chinese listeners rely strongly on the integration of fragment information in sentence processing. Daily spoken Chinese contains many sentences that are either incomplete or ungrammatical when literally translated into Indo-European languages, due to frequent omissions of subject and object and a flexible word order. For instance, both baNV and VN can stand alone as complete sentences with omitted subject, and NV as a complete sentence with omitted object (SV) or subject (OV). Thus, Chinese listeners have to rapidly incorporate partial sentence information rather than to wait for "complete sentences" to occur, whenever fragment interpretations become available. For example, in this experiment, the fragments baN and baNV were interpreted as part of the complete NbaNV, and Nbei and NbeiV were interpreted as part of the complete NbeiNV. Moreover, the results with ba also indicate that the grammatical violation produced by the wrong positioning of the marker does not disrupt comprehension, showing that Chinese listeners are less sensitive to grammaticality and can interpret ungrammatical sentences by reference to grammatical models. This contrasts with sentence processing in English, in which grammaticality plays a more important role (Bates, 1991; von Berger, Wulfeck, Bates, and Fink, 1996). These properties of processing in Chinese reflect language-specific properties of the language, in which morphological, grammatical, and syntactic constraints on sentence constituents are very weak. The above experiments were concerned with how Chinese speakers determine the global relationships among different sentence constituents (nouns and verbs).

44

Ping Li

The next two experiments are concerned with how Chinese speakers identify locally ambiguous items, homophones, in a sentence.

4. EXPERIMENT 3: CROSS-MODAL NAMING 4.1. Method 4.1.1. PARTICIPANTS Thirty native Cantonese Chinese speakers from Hong Kong participated in this experiment. 4.1.2. MATERIALS AND DESIGN Thirty spoken homophone nouns were selected, each with at least two different meanings that share the same syllable and the same tone. Each homophone was embedded in two different sentences, one biasing either of the two selected meanings. A separate group of 20 speakers was asked to judge the degree of constraint of the prior context on the target homophone. They were given the test sentences with the prior context but without the homophone, and were asked to fill in a word that naturally completes the sentence. Their responses were scored on a 1 to 4 scale, depending on how close the word that they chose matched the homophone (Marslen-Wilson and Welsh, 1978). A high-constraint score (mean score 1.6) was derived from these responses, indicating that the sentence contexts were highly constraining. Three independent variables were manipulated in this experiment. 1. Probe type. The visual probe was (a) biased, which was related to the contextually biased meaning of a homophone, or (b) unbiased, which was related to the second meaning not biased by the context, or (c) unrelated control. 2. Dominance. Half of the prior contexts biased the dominant (more frequent) meaning of the homophone, and the other half the subordinate meaning (less frequent). The frequency information was based on Ho and Jiang (1994). 3. Homophone density. Half of the homophones had many competitors (four or more alternative meanings—high density) and the other half had few competitors (two to three alternative meanings—low density). The 30 participants were randomly assigned to six groups of five. Each group received an equal number of sentences in the 3 (probe type) X 2 (dominance) X 2 (homophone density) design. The order of presentation was counterbalanced across participants. 4.1.3. PROCEDURE The test sentences were read by a native Cantonese speaker and were digitized into a Macintosh computer with a sampling rate of 22 kHz. The presentation of

Variation and Sentence Processing in Chinese

45

auditory and visual stimuli was controlled by the PsyScope program (Cohen et al., 1993). Naming latencies were recorded by the CMU button-box with a voiceactivated relay (Cohen et al., 1993). A cross-modal naming task (cf. Seidenberg, Tanenhaus, Leiman, and Bienkowski, 1982; Swinney, 1979) was used as follows. Participants saw a fixation point, and immediately heard on a pair of headphones the sentence in which the homophone was embedded. A visual probe that was either related (i.e., biased or unbiased) or unrelated to the homophone occurred at the offset of the homophone.4 The participant's task was to, as accurately and quickly as possible, name the visual probe aloud. 4.1.4. DATA ANALYSIS The dependent variable was participants' response latencies to each visual probe. The latency was measured from the onset of the visual probe to the onset of the vocal response (see Li and Yip, 1996, for a detailed discussion). 4.2. Results and Discussion Figure 3 presents the naming latencies as a function of probe type, dominance, and homophone density. A 3 X 2 X 2 (probe type X dominance X homophone density) ANOVA indicated a main effect of probe type (F(2,28) = 8.84, p < .01) and an interaction between probe type and dominance (F(2,58) = 4.62, p < .05). No other effects were significant. First, the main effect of probe type shows that the naming of the visual probes differed in speed as a function of whether the probe was related to a contextually biased, unbiased, or totally unrelated meaning of the homophone. This result pro-

Figure 3. Naming latencies as a function of probe type, dominance, and homophone density. (From Li and Yip, 1996. Reprinted with permission of Lawrence Erlbaum Associates, Inc.)

46

Ping Li

vided evidence for the operation of context effects, since a contextually biased meaning elicited the fastest response to the visual probe, as compared with the unbiased and the unrelated meaning. Second, the interaction between probe type and dominance was due to the pattern that dominant (more frequent) meanings were in general accessed faster than subordinate (less frequent) meanings, but when the context biased the homophone meaning (i.e., in the biased probe condition), dominant and subordinate meanings did not differ (F(l,59) = 2.65, ns). This pattern indicates that context and frequency can mutually interact during the identification of homophone meanings, and that context may be sufficiently constraining to act early on to reduce the effect of the frequency of homophone meanings. Finally, the absence of a homophone density effect shows that the number of competing meanings in a homophone did not affect listeners' responses, suggesting that the context effect operated early enough, probably within the acoustic boundary of the homophone, to select the contextually appropriate meaning and suppress the inappropriate meanings. Our results show that Chinese speakers are sensitive to the contextually biased meaning at an early stage, at least immediately following the occurrence of the homophone (i.e., the acoustic offset of the word). This context effect occurs much earlier than what has been previously assumed (e.g., about 1.5 sec following the occurrence of the ambiguous word, e.g., as shown in Onifer and Swinney, 1981). It seems that Chinese listeners, to cope with the extensive ambiguity created by massive homophones, can rapidly disambiguate alternative meanings as they hear a homophone in a sentence. Our results are consistent with the contextdependency hypothesis, which argues that ambiguous meanings of a word may be selectively accessed at an early stage according to prior sentential context (Simpson, 1981; Simpson and Krueger, 1991; Tabossi, 1988). In a recent study Moss and Marslen-Wilson (1993) argued that the offset of an ambiguous word may not be the critical point for tapping into the locus of context effect, because many words in context could be recognized before the acoustic offset of the word. They suggested that the initial access and selection of contextually appropriate meaning may occur earlier than the word offset. Since the crossmodal method in Experiment 3 tapped the recognition process only at the offset of the homophone, in Experiment 4I used a word gating-method (Grosjean, 1980) to examine how early context effects can be observed within the acoustic boundary of the homophone. 5. EXPERIMENT 4: WORD GATING 5.1. Method 5.1.1. PARTICIPANTS Eighteen native Cantonese Chinese speakers from Hong Kong participated in this experiment. None had participated in Experiment 3.

Variation and Sentence Processing in Chinese

47

5.1.2. MATERIALS AND DESIGN The same 30 homophones and their biasing sentence contexts from Experiment 3 were used. A neutral sentence context that biased neither meaning of the homophone was added to the test sentences for each homophone. Thus, there were three levels of sentence context in its bias to homophone meanings: biasing the dominant (more frequent) meaning, biasing the subordinate (less frequent) meaning, and neutral.5 There were two levels of homophone density, as in Experiment 3: high density, in which a homophone had many competitors (four or more), and low density, in which it had few competitors (two to three). 5.1.3. PROCEDURE Each homophone embedded in a sentence context was gated with a digital editing program, and presented to listeners as follows (see Grosjean, 1980,1988; Li, 1996c). Listeners heard the preceding sentence context plus a gated portion of the homophone, one at a time in increasing length, until the whole homophone was presented. The first gate contained the preceding context up to, but not including the homophone, the second gate contained the first gate plus the first 40 ms of the homophone, the third gate contained the second gate plus an additional 40 ms of the homophone, and so on, until the last gate reached the end of the homophone. The presentation of successive gates was controlled by the PsyScope program (Cohen et al., 1993). Listeners were asked to identify, for each presentation, the last word (i.e., the homophone) being presented. The 18 participants were randomly assigned to three groups of six. Each group received an equal number of sentences in the 3 (sentence context) X 2 (homophone density) design. Participants wrote down their answers in Chinese characters each time they heard a gate of the homophone along with its sentence context, and then pressed the computer spacebar to hear the next gate. 5.1.4. DATA ANALYSIS The dependent variable was the amount of acoustic information that listeners needed to correctly identify the homophone in context; in the case of a neutral context, any meaning of the homophone was regarded as correct. The correct identification was expressed as "percent word needed," that is, the amount of acoustic signal of the homophone that listeners needed divided by the total length of the homophone. 5.2. Results and Discussion Figure 4 presents the percentage of words needed for the correct identification of the homophone, as a function of sentence context and homophone density.

48

Ping Li

Figure 4. Percent word needed for the identification of the homophone, as a function of context and homophone density.

A 3 X 2 (context X homophone density) ANOVA revealed main effects of both context (F(2,34) = 51.28, p < .01) and homophone density (F(l,17) = 13.15, p < .05), and a significant interaction between the two (F(2,34) = 14.29, p < .01). As seen in Figure 4, the main effect of context came from the difference between the semantically neutral context and the biased contexts. A post hoc test (Tukey HSD) revealed no difference between the two biased contexts, but differences between the neutral context and either of the biased contexts. The main effect of homophone density was due to the difference between the low- versus high-density homophones in the neutral context (F(l,17) = 16.93, p < .01); the other two contexts showed no difference in homophone density. The result that the homophone density effect existed only in the semantically neutral context but not in the semantically biased contexts also accounted for the interaction between context and homophone density. The interaction is consistent with the absence of homophone density effects in Experiment 3, in which I suggested that context effects can operate early to select only the contextually appropriate meanings and suppress the inappropriate meanings. These results indicate that when the context is semantically biased toward a specific meaning of the homophone, Chinese speakers can identify the appropriate meaning with a minimal amount of acoustic-phonetic information. On the average (averaging over the two biasing conditions), they needed only 25% of the homophone's acoustic signal. Note that 25% does not provide sufficient acoustic information about the status of the word. The processing system thus must be contex-

Variation and Sentence Processing in Chinese

49

tually driven, relying on top-down information to identify the word on the basis of only a quarter of the word's information. The results match up well with established estimates of word recognition times in English: Grosjean (1980) and Marslen-Wilson (1987) showed that when spoken-word recognition takes place in context, only half or even less of the acoustic information of a word is needed for correct identification. Our results also indicate that, consistent with the contextdependency hypothesis, the access of the appropriate meaning of the homophone can take place well before the acoustic offset of the homophone.

6. GENERAL DISCUSSION This chapter presents results from a number of experimental studies of sentence processing in Chinese. In Experiment 1, I examined several basic cues that are important to Chinese speakers' identification of sentence roles, and the ways in which these cues interact to determine the final outcome of the identification process. The results suggest that Chinese speakers use a weighted array of linguistic cues to guide sentence role assignment, including noun animacy, word order, and grammatical markers. In Experiment 2,1 studied interpretations associated with sentence fragments at various temporal locations of the sentence, and showed that the rapid integration of fragment interpretations and the interactions between these interpretations characterize much of sentence processing in Chinese. These processing characteristics are described with reference to the weak morphological and syntactic constraints in Chinese. The results also suggest that Chinese speakers show less sensitivity to grammatical violations than do English speakers. In Experiments 3 and 4,1 examined how Chinese speakers resolve extensive homophone ambiguities within a sentence. The results suggest that Chinese speakers rely on sentential context at an early stage to disambiguate homophone meanings, at least at the acoustic offset of the homophone (according to the cross-modal naming method), and possibly earlier (according to the gating method). The results are consistent with an interactive, context-dependency hypothesis but inconsistent with a modular, exhaustive-access hypothesis. A seemingly mysterious question that initiated my inquiries into Chinese sentence processing is this: how can Chinese speakers ever understand each other, if the language does not have clear inflectional markings (in contrast to many IndoEuropean languages), lacks a stable word order (in contrast to English), involves a high degree of ellipsis (to the extent contrary to any languages known), and has a high degree of ambiguity with lexical items (again to the extent contrary to any languages known). The fact is, that Chinese speakers do not seem to have more trouble understanding each other than do speakers of English or other IndoEuropean languages. That mysterious question seems to be less mystical now,

50

Ping Li

given that Chinese speakers do have in stock a number of grammatical devices (though not always available, such as ba and bei), strategies to cope with variable word orders (default SVO and pragmatic cues to other orders), strategies to rely on semantic (such as noun animacy) and pragmatic information (context in which the sentence is uttered), strategies to integrate incomplete or ungrammatical constructions on the basis of complete and grammatical models, and strategies to rely on prior sentential context to resolve lexical-morphemic ambiguities. The use of these strategies and interpretations in real-time sentence processing suggests interactive processes in which multiple sources of information (i.e., syntactic, semantic, and contextual) compete and collaborate in sentence processing (MarslenWilson, 1987; McClelland, Rumelhart, and the PDF Research Group, 1986; MacWhinney and Bates, 1989). Crosslinguistic variations can reveal important aspects about sentence processing, clearly in unraveling the specific strategies used to cope with languagespecific properties, such as those associated with grammatical morphology, word order, ellipsis, and homophone ambiguities in Chinese. For example, to deal with the high degree of homophony, Chinese speakers rely strongly on contextual information in disambiguating various meanings in a homophone. To deal with the high degree of syntactic variability, they rely strongly on using fragment interpretations and semantic and pragmatic information; they depend less on syntactic or grammatical relations and show less sensitivity to grammatical violations. These processing patterns run at odds with traditional emphasis: linguistic theories have assigned a central role to grammaticality and syntactic structure. Meaningless sentences like "colorless green ideas sleep furiously" are used to demonstrate that it is possible for the speaker to judge the grammaticality and understand the structure of a sentence without retrieving its meaning (Chomsky, 1957). Our studies seem to suggest the opposite: Chinese speakers can retrieve the meaning of a sentence without heavily relying on grammatical relations or being disturbed by grammatical violations, and partial structures or fragments lend themselves to probabilistic interpretations based on complete and grammatical models in the language.

ACKNOWLEDGMENTS The preparation of this chapter was supported by a Faculty Research Initiation Grant from the School of Arts and Sciences, University of Richmond. Partial support for the preparation of Experiments 1 and 2 was provided by the Human Frontier Science Program to the author and by NIH/NIDCD (#R01-DC00216) to Liz Bates, and Experiments 3 and 4 by the Chinese University of Hong Kong to the author and Michael Yip. I would like to thank Liz Bates, Francois Grosjean, and Brian MacWhinney for insightful discussions on the materials presented here. I also thank Hua Liu, Yeung Lai Ping, Wong Wing, Edward Yang, and Michael Yip for their assistance to the experiments.

Variation and Sentence Processing in Chinese

51

REFERENCES Bates, E. (1991). On-line studies of sentence processing and the competition model. Unpublished manuscript, University of California, San Diego. Bates, E., and MacWhinney, B. (1982). Functionalist approaches to grammar. In E. Wanner and L. Gleitman (Eds.), Language acquisition: the state of the an. New York: Cambridge University Press. Bates, E., and MacWhinney, B. (1987). Competition, variation, and language learning. In B. MacWhinney (Ed.), Mechanisms of language acquisition. Hillsdale, NJ: Erlbaum. Bates, E., and MacWhinney, B. (1989). Functionalism and the competition model. In B. MacWhinney and E. Bates (Eds.), The crosslinguistic study of sentence processing. New York: Cambridge University Press. Bates, E., McNew, S., MacWhinney, B., Devescovi, A., and Smith., S. (1982). Functional constraints on sentence processing: A cross-linguistic study. Cognition, 11,245-299. von Berger, E., Wulfeck, B., Bates, E., and Fink, N. (1996). Developmental changes in real-time sentence processing. First Language, 16, 193-222. Chen, S., Tzeng, O., and Bates, E. (1990). Sentence interpretation in Chinese aphasia. In Proceedings of the Tenth Meeting of the Second Language Research Forum, Department of Linguistics, University of Oregon, Eugene, Oregon. Chomsky, N. (1957). Syntactic structures. The Hague: Mouton and Co., Publishers. Clynes, D., and MacWhinney, B. (1990). The ECS software. Department of Psychology, Carnegie Mellon University. Cohen, J., MacWhinney, B., Flatt, M., and Provost, J. (1993). PsyScope: A new graphic interactive environment for designing psychology experiments. Behavior Research Methods, Instruments, and Computers, 25, 257-271. Cotton, S., and Grosjean, F. (1984). The gating paradigm: A comparison of successive and individual presentation formats. Perception and Psychophysics, 35, 41-48. Frazier, L. (1987). Sentence processing: A tutorial overview. In M. Coltheart (Ed.), Attention and performance (Vol. XII). Hillsdale, NJ: Erlbaum. Frazier, L. (1990). Exploring the architecture of the language-processing system. In G. Altmann (Ed.), Cognitive models of speech processing. Cambridge, MA: The MIT Press. Grosjean, F. (1980). Spoken word recognition processes and the gating paradigm. Perception and Psychophysics, 28, 267-283. Grosjean, F. (1988). Exploring the recognition of guest words in bilingual speech. Language and Cognitive Processes, 3, 233-274. Ho, H.-H., and Jiang, Y.-H. (1994). Word frequency in Hong Kong in the 90's. Research Institute of Humanities, Chinese University of Hong Kong. Institute of Linguistics, The Academy of Social Sciences. (1985). Xianda Hanyu Cidian [Modern Chinese Dictionary]. Beijing: Commercial Press. Li, C., and Thompson, S. (1981). Mandarin Chinese: A functional reference grammar. Berkeley: University of California Press. Li, P. (1994). Understanding the time course of sentence comprehension: A sentence gating study in Mandarin Chinese. In H.-W. Chang, J.-T. Huang, C.-W. Hue, and O. Tzeng (Eds.), Advances in the study of Chinese language processing (Vol. I). Taipei: National Taiwan University Press.

52

Ping Li

Li, P. (1996a). The temporal structure of spoken sentence comprehension in Chinese. Perception and Psychophysics, 58, 571-586. Li, P. (1996b, November). Spoken word recognition of homophones in Chinese. Poster presented at the Thirty-seventh Annual Meeting of the Psychonomic Society, Chicago, 1996. Li, P. (1996c). Spoken word recognition of code-switched words by Chinese-English bilinguals. Journal of Memory and Language, 35, 757-774. Li, P., Bates, E., Liu, H., and MacWhinney, B. (1992). Cues as functional constraints on sentence processing in Chinese. In H. C. Chen and O. Tzeng (Eds.), Language processing in Chinese. Series of Advances in Psychology 90. Amsterdam: Elsevier Science Publishers. Li, P., Bates, E., and MacWhinney, B. (1993). Processing a language without inflections: A reaction time study of sentence interpretation in Chinese. Journal of Memory and Language, 32, 169-192. Li, P., and Yip, M. (1996). Lexical ambiguity and context effects in spoken word recognition: Evidence from Chinese. In G. Cottrell (Ed.), Proceedings of the 18th Annual Conference of the Cognitive Science Society. Hillsdale, NJ: Erlbaum. Li, P., and Yip, M. (in press). Context effects and the processing of spoken homophones. Reading and Writing: An Interdisciplinary Journal. Lu, J-M. (1980). Hanyu kouyu jufa-li -de yiwei xianxiang [On reversed sentence patterns in spoken Chinese]. Zhongguo Yuwen [The Chinese Language], 1, 28-41. MacWhinney, B., and Bates, E. (Eds.). (1989). The crosslinguistic study of sentence processing. New York: Cambridge University Press. Marslen-Wilson, W. (1987). Functional parallelism in spoken word-recognition. Cognition, 25, 71-102. Marslen-Wilson, W. D. (1990). Activation, competition, and frequency in lexical access. In G. T. Altmann (Ed.), Cognitive models of speech processing. Cambridge, Mass.: The MIT Press. Marslen-Wilson, W. D., and Welsh, A. (1978). Processing interactions and lexical access during word recognition in continuous speech. Cognitive Psychology, 10, 29-63. McClelland, J., Rumelhart, D., and the POP Research Group (1986). Parallel distributed processing: Explorations in the microstructure of cognition (Vol. II). Cambridge, Mass.: The MIT Press. Miao, X-C. (1981). Word order and semantic strategies in Chinese sentence comprehension. International Journal of Psycholinguistics, 8, 109-122. Moss, H., and Marslen-Wilson, W. (1993). Access to word meanings during spoken language comprehension: Effects of sentential semantic context. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 1254-1276. Onifer, W, and Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 9, 225-236. Seidenberg, M. S., Tanenhaus, M. J., Leiman, J. M., and Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 538-559. Simpson, G. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136.

Variation and Sentence Processing in Chinese

53

Simpson, G. B., and Krueger, M. A.(1991). Selective access of homograph meanings in sentence context. Journal of Memory and Language, 30, 627-643. Small, S., Cottrell, G., and Tanenhaus, M. (1988). Lexical ambiguity resolution: perspectives from psycholinguistics, neuropsychology, and artificial intelligence. SanMateo: Morgan Kaufmann Publishers. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645 - 659. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27, 324-340. Tyler, L., and Wessels, J. (1985). Is gating an on-line task? Evidence from naming latency data. Perception and Psychophysics, 38, 217-222.

NOTES 1

The sentence-role assignment task differs from sentence-processing tasks involved in several other models of sentence processing (e.g., Frazier, 1987) in which speakers read locally ambiguous structures in a sentence. For this reason, it was often difficult to directly compare our model with other models. Moreover, our studies have focused on the spoken modality whereas most other models focus on the visual modality; note that it is debatable whether visual and auditory processes are identical (Marslen-Wilson, 1990). 2 A fourth type of sentence was also tested in this experiment: sentences with the indefinite marker yi. (See Li et al., 1993, for details.) 3 For the NVN and NNV sentences, there was an initial bias toward first-noun choice. That is, upon hearing only the first noun, listeners found it easier to select the first noun as the agent than to select another noun that had not yet appeared. The bias was toward what has been heard, in the absence of an alternative form in the input. 4 A condition in which the visual probe occurred at the onset of the homophone was also tested originally. The results are not reported here because we suspect that the auditory signal of the homophone may cause interferences to listeners' naming of the visual probe that occurs at the onset of the homophone (also Grosjean, personal communication). 5 The sentence context variable here is comparable to the dominance variable in Experiment 3. Sentence context makes more sense than dominance in this experiment because of the addition of the neutral context that biased neither meaning of the homophone.

This page intentionally left blank

COMPREHENSION REPAIR IN THE PROCESSING OF A SHORT ORAL DISCOURSE INVOLVING A LEXICALLY AMBIGUOUS WORD GIYOOHATANO* KEIKO KUHARA-KOJIMA *Department of Human Relations Keio University Tokyo, Japan Department of Psychology Tokyo Woman's Christian University Tokyo, Japan

There are far fewer kinds of syllables in the Japanese language than in the European languages, because (a) no consonants other than a nasal coda come after a vowel, (b) consonants seldom cluster before a vowel, and (c) there are only five vowels. As a result, there are many homonyms in Japanese. This is especially true for Chinese loan words and invented words consisting of kanji (Chinese characters), because the Japanese language also has a much more restricted phonological inventory than the Chinese language. In Japan, about 2,000 kanji are designated as "kanji for daily use," but many of them share one and the same (Chinese) reading. For example, more than 70 of the 2,000 kanji have the Chinese reading of kou. Technical terms and other infrequently used words, most of which have been invented in the course of translating Western literature, can often be distinguished from their homonyms only by being written in kanji. It is hard to find the correct word or the right sequence of kanji for a given pronunciation, without considering semantic contexts. This can best be illustrated by describing how Japanese word-processing software operates. To use a contemporary word processor, one inputs every word or sentence segment in kana Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

55

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

56

Giyoo Hatano and Keiko Kuhara-Kojima

(syllabaries), which represent the pronunciation, and then hits the conversion key. When there are two or more alternatives, as is often the case, they are shown in the order of frequency of use or by some other measure of likelihood, and one chooses the right word by hitting the proper selection key. Let us take as an example the following message, which we think sounds natural as a somewhat formal sentence (e.g., used in a letter or left on a voice recorder): senseiga kikousareta kouzaha saikin kankousaremashita. The handbook you contributed to has recently come out.' This sentence includes five kanji compound words or word stems (underlined parts). When one enters them in isolation, the Japanese version of the Mac Write II software in the first author's personal computer displays 3, 10, 3, 2, and 9 alternatives of words or kanji combinations. Even when aspects of the syntactic structure of the sentence are considered (i.e., word stems are entered with their inflections after being segmented properly), the numbers of alternatives can only be reduced to 3 and 4 for the second and last selections. To choose the right alternative, semantic contexts have to be used skillfully. It should also be noted, however, that the target (correct) word appears first in four out of the five selections in the above message. This is mainly because the alternative that was chosen previously by the user comes to the top of the list of alternatives. The order of alternatives reflects how often and how recently each alternative has been chosen. It implies that, even if the receiver of a message is as mindless as the word-processing software, that person will be able to choose the right word if she or he has been exposed to many messages from the same sender and has given appropriate feedback. Likewise, when people listen to news or other messages about currently popular topics, they can find the correct word among those pronounced identically just by choosing the most salient alternative. In fact, we confirmed that college students could readily transcribe radio news or taperecorded editorial articles from newspapers if they could correctly recognize phonemes and properly segment their series (Kuhara-Kojima, 1997). In other words, they seldom chose homonyms of the target words.

1. STRATEGIES FOR REDUCING LEXICAL AMBIGUITY Recognizing a correct target word that has many homonyms is more difficult when people engaging in oral communication are just passers-by or in contact for the first time. Even when the listener is allowed to ask the speaker which word is referred to, she or he cannot do so very often because this interferes with smooth conversation. There are many situations in which individual listeners cannot question the speaker (e.g., when listening to a talk given through audio devices). In such cases, how can Japanese people reduce the amount of lexical ambiguity? We will first discuss speakers', and then listeners', strategies.

Comprehension Repair

57

A speaker uses at least three strategies (Kuhara-Kojima, 1997). First, she or he may offer a semantic context that triggers the correct word, even though it is a bit redundant. This attempt seems universal. For example, it is likely that an English speaker would say "a river bank" or "the bank near a river," to make clear that she or he is referring to "bank as a dike," and knows that it is less readily retrievable than its homonym, "bank as an institution to deposit money." The two other strategies are more or less peculiar to the Japanese language. A speaker may use, as the second strategy to avoid the homonymic miscontrual, a colloquial form that does not involve many Chinese loan words or kanji compound words. Cases where two or more different words share their pronunciation and context for use have survived in the Japanese language, probably because they can readily be distinguished when transcribed in kanji (Suzuki, 1975). Therefore, colloquial sentences that do not involve many kanji compound words are free from the problem of solving lexical ambiguity, as far as a listener is concerned. Third, people sometimes refer to how a target word is written in kanji when they want to specify it, even when they speak. Because most kanji have the Japanese as well as the Chinese reading, they can indicate the target word by referring to the Japanese reading of its component characters. For example, science and chemistry have the same pronunciation (kagaku) in Japanese and may appear in similar contexts. In this case, the best way to differentiate chemistry is to repronounce it by using the Japanese reading for the initial character (bakegaku). Experienced Japanese speakers often use these strategies unconsciously. When they are shown a pair of sentences and required to judge which one clearly conveys the intended meaning, they not only choose the one that relies on these strategies but also explain its advantage in terms of the strategies (Kuhara-Kojima, 1997). When a speaker does not use any of these strategies, how can a listener effectively process a discourse involving several lexically ambiguous words? We investigated this question using a series of psycholinguistic experiments, two of which will be described in the remaining part of this chapter.

2. REPAIRING COMPREHENSION: EXPERIMENT 1 In the first experiment, college students listened to a short oral passage, then they were asked to summarize it and were tested for their interpretation of an ambiguous word. The passage consisted of an initial part containing a target word that tended to be misconstrued as its homonym, and a later contextual sentence that induced inconsistencies with the initial homonym construal. The inconsistencies could be resolved completely by changing the initial interpretation of the ambiguous word. We investigated whether college students would repair their comprehension of the whole passage and change the homonym construal.

58

Giyoo Hatano and Keiko Kuhara-Kojima

2.1. Method 2.1.1. MATERIALS The passages used in this experiment, consisting of three sentences, were constructed as follows. After the initial sentence introducing the topic of the passage, the second sentence elaborated on the topic using an ambiguous target word that was likely to be construed as its homonym in this context. The last sentence (postcontextual sentence) differed, depending on whether the passage was incongruent or neutral. In the neutral passages, the postcontextual sentence was consistent with either the target-word interpretation or the homonym interpretation; therefore, it would not induce incongruity or an attempt to repair comprehension of the passage. In the incongruent passages, however, the last sentence provided an unambiguous context for the correct construal of the target word. When listeners interpret the ambiguous word incorrectly and then find their interpretation to be inconsistent with the postcontextual sentence, they must somehow repair their comprehension of the passage. In the ideal case, they change their interpretation of the ambiguous word in addition to reconstructing the representation of the entire passage. Three types of the target words were distinguished in terms of their likelihood of being interpreted as their homonym and the compatibility of that interpretation with the postcontextual sentence. The first type of target word has a salientcontradictory homonym. More specifically, according to our pilot survey with another group of college students, the corresponding homonym is much more readily retrieved than the target word for a given pronunciation (Table 1), but the TABLE 1 TARGET WORDS AND THEIR HOMONYMS Pronunciations or kana representation

Targets

Proportion target first"

Targets with a salient-contradictory homonym kouten stormy weather 3% reiguu warm treatment 3% Targets with a salient-irrelevant homonym hisho summering 5% shokki loom 0% Targets with a comparable-irrelevant homonym engei entertainment 35% ondo temperature 47%

Homonyms

good weather cold treatment secretary tableware gardening local dancing song

Proportion homonym firstb

92% 87% 95% 100%

60% 51%

a Proportion of the students who retrieve the target word first for a given pronunciation in the context of an ambiguous sentence. b Proportion of the students who retrieve the homonym word first for a given pronunciation in the context of an ambiguous sentence.

Comprehension Repair

59

homonym interpretation contradicts the indicated meaning in the postcontextual sentence, because its meaning is almost the opposite of the target's. The second type has a salient-irrelevant homonym. In this case, the homonym interpretation is highly likely, but is irrelevant to the postcontextual sentence because its meaning is completely unrelated to the meaning of the target. Thus, if the homonym interpretation is adopted, the passage may be taken to consist of two rather different parts. The third and last type of target word has a comparable-irrelevant homonym. That is, the homonym interpretation, which is comparable in likelihood to the target-word interpretation, is irrelevant to the postcontextual sentence. Two target words of each of these three types were used in this experiment. They are shown in Table 1. We had confirmed for each target word that, when the ambiguous and postcontextual sentences were presented simultaneously, more than 93% of another group of students chose a paraphrase of the target construal for the ambiguous sentence. Two examples of incongruent and neutral passages (with the target having a salient-contradictory or salient-irrelevant homonym) are shown in Table 2. 2.1.2. PROCEDURES College students were aurally presented with eight passages and required to summarize each of them in turn. They were administered the recognition test of the contents of the passages and the kanji-encoding test following the presentation and summarization of all eight passages. To conceal the purpose of the experiment, we wanted to avoid presenting many passages of the same type. Otherwise, we were concerned that listeners would find that their initial construal tended to be wrong, become unduly cautious, and adopt an unusual strategy (e.g., trying to maintain different interpretations of an ambiguous word). Thus two each of the incongruent and neutral passages were given with four filler passages, of which two did not involve any ambiguous word and two did not contain a postcontextual sentence. Therefore, the listeners had to revise their initial homonym interpretation and repair the comprehension of the whole passage twice, at most. Eighty-nine college students were randomly divided into three groups and assigned one of three sets of passages, constructed by combining the different types of passages and target words. The experiment was conducted in small groups of 6-8 students. We compared their performances for incongruent and neutral passages for each type of target word. 2.1.3. TESTS Students' representations of the whole passage and of the target word were assessed by three written criteria. The first was a summary of the passage, which

60

Giyoo Hatano and Keiko Kuhara-Kojima

was written just after listening to it. The summary had to be written in 40 characters or less, and within a time limit of 2 min. The second was the recognition test for the content of the sentence, including the ambiguous word. The students were asked to judge whether each of 14 sentences was semantically the same as had been given in one of the passages. Eight of them were semantically equivalent either to the second sentence with the targetword interpretation or to that with the homonym interpretation (of the ambiguous word) in the two incongruent and neutral passages (Table 2), and the remaining six referred to the filler passages. These sentences were aurally presented in a mixed order with an interval of 5 sec. TABLE 2 EXAMPLES OF PASSAGES AND SENTENCES FOR THE RECOGNITION TEST Example 1: Target with a salient-contradictory homonym Introductory sentence The weather of the Himalayas often changes every several days. Ambiguous sentence The weather forecast on the radio called for kouten (stormy weather or good weather) lasting for several days. Postcontextual sentence in the incongruent passage The mountaineers were forced to delay their departure, though they wanted to reach the mountaintop without break. Postcontextual sentence in the neutral passage Success or failure of the mountaineers depends on whether or not they can avoid stormy weather and make the best use of good weather. Sentences for the recognition test The weather forecast called for rough weather lasting for several days (a paraphrase of the target construal). The weather forecast called for fair weather lasting for several days (a paraphrase of the homonym construal). Example 2: Target with a salient-irrelevant homonym Introductory sentence A series of light articles on business topics featured in Newspaper H has been popular. Ambiguous sentence The topic of the next special series will be hisho (summering or secretaries) of presidents. Postcontextual sentence in the incongruent passage The series will focus on presidents who work out their plans at their own summer houses during vacation. Postcontextual sentence in the neutral passage This series will make pleasant reading among white-collar workers. Sentences for the recognition test The topic of the next special series will be lives of presidents in summer resort (a paraphrase of the target construal). The topic of the next special series will be human relationships between presidents and their assisting officers (a paraphrase of the homonym construal).

Comprehension Repair

61

In the third test, the kanji-encoding test, the students were given the second sentence of the critical four passages, including the target word transcribed in kana (syllabaries), with three other filler items, and asked to change the kana into kanji. A list of kanji characters having the pronunciation corresponding to each segment of the target word (e.g., kou in kouteri) was given to help the students. When there were many kanji sharing the pronunciation, we selected 10, including the correct one. The time limit was 2.5 min. 2.2. Results and Discussion We first classified the students' summaries into five categories: a summary based on the target-word interpretation (i.e., a summary using the target word or its paraphrase); a summary using the homonym (or its paraphrase); a summary using both the target word and its homonym; a general summary referring to neither the target word nor its homonym; and a summary that does not fit into any of the above categories. Because in most cases the college students wrote the target word or homonym in kanji, we could readily recognize which word they referred to. The distribution of the summaries by category for each type of target word is shown in Table 3. We also calculated an individual score for the summaries by assigning +1 to a summary using the target word, — 1 to a summary using the homonym, and 0 to all other summaries. The means and standard deviations for these values are also shown in Table 3. As can be seen, the means were clearly (and significantly) different between the incongruent and neutral passages. Our results show that when the homonym was salient (either contradictory or irrelevant to the postcontextual sentence), for neutral passages, as expected, many TABLE 3 FREQUENCY DISTRIBUTION OF TYPES OF SUMMARIES (EXPERIMENT 1) Types of summaries Passages (Score)

a

n

Target Homonym Both

Targets with a salient-contradictory homonym Incongruent 21 6 28 31 Neutral 6 28 Targets with a salient-irrelevant homonym 19 12 Incongruent 30 1 32 Neutral 28 Targets with a comparable-irrelevant homonym 31 41 4 Incongruent Neutral 30 15 12 a

General Others

M

SD

0 0

20 22

9 6

0.54 -0.71

0.91 0.99

2 0

26 22

1 1

0.23 -1.11

0.96 0.82

6 0

8 32

3 1

1.19 0.10

0.86 0.87

n = number of subjects; total number of frequencies = n X 2.

62

Giyoo Hatano and Keiko Kuhara-Kojima

more summaries using the homonym were generated than those using the target word, whereas the distribution was more or less even between these two types of summaries when the homonym was comparable in likelihood. For incongruent passages, a marked shift toward summarizing the passage using the target word or its paraphrase occurred. In other words, a large proportion of the students gave summaries congruent with the postcontextual sentence. However, there were not many summaries using the target word itself, even for incongruent passages. Next, we examined students' performances on the recognition and kanjiencoding tests. For the former, we classified the responses into the four possible patterns of OLD (i.e., included in one of the preceding passages)-NEW (i.e., not included in the passages) pairs, because two sentences were included in the test for each target word. (One sentence was consistent with the target-word interpretation and the other consistent with the homonym interpretation.) To run statistical analyses, we calculated the recognition score for individuals by giving +1 to OLD-NEW (only the sentence consistent with the target word interpretation was recognized), — 1 to NEW-OLD (only the sentence consistent with the homonym interpretation was recognized), and 0 to the two remaining patterns (i.e., OLDOLD in which both sentences are recognized as semantically equivalent to what the students had seen previously, and NEW-NEW in which neither sentence was recognized as semantically equivalent to what the students had seen previously), separately for each type of target word (Table 4). It can be seen that the proportion of OLD-NEW patterns was much larger for the incongruent passages than for the neutral passages. The recognition score differed significantly between these two types of passages. Students' answers on the kanji-encoding test were assigned as the target, the TABLE 4 FREQUENCY DISTRIBUTION OF RECOGNITION RESPONSE PATTERNS (EXPERIMENT 1) Response patterns Passages

na

OLD-NEW NEW-OLD OLD-OLD NEW-NEW

Targets with a salient-contradictory homonym 34 Incongruent 28 Neutral 7 31 Targets with a salient-irrelevant homonym Incongruent 30 37 Neutral 4 28 Targets with a comparable-irrelevant homonym Incongruent 31 30 Neutral 30 13 a

M

SD

16 43

4 6

2 6

0.64 -1.16

1.20 1.05

5 34

4 1

14 17

1.07 -1.07

1.06 0.84

18 27

4 4

10 16

0.39 -0.47

1.10 0.96

n = number of subjects; total number of responses = n X 2.

Comprehension Repair

63

TABLES FREQUENCY DISTRIBUTION OF kanji Encodings (Experiment 1) Passages

na

Targets

Targets with a salient-contradictory homonym Incongruent 28 34 Neutral 31 12 Targets with a salient-irrelevant homonym Incongruent 30 34 Neutral 28 2 Targets with a comparable-irrelevant homonym Incongruent 31 37 Neutral 30 26

Homonyms

Neither

21 48

1 2

25 54

1 0

24 34

1 0

a

n = number of subjects; total number of responses = n X 2.

homonym, or neither. The frequency distributions of these interpretations are shown in Table 5, revealing a clear difference between the incongruent and neutral passages. The proportion of the responses giving the target word interpretation for either the recognition or the kanji-encoding tests was 60% for the incongruent passages. If we use the more stringent criterion of both the OLD-NEW recognition pattern and the target-word transcription for attributing the target-word interpretation, the proportion diminished to 52, 39, and 40%, for the target words having salient-contradictory, salient-irrelevant, and comparable-irrelevant homonyms, respectively. Therefore, we examined the relationships between types of summaries and the construal of the ambiguous word for the incongruent passages. When the homonym was salient (contradictory and irrelevant combined), only nine summaries were written using the target word itself, and in all cases the ambiguous word was construed as the target on both tests. However, among the 31 summaries using a paraphrase of the target word, only 17 were accompanied by the target-word interpretation on both tests. When the homonym was comparable, 16 out of the 21 summaries involving the target-word itself were accompanied by the targetword interpretation, whereas eight of 20 paraphrasing summaries did so. We interpret the above findings as indicating that, whereas the students did modify their representation of the whole passage relying on the incongruent postcontextual sentence, they often did not change their initial interpretation of the ambiguous word. In other words, the processes of passage comprehension and interpretation of the target word, though related, may not be identical. It is possible to adjust one's comprehension of the ongoing discourse by using the latest information, without correcting the misconstrual of an ambiguous word presented earlier.

64

Giyoo Hatano and Keiko Kuhara-Kojima

3. RECOGNIZING INCONGRUITY AND REVISING THE INITIAL CONSTRUAL: EXPERIMENT 2 The results of the first experiment indicate that the college students could build a coherent representation covering the main elements of the whole passage, which could have served as the basis for future conversation (e.g., answering questions about the discourse), even when the passage included an ambiguous word that was likely to be misconstrued. Fewer than 20% of the students failed to incorporate the incongruent postcontextual information into the summary. At the same time, however, the results strongly suggest that a majority of the students failed to change their initial interpretations of the ambiguous word. In the second experiment we investigated whether college students would tend to recognize the incongruity induced by the postcontextual information and repair their comprehension of the whole passage without trying to discover where they had misconstrued. 3.1. Method The method used in this experiment was similar to that of the first experiment, with the same target words and homonyms, but several important features were changed. First, we used incongruent passages only, because here we were interested in how people would deal with incongruous information and because the effects of the postcontextual sentence had been demonstrated in the first experiment. Each student was presented with and asked to summarize three incongruent passages containing an ambiguous target word, the homonym of which was salient-contradictory, salient-irrelevant, or comparable-irrelevant. The choice and order of the target words were individually randomized. Second, we examined whether students would recognize that the postcontextual sentence was incongruent with the established representation of the passage based on the misconstrual of the ambiguous word, by observing directly whether they would look back at preceding sentences, as well as by asking later whether they had been puzzled during the presentation of the passage. For this reason, this experiment was conducted individually, and the passage was presented visually, by using cards on each of which a constituent sentence was printed. The ambiguous word had to be written in kana, so some other important words that would normally be written as kanji were also written in kana. Thus, this presentation was close to the oral one in the sense that only the pronunciation of critical words was given. Students were asked to read aloud each card in order, and turn it over when finished, but allowed to return to any earlier card whenever they wanted. Third, to clarify how students would repair their comprehension of the whole passage, they were required to summarize the passage before and after the contextual sentence. Each passage consisted of five sentences: the sentence involving an

Comprehension Repair

65

ambiguous word as the second or fourth sentence (a variable that has proven to be insignificant), the postcontextual sentence inducing incongruity with the homonym interpretation as the last one, and three other sentences that were neutral in terms of the construal of the ambiguous word. Compared with the first experiment, we expected the students in this experiment to be more likely to recognize the homonym construal of the ambiguous word as the source of the comprehension failure, and thus to revise it. This was because the ambiguous word was transcribed unexpectedly in kana, the students were allowed to read earlier sentences again, and all the passages were designed to induce incongruity. Thus, if students were unable to discover where they had misconstrued here, we could conclude that their counterparts in the first experiment would be even less likely to do so. Eighteen college students, who had not participated in the first experiment, served as subjects. After reading and summarizing the three passages, they were, also individually, questioned about any puzzlement or embarrassment that might have occurred during their reading. They then were indirectly asked about their interpretation of the ambiguous words (e.g., "What is the topic of the special series of articles in Newspaper H?"). If they used the target word ambiguously, they were required to paraphrase it. Finally, they were given a kanji-encoding test that was similar to the one used in the first experiment. The entire experimental session, which took 30-40 min for each student, was tape-recorded and transcribed. 3.2. Results and Discussion The 18 students generated 54 final summaries altogether. Of these, 52 were summaries that included the target word itself (44) or its paraphrase (8) and had incorporated the content of the last, postcontextual sentence. A typical one stated, "Newspaper H plans to take up hisho (summering or secretaries) as the next topic. It is about presidents who elaborate their plans during the summering period ..." Because the summaries were given orally, the students could use the ambiguous word without specifying their interpretation, either as the target or homonym. Considering the responses to the test and inquiries after reading and summarizing the passages, 9 of the 52 summaries were written based on the correct initial interpretation of the ambiguous word. The students reported that they had been puzzled for 36 (84%) of the remaining 43 summaries. Although judging coherence of a given text is believed to be difficult (e.g., Baker, 1984, 1985), this was not the case here, probably because the postcontextual information was clearly inconsistent with the homonym interpretation. Also, for 19 of the 43 passages, the students had returned to one of the earlier sentences before they started summarizing after the postcontextual sentence. They looked once again at either the last

66

Giyoo Hatano and Keiko Kuhara-Kojima

(postcontextual) sentence or the sentence involving the ambiguous word. These looking-back behaviors seem to exemplify the error-recovery heuristics listed by Carpenter and Daneman (1981), such as checking inconsistent information, and checking previous pieces of information selectively. However, only seven (16%) of the 43 summaries seemed to be written with the shift from the homonym interpretation to the target-word interpretation. For these, the students replied that they had recognized their misconstrual, and gave the target word consistently both in the summary and in the interview. For example, one student replied: "I thought hisho written in kana in the fourth sentence meant secretaries of presidents, and expected a special series on their jobs and others. I was a bit puzzled when I learned that the series would be concerned with leisure." A large majority (36 or 84%) of the 43 summaries were apparently well written, yet were judged as not involving the correction of the misconstrual of the ambiguous word, from the students' responses during the interview. These summaries were divided into three categories. The first category included 12 summaries, in which the students expressed strong incongruity, because their representations involved both the homonym interpretation and the postcontextual information, though they eventually wrote a modal summary without revising the misconstrual. They could readily revise their interpretation when some cue was given during the interview. For example, the student who offered the "typical" summary shown above admitted that its content was contradictory and stated that the next series would introduce secretaries of presidents. When the list of kanji was given at the fczn/z-encoding test, he realized that hisho must be summering, saying that he had been completely wrong. The second category involved eight summaries in which students tried to make the homonym interpretation and the postcontextual information compatible by producing some additional pieces of information. They retained the homonym interpretation during the test and inquiries. For example, one student explained that the passage was about presidents' consideration of how to treat or cope with secretaries during the summer vacation. Another student interpreted the passage about the mountaineers as including two kinds of forecast, the first for good weather, and the second, stormy weather, though the latter was not explicitly referred to. The third and last category included 16 summaries, in which students simply juxtaposed the homonym interpretation and the postcontextual information, without any sign of feeling incongruity or any attempt to reconcile their interpretations. In short, these results reveal that most of the students noticed inconsistencies, but maintained their initial interpretation of the ambiguous word, recognizing a need to resolve the incongruity somehow or actually resolving the inconsistencies by making additional inferences.

Comprehension Repair

67

4. ECONOMY OF LISTENERS' REPAIR STRATEGIES Through reading or listening to a passage, we build a representation of the microworld described in it. As we incorporate a new piece of information, we elaborate or revise this microworld (Collins, Brown, and Larkin, 1980), which is probably represented in intermediate memory (Tanaka, 1992) or long-term working memory (Ericsson and Kintsch, 1995). Although this process is basically incremental—a new piece of information is appended to the already constructed representation—this is not always the case. The new information may seem inconsistent with the old representation, requiring some attempt to resolve the inconsistency. One source of such inconsistency is the misinterpretation of a word or phrase. For example, the same sequence of characters or phonemes embedded in a context may stand for different words, but receivers of the message tend to adopt a particular interpretation that seems to fit best with the context (Hatano et al., 1990; Onifer and Swinney, 1981), irrespective of the modes of access they may adopt (Simpson, 1984). They seldom withhold their interpretation or maintain multiple interpretations beyond a few seconds, because doing so might unduly increase their information-processing load. Because humans are sensitive to contexts and thus good at selecting an interpretation, lexical ambiguity can be resolved promptly and without confusion in most cases (Bubka and Gorfein, 1989). However, people sometimes have to resolve the inconsistency produced by an earlier misconstrual. This must be true for every language, but the process of inconsistency resolution can readily be studied for the Japanese language, because it has so many words with homonyms. The results of the two experiments described here suggest that (a) Japanese people tend to accept a sequence of phonemes or characters as the most likely word in the context and often proceed without marking it as potentially having a homonym; (b) They often seem to recognize the inconsistency induced by their earlier misconstrual, but cannot identify the source of their misinterpretation; (c) When they find inconsistency between their mental representation of a passage constructed earlier and the information coming later, they repair their comprehension by modifying the representation, without changing the initial interpretation of the ambiguous word; and (d) Their attempts at comprehension repair seem to be constrained by two principles, the "new information priority principle" and the "minimal modification principle." It may appear amusing at first that some of the participants in our experiments, who had construed hisho as secretaries, added the new piece of information about presidents' ways of summering as an additional topic without revising the construal of the ambiguous word. However, such error-recovery heuristics are quite understandable in terms of economy of processing. In most discourse processes,

68

Giyoo Hatano and Keiko Kuhara-Kojima

paying maximal attention to the latest pieces of information is required for smooth conversation. For example, we are expected to respond to the latest, and not to previously presented, pieces of information. This can explain why the attempts at comprehension repair by the participants in our experiments were constrained by the new information priority principle. Our experimental results in general, and their interpretations in terms of the minimal modification principle in particular, are quite consistent with the minimalist hypothesis proposed by McKoon and Ratcliff (1992), which indicate that people establish locally coherent representations of parts of a text during reading. As in the oral presentation of the passage, where details of the earlier pieces of information used to construct the present representation of the whole passage would be lost or stored in a less retrievable part of long-term memory, the pieces of information had to be sought on the cards read much earlier when the passage was given in print. In both cases, it required less mental effort to try to recover coherence within the latest set of pieces of information. Even when the inconsistency could also be resolved by offering an auxiliary inference, it might still require less effort than looking back. In fact, adding a piece of new information by inference, such as inferring that the second weather forecast called for stormy weather, could be done without looking back or searching the possible source of error. Similar attempts have been observed in a series of studies by Ackerman (1982, 1984) and his associates (Ackerman et al., 1991) on comprehension repair in children as well as adults. Thus far we have discussed strategies used by Japanese people for reducing lexical ambiguities and for recovering from misconstruals in processing the Japanese language, which has many lexically ambiguous words when they are orally presented. How universal are these strategies? We would speculate, because these strategies are the natural products of human minds, they would be more or less universal, though people using different languages may differ in their skills for using the strategies. However, it would not be easy to run very similar experiments in those languages in which target words and their homonyms do not often appear in the same contexts. We believe that a contribution of crosslinguistic studies is to find a universal phenomenon that is more readily recognizable in some languages than others.

REFERENCES Ackerman, B. P. (1982). Children's use of contextual expectations to detect and resolve comprehension failures. Journal of Experimental Child Psychology, 33, 63-73. Ackerman, B. P. (1984). The effects of storage and processing complexity on comprehension repair in children and adults. Journal of Experimental Child Psychology, 37, 303-334.

Comprehension Repair

69

Ackerman, B. P., Jackson, M., and Sherrill, L. (1991). Inference modification by children and adults. Journal of Experimental Child Psychology, 52, 166-196. Baker, L. (1984). Children's effective use of multiple standards for evaluating their comprehension. Journal of Educational Psychology, 76, 588-597. Baker, L. (1985). How do we know when we don't understand? Standards for evaluating text comprehension. In D. L. Forrest-Pressley, G. E. MacKinnon, and T. G. Waller (Eds.), Metacognition, cognition and human performance (pp. 155-205). New York: Academic Press. Bubka, A., and Gorfein, D. S. (1989). Resolving semantic ambiguity: An introduction. In D. S. Gorfein (Ed.), Resolving semantic ambiguity. New York: Springer-Verlag. Carpenter, P. A., and Daneman, M. (1981). Lexical retrieval and error recovery in reading: A model based on eye fixations. Journal of Verbal Learning and Verbal Behavior, 20, 137-160. Collins, A., Brown, J. S., and Larkin, K. M. (1980) Inference in text understanding. In R. J. Spiro, B. C. Bruce, and W. F. Brewer (Eds.), Theoretical issues in reading comprehension (pp. 387-407). Hillsdale, NJ: Erlbaum. Ericsson, K. A., and Kintsch, W. (1995). Long-term working memory. Psychological Review, 102,211-245. Hatano, G., Kuhara-Kojima, K., and Saito, H. (1990). Comprehension monitoring and repair in the processing of a lexically ambiguous word and its embedding text. Japanese Cognitive Science Society Technical Report, No. 17. [in Japanese with an English summary] Kuhara-Kojima, K. (1998). Strategies for reducing the amount of lexical ambiguity by Japanese speakers. To be presented at the 28th Annual Symposium of the Jean Piaget Society, June 1998, Chicago. McKoon, G., and Ratcliff, R. (1992). Inference during reading. Psychological Review, 99, 440-466. Onifer, W., and Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory & Cognition, 9, 225-236. Simpson, G. B. (1984). Lexical ambiguity and its role in models of word recognition. Psychological Bulletin, 96, 316-340. Suzuki, T. (1975). On the twofold phonetic realization of basic concepts: In defense of Chinese characters in Japanese. In P.C.C. Peng (Ed.), Language in Japanese society (pp. 175-192). Tokyo: University of Tokyo Press. Tanaka, H. (1992). Discourse understanding and memory models. In Y. Anzai, S. Ishizaki, Y. Otsu, G. Hatano, and F. Mizoguchi (Eds.), Handbook of cognitive science (pp. 203-210). Tokyo: Kyoritsu-shuppan. [in Japanese]

This page intentionally left blank

AMBIGUITY OF REANALYSIS IN PARSING COMPLEX SENTENCES IN JAPANESE YUKIHIROSE* ATSU INOUE** *Program in Linguistics Graduate Center The City University of New York New York, New York Faculty of Economics Kanto Gakuin University Yokohama, Japan

1. INTRODUCTION The human sentence-processing mechanism is not an error-free device. It sometimes makes wrong decisions at choice points. When it does, a garden path can result. A garden path occurs when the parser selects an incorrect analysis at the ambiguity onset but can continue to proceed with that analysis in assigning structure to the subsequent input word string. The parser recognizes a garden path when it later discovers an input word that does not fit into the current analysis. As has been discussed in the recent literature (Inoue, 1991; Inoue and Fodor, 1995; Mazuka, 1991; Mazuka and Itoh, 1995; Yamashita, 1994; Yamashita, Stowe, and Nakayama, 1993; among others), processing relative clauses in Japanese inevitably causes a garden path in many cases. Japanese is a head-final SOV language. Therefore, in processing relative clause constructions, the relative clause is encountered before the head noun that it modifies. It is not always obvious to the parser whether an initial string of words such as (1), where the argument structure of the verb is consistent with the NPs that precede it, forms one single Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

71

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

72

Yuki Hirose and Atsu Inoue

clause ('Yoko saw the child at the intersection'), or whether instead there is a clause boundary inside that string as in (2). [Examples (l)-(3) are originally from Mazuka and Itoh, 1995.] (1)

Yoko-ga kodomo-o koosaten-de mikaketa -NOM child-ACC intersection-LOC saw

(2)

Yoko-ga [ ei kodomo-o koosaten-de mikaketa] onnanokoi-ni -NOM child-ACC intersection-LOC saw girl-DAT koe-o kaketa. called (to) 'Yoko called (to) the girl who saw the child at the intersection.'

We assume, following Inoue and Fodor (1995), that the parser makes its structural decisions incrementally left to right. Such a parser will first build a single clause, in accord with general principles such as Minimal Attachment (Frazier and Fodor, 1978). Only later, when the appearance of the head noun signals that the structure involves a relative clause construction, a garden path is detected and reanalysis is initiated. In most cases, no severe difficulty is perceivable at this point. There are, however, relative clause structures that cause severe processing difficulty for perceivers. This chapter addresses the question of why a certain type of relative clause is more difficult to process than another type. It will be argued that thematic ambiguity of the head noun increases the cost of the reanalysis process, contrary to existing views that the number or kinds of noun phrases (NPs) to be reanalyzed is the only deciding factor in assessing the cost of reanalysis. Section 2 introduces two types of relative clauses that are widely believed to show a contrast in difficulty, a contrast that can be predicted from two different perspectives, considered in section 3. Section 4 presents intuitive and experimental evidence for the "preference-based" account, supported by a recent frame-by-frame self-paced reading experiment reported in section 5. Finally, section 6 discusses the general implications of this experiment for reanalysis processing in Japanese.

2. EASY VERSUS DIFFICULT REANALYSIS Let us start by considering what revision processes are involved in obtaining a relative-clause analysis out of an initial single-clause analysis. Obviously, any revision has to locate the clause boundary at which the relative clause begins. The offset of the relative clause is syntactically marked by the presence of the head noun. In a relative clause such as the one we saw in (2) and the one in (3) below,

Ambiguity of Reanalysis

73

there has to be a syntactic gap, or an empty category, in an argument position coindexed with the head noun. In (2), the gap is in the subject position, and in (3) the gap is in the object position of the relative clause. (The subject empty category of the relative clause in [3] is an empty argument bound by the matrix subject.) Because in these examples all the argument positions are filled with NPs in the first-pass single-clause analysis, the parser has to make room for the gap by removing an NP.1 In Japanese there is no syntactic cue signaling the exact position and the role of the relative clause gap, whereas in English, for example, the gap position is usually clear from certain syntactic cues, such as the form of relative pronouns (for more discussion, see section 6). So the parser would have to guess the best position for the relative clause gap in Japanese. Our concern is the contrast exemplified by (2) (repeated below) versus (3), originally discussed by Mazuka and Itoh (1995). (2) Yoko-ga [ ei kodomo-o koosaten-de mikaketa] onnanokoi-ni -NOM child-ACC intersection-LOC saw girl-DAT koe-o kaketa. called (to) 'Yoko called (to) the girl who saw the child at the intersection.' (3) Yokoi-ga kodomo-o [ ei ei koosaten-de mikaketa] takusiii-ni noseta. -NOM child-ACC intersection-LOC saw taxi-DAT put on 'Yoko put the child into the taxi she saw at the intersection.' Mazuka and Itoh refer to relative clause structures such as (2) as involving noncostly reanalysis, in contrast to (3), which they claim involves costly reanalysis, and which they further suggest leaves an "impression of conscious reanalysis" (p. 305). They propose that reanalysis is not costly in (2) because only the subject NP has to be reanalyzed, whereas reanalysis in (3) is costly because both the subject NP and the object NP have to be reanalyzed. Based on Mazuka and Itoh's generalization, we will henceforth for convenience refer to structures such as (2) as SR (Subject Reanalysis) structures and to structures such as (3) as SOR (Subject and Object Reanalysis) structures, although for some theories (e.g., Gorrell, 1995, see section 3) such terms may not accurately describe the process by which a relative clause analysis is obtained. In this chapter we limit our discussion to the cases in which a single-clause analysis was initially pursued and reanalysis is initiated by the head noun that follows. The crucial question to be addressed is whether or not the SR versus SOR distinction correctly predicts the contrast in difficulty. If it does, we must then explain why an SOR relative clause is more difficult to arrive at than an SR relative clause; if not, we need to establish what would be the right characterization. Section 3 discusses some theories in the literature that are relevant for answering these questions.

74

Yuki Hirose and Atsu Inoue

3. PREVIOUS PROPOSALS Theories differ with respect to how they characterize the revision process for arriving at a relative clause structure. Some regard reanalysis as a matter of lowering material from the original clause to form a relative clause, leaving elements on the left in the now higher, original clause (Gorrell, 1995; Sturt and Crocker, 1996). Others portray reanalysis as expelling elements on the left into a higher clause to convert the original main clause into a relative clause (Inoue, 1991). Yet other theories assume parallel processing, whereby the parser computes several structures simultaneously; therefore, no reanalysis is needed at all, but only selection among the alternatives (Gibson, 1991; Gorrell, 1987). More importantly, these theories differ with respect to the status of what Mazuka and Itoh called "costly reanalysis." One perspective is that SOR structures are unpreferred and are therefore built only after the preferred alternative structure (i.e., SR) fails (Inoue, 1991; Sturt and Crocker, 1996, Frazier and Clifton, in press). The other perspective is that the parser is not capable of computing the SOR structure (Gibson, 1991; Gorrell, 1995). For Inoue (1991), the preference would be accounted for by the Minimal Expulsion strategy, (i.e., expulsion of the least number of arguments so that the resulting complex NP is well formed) (p. 138). Inoue's claim is that this follows from the parser's general preference towards minimal revision; the parser avoids unnecessary effort by not reprocessing any part of a sentence unless it has to. Sturt and Crocker (1996) assume a tree-lowering mechanism for reanalysis, which, in the case of relative clause reanalysis, can be described as pushing the right side of the original clause down to create a relative clause. They explain the preference for SR structures by reference to the point of reanalysis, that is, the point at which the tree-lowering applies. The SR structure is obtained when the parser applies lowering at the V node, which dominates both kodomo-o and the verb mikaketa, whereas the SOR structure is obtained by lowering the V node that only dominates the verb mikaketa. Sturt and Crocker propose a top-down initial search strategy for Japanese, as opposed to a bottom-up search strategy in English parsing. At the head noun in sentences such as (2) and (3), at which reanalysis is initiated, the parser searches downward for a node at which to apply lowering. The V node (dominating both the object NP kodomo-o and the verb mikaketa} is selected first because it is in a higher position than the lowest V node, and an SR structure such as (2) results. The correct SOR structure for (3) is obtained if lowering is applied at the V node, but this is tried only if the initial search strategy (i.e., lowering at the higher V') has failed. (See section 4 for a possible SR analysis at the head noun 'taxi' in [3].) Frazier and Clifton (in press) argue that the preference follows from a principle they call Minimal Revisions. Although Frazier and Clifton illustrate the revision process in the same manner as Sturt and Crocker (1996), it is not clear that they

Ambiguity of Reanalysis

75

assume tree-lowering as the reanalysis mechanism. Rather, they apparently assume that the subject NP (and the object NP) of the original clause are raised to a higher clause; in this sense, the operation is similar to what Inoue (1991) proposes. According to their Minimal Revisions Principle, revision to the SR structure is preferred because it is more minimal than revision to the SOR structure in terms of both the number of new nodes that are added and the number of lexical phrases that are reanalyzed. In the SR revision as in (2), one new node for the subject gap of the relative clause is added and the original subject NP (Yoko-ga) is raised, whereas in the SOR revision as in (3), two new nodes for the empty subject and the object gap of the relative clause have to be added, and both the subject NP (Yoko-ga) and the object NP (kodomo-o) have to be raised. Hence analysis (2) is predicted to be preferred over (3). Even though the explanations differ among the three theories, the common idea behind them is that in (3) the SR analysis is temporarily possible and this analysis is preferred to the SOR analysis before the ambiguity is resolved by the matrix verb. This view is crucially based on the presence of temporary ambiguity with respect to the revision of the tree. The SOR analysis is difficult because the parser ignores it in favor of the SR analysis. Hence these theories might be called "preference-based" accounts of reanalysis. For the other view, to be discussed below, temporary availability of the SR analysis is irrelevant for explaining the difficulty of the SOR structure. Gorrell (1995) assumes that the parser is subject to structural determinism. Although structural relations such as government and thematic role assignment can be altered during the course of processing, dominance and precedence relations are strictly deterministic. It follows that only lowering is a possible revision, and not all lowering is permitted. Under structural determinism, (3) is unprocessable because the dominance relation between the object NP kodomo-o and the VP that dominated it in the initial analysis is not preserved in the final SOR structure.2 The correct SOR analysis for (3) would be literally beyond the capacity of the structural determinism parser. By contrast, for (2), all the dominance and precedence relations that were established in the initial single-clause analysis remain unchanged after the SR structure is obtained (though additional relations have been added); thus there is no violation of structural determinism. Gibson (1991) assumes a parallel computation model. Several possible structures for a given input word string are assumed to be computed simultaneously. Since human memory resources are finite, there has to be a limitation on the extent to which such parallelism is maintained. Gibson proposed the Processing Load Unit (PLU) to evaluate memory load for each analysis. When one of the possible analyses is associated with two more PLUs than its best competitor, it is considered too costly for the parser to maintain, and thus the structure is rejected, or "pruned." If that rejected analysis later turns out to be the correct analysis, it cannot be recovered as part of the regular processing routine. An increase in PLUs

76

Yuki Hirose and Atsu Inoue

is associated with various factors. The PLU factor relevant to the current discussion is the one associated with an NP that has not received a thematic role from its thematic role assignor. Consider the point at which a word string such as (1), repeated below, has been received. (1) Yoko-ga kodomo-o koosaten-de mikaketa -NOM child-ACC intersection-LOC saw For the single-clause analysis, which is the simplest, there should no longer be a PLU associated with either Yoko-ga or kodomo-o, as all the thematic role requirements are satisfied at the verb mikaketa. For an SR structure as in (2), one PLU would be associated with Yoko-ga after mikaketa has been processed because its thematic role requirement is not satisfied by the lower verb mikaketa; the difference between the SR structure and the simplest analysis would be one PLU, and the parser would maintain both analyses. When the next noun is received, the simplest analysis of the string, as indicated in (1), is discovered to be wrong and is discarded, so then SR becomes the best structure. For an SOR structure such as (3), since neither Yoko-ga nor kodomo-o is an argument of mikaketa, the PLU associated with Yoko-ga and the PLU associated with kodomo-o would not be discharged by the verb. As a result, the correct SOR analysis for (3) would carry two more PLUs than the simplest analysis. Thus this competing SOR analysis would be pruned as soon as mikaketa is received, and so it does not survive; when all other analyses are ruled out by subsequent lexical items, the parse fails. To summarize, preference-based theories predict that SOR resolutions of SR/ SOR temporary ambiguities are more difficult to process than SR resolutions because the wrong SR analysis attracts the parser on the first attempt of reanalysis. By contrast, according to Gorrell's structural determinism and Gibson's parallel computation theory, the SOR analysis is impossible to compute regardless of whether the SR analysis is temporarily available or not. Section 4 presents some evidence for the former approach and against the latter approach.

4. EFFECT OF THEMATIC AMBIGUITY OF THE HEAD NOUN Native speakers report that an SOR sentence such as (3) becomes notably easier with the following minimal change, shown in (4). (4) Yokoi-ga kodomo-o [ ei ei hiroba-de mikaketa] burankoi -ni noseta. name-NOM child-ACC square-LOC saw swing-DAT put on 'Yoko put the child onto the swing she saw at the square.' In (4), the head noun taxi is replaced with swing, and in order to maintain plausibility, intersection is replaced with square. The reason (4) is easier than (3)

Ambiguity of Reanalysis

77

seems to have something to do with the characteristics of the head noun. In (3), as Mazuka and Itoh noted, the head noun taxi can be interpreted not only as THEME of the verb see, but also as AGENT, metaphorically meaning the driver of the taxi. Therefore taxi as the head noun does not uniquely signal what role it should play in the relative clause; the head noun can be construed as associated with either the subject gap (with AGENT role) or the object gap (with THEME role) in the relative clause. The head noun does not provide definitive information regarding what kind of relative clause the reanalysis routine should construct. On the other hand, swing can never be an AGENT of the verb see. If the parser is sensitive to such information about possible thematic roles, it can tell at this point that both the subject and the object will have to be reanalyzed. In other words, there is an ambiguity even at the reanalysis stage in (3) but not in (4), if thematic information is used. In the rest of this section we will present experimental evidence confirming these native intuitions and discuss its implications for the theories discussed in section 3. The intuition that an SOR sentence is easier or harder depending on the possible thematic roles of the head noun was confirmed in a whole-sentence reading time experiment reported in Hirose and Chun (in press). The following three types of sentences, all involving a relative clause, were tested. (5) a. Yamaokai-ga kakusiisan-o [ eiej anotekonotede sagasidasita] -NOM hidden fortune-ACC after great effort discovered mogurino bengosij-ni yamunaku azuketa. unlicensed lawyer-DAT unwillingly entrusted 'Yamaoka unwillingly entrusted his hidden fortune to the unlicensed lawyer who he discovered after great effort.' b. Yamaokai-ga kakusiisan-o [ ei ej anotekonotede sagasidasita] -NOM hidden fortune-ACC after great effort discovered mogurino kasikinkoj-ni yamunaku azuketa. unlicensed safe-DAT unwillingly entrusted 'Yamaoka unwillingly entrusted his hidden fortune to the unlicensed safe that he discovered after great effort.' c. Yamaoka-ga [ ei kakusiisan-o anotekonotede sagasidasita] -NOM hidden fortune-ACC after great effort discovered mogurino bengosii-ni yamunaku ayamatta. unlicensed lawyer-DAT unwillingly apologized 'Yamaoka unwillingly apologized to the unlicensed lawyer who discovered his hidden fortune after great effort.'

78

Yuki Hirose and Atsu Inoue

The sentence in (5a) is disambiguated by the final matrix verb azuketa ('entrusted') as an SOR sentence, in which both the first NP Yamaoka-ga and the second NP kakusiisan-o have to be matrix clause arguments. The head noun bengosi ('lawyer') has to be interpreted as THEME of the embedded verb sagasidasita ('discovered'), but its role is ambiguous when it is first encountered because it is animate and so might be either AGENT or THEME of 'discovered.'3 Sentence (5b) has the same SOR structure as (5a), but the head noun is not ambiguous; the inanimate noun kasikinko ('safe') can only be interpreted as THEME of the embedded verb. Sentence (5c) is identical to (5a) except that it is disambiguated in the opposite direction, by the matrix verb ayamatta ('apologized'), which does not take an accusative (NP-o) argument. Hence, the first NP Yamaoka-ga is a matrix clause constituent, but the second NP kakusiisan-o must be retained in the relative clause. The head noun in (5c) thus has to be interpreted as subject (AGENT) of the embedded verb 'discovered.' However, like (5a), the sentence is temporarily ambiguous when the animate head noun is first encountered, and is not disambiguated until the matrix verb. Hirose and Chun report that for two types of SOR structure, the whole-sentence reading time was significantly lower when the head noun did not allow an AGENT interpretation for the embedded verb, as in (5b), than when the head noun could be either the AGENT or the THEME of the embedded verb, as in (5a). The (5b)-type (SOR, head unambiguous) sentences were also read faster than the (5c)-type (SR, head ambiguous) sentences, although this difference was not significant. Interestingly, a comparison of the SR/SOR minimal pair provided some slight support for Mazuka and Itoh's generalization that SR-type sentences such as (5c) are easier than SOR-type sentences such as (5a) when the head nouns are both ambiguous; (5c)-type sentences were read faster than (5a)-type sentences, though the difference was not significant. The experiment thus confirmed native intuitions for these sentences and also found evidence for the existence of an easy SOR-type sentence. At least, the difficulty attributable to the SOR analysis is less than that due to ambiguity. This finding poses a serious problem for those theories which claim that SOR is not merely unpreferred, but inherently problematic. Gorrell (1995) and Gibson (1991) predict difficulty of the SOR sentence independent of any information in the head noun. In Gorrell's model the only factor discriminating difficult from easy revisions is the nature of the structural change between the initial analysis and the final analysis. In Gibson's model, a pruning decision is made at the verb before the head noun. Hence the presence or absence of ambiguity in the head noun should have no effect on processing difficulty in either model. For preference-based theories, by contrast, the existence of easy SOR sentences can be accommodated. SOR sentences can be easy when the head noun is unambiguous because there are no chances for the parser to adopt the incorrect SR analysis at the head noun (and hence no need for reanalysis at the end of the

Ambiguity of Reanalysis

79

sentence). Therefore, in the sentences with an unambiguous head noun, such as (5b), the SR analysis would be filtered out at the noun position, for example, by thematic considerations, and the SOR analysis would be adopted instead. The head noun ambiguity could affect the processing speed either at the site of the ambiguity or at a later disambiguation point in the sentence, or possibly at both positions. The ambiguity effect would show up at the disambiguation point as a result of a garden path. SOR sentences with an ambiguous head noun would be difficult because the parser would have been led down the garden path, assuming that the parser prefers the SR analysis at the head noun. The analysis would therefore have to be revised again at the matrix verb. The matrix verb is thus the point at which the difficulty would manifest itself. A difficulty might also show up at the site of the ambiguity, reflecting the ambiguity resolution process itself. SOR sentences with an ambiguous head noun may be difficult at the noun position because the parser may need some time to make a choice between the alternatives.4 The preference-based theories may or may not be compatible with this account, which predicts difficulty at the site of ambiguity, depending on whether the ambiguity resolution process at the ambiguous head noun can allow a period of indecisiveness before adopting the SR analysis. Sturt and Crocker's model does not seem to be able to assume a slow decision at the ambiguous head noun. In Sturt and Crocker's model the parser first attempts to lower a node at each point in a top-down search. Then it evaluates the plausibility of the interpretation of the resulting structure possibly by thematic considerations.5 When the interpretation is implausible, it is ruled out and the parser attempts to lower the next lower node in the tree. That is, the thematic role information is used only after the structural reanalysis (by a tree-lowering operation). Therefore the fact that two roles (AGENT and THEME) are available at the head noun in ambiguous SOR sentences should have no or very little effect on the ambiguity resolution time, because the parser is not choosing between the two roles but merely checking if a plausible role is available to be adopted for the analysis already chosen. Therefore, Sturt and Crocker's model seems to predict that processing proceeds even more smoothly at an ambiguous head noun than at an unambiguous head noun that goes against the parser's preference induced by its top-down search strategy. That is, there could be a processing disadvantage for the unambiguous head noun in an SOR sentence. But there should never be a disadvantage for the ambiguous head noun even if the cost may be too small to measure experimentally. From the whole-sentence reading time experiment reported in Hirose and Chun, it is impossible to tell at which point in the sentence the processing is slowing down, and therefore it is difficult to decide which of the above accounts is correct. In the experiment reported in section 5, the same set of structure types (5a-c) were examined with a frame-by-frame self-paced reading task.

80

Yuki Hirose and Atsu Inoue

5. FURTHER EXPERIMENTAL EVIDENCE A frame-by-frame self-paced reading task was employed to measure processing load at different positions of the sentences in question. The three types of sentence, comparable to (5a-c), were divided into eight positions (P1-P8), illustrated in (6a-c) below. P1 through P4 were identical across sentence types. The positions of interest in this experiment were P5, the head noun; P6, an adverbial or prepositional phrase that is unambiguously a matrix clause constituent serving as "padding" (to allow the experiment to catch possible spillover effects); and P7, the matrix verb. P8 was a maru, the Japanese equivalent to the English period symbol. P1 P2 P3 P4 roojin-ga I sutego-o I atikoti kiite I sagasita old man-NOM orphan-ACC by asking around looked for

(6)

a.

P5 P6 P7 P8 / bokusi-ni I tootoo I takusukotonisita I ° priest-DAT finally decided to entrust 'The old man finally decided to entrust the orphan to the priest who he looked for by asking around.'

b.

P5 P6 P7 P8 / sisetu-ni I tootoo I takusukotonisita I ° orphanage-DAT finally decided to entrust 'The old man finally decided to entrust the orphan to the orphanage which he looked for by asking around.'

c.

P5 P6 P7 P8 / bokusi-ni I tootoo I aukotonisita I ° priest-DAT finally decided to meet 'The old man finally decided to meet the priest who looked for the orphan by asking around.'

5.1. Method 5.1.1. SUBJECTS Forty-five college students from Shoin Women's University and Daito Bunka University in Japan participated in the experiment. All were native speakers of Japanese. 5.1.2. MATERIALS AND DESIGN The experimental sentences were constructed as intersecting pairs, forming triplets. There were 24 such triplets. One pairing within each triplet took two

Ambiguity of Reanalysis

81

sentence types requiring an SOR analysis, differing with respect to thematic ambiguity of the head noun. For the ambiguous type (e.g., [6a]), an animate head noun created an AGENT / THEME ambiguity with respect to the lower clause verb (or an EXPERIENCER / THEME ambiguity where the relative clause used a "psych" verb). For the unambiguous type (e.g., [6b], an inanimate head noun could only be interpreted as THEME in the relative clause). An orthogonal pairing took sentence types which both had the head-noun thematic role ambiguity, and which differed with respect to whether the relative clause was eventually disambiguated to the SOR or the SR analysis by the matrix verb, (e.g., [6a] and [6c]), respectively. At each position, length was matched across sentences in a triplet, length being measured in number of characters (both kanji and hand). The experimental sentences were distributed in a counterbalanced design to form an experiment in three versions. An additional 66 sentences varying in length and structure were constructed as fillers and interspersed among the experimental items, so that two experimental items were never presented adjacently. The order of presentation of sentences was identical across the three versions. For all sentences used in the experiment, comprehension questions were prepared, to which a YES/NO response was required. These questions were used to ensure that sentence reading was appropriately accurate. 5.1.3. PROCEDURE Each sentence was presented frame-by-frame, in Japanese script. Each frame was displayed in the fixed position on a computer screen. Presentation of a sentence was initiated when the subject first pressed the return key on a standard computer keyboard labeled YES. Each time the YES key was pressed, a new frame replaced the previous display. The key press following the final display (maru), initiated display of a question about the content of the sentence, to which the subjects were instructed to respond using either the YES key or the tab key, which was labeled NO. The time between the onset of presentation of any frame and the key press initiating the next frame was recorded by the computer's internal clock, as reading time. 5.2. Results Table 1 shows the mean reading time for the four positions of interest (i.e., P5 through P8), for all sentence types. The data for the two SOR types to be discussed in the next paragraph are also presented in Figure 1. The comparison between SOR ambiguous and SOR unambiguous sentences, (6a) versus (6b), showed that at P5, the head noun,6 thematic role ambiguity led to longer reading times, F 1 4 2 ) = 6.89, p < .05, F2(l,21) = 4.37, p < .05. At

82

Yuki Hirose and Atsu Inoue TABLE 1 READING TIMES PER FRAME (ms)a P5 P6 P7 P8 HeadNP AdvP Matrix V maru

SOR, ambiguous SOR, unambiguous SR, ambiguous

989 912 977

663 680 708

814 767 792

512 460 490

a NP, noun phrase; AdvP, adverbial phrase; SOR, Subject and Object reanalysis; SR, Subject reanalysis.

Figure 1. Data for the two Subject and Object Reanalysis (SOR) types.

the following PP/AdvP position, P6, no difference in reading times was found, F < 1 for both subjects and items analyses. At the matrix verb position, P7, SOR sentences with an ambiguous head noun yielded longer reading times than those with an unambiguous head noun, F1(1,42) = 6.68, p < .05; F2(l,21) = 17.21, p < .001. A difference in the same direction was also found at the maru frame, P8, F1(1,42) = 7.06, p < .05; F2(l,21) = 7.26, p < .05. In the orthogonal comparison over SOR and SR sentences with role-ambiguous head nouns, (6a) versus (6c), the outcome is not conclusive. Under the materials design for this experiment, these pairs were identical except for the matrix verb frame (P7). Therefore, the positions of interest were P7 and P8.7 Figure 2 shows the data at the matrix verb position (P7) and the maru position (P8) for all three sentence types. At the matrix verb frame (P7), where the ambiguity between the SR and SOR analyses is resolved, no significant difference between SR and SOR structures emerged, nor was any significant difference evident at the sentence-final maru frame, F < 1 in all analyses. However, as Figure 2 shows, in terms of numerical value, the reading time at these positions for the head-ambiguous SR sentence is

Ambiguity of Reanalysis

83

Figure 2. Reading times (ms) for head-ambiguous and unambiguous Subject and Object Reanalysis (SOR) and Subject Reanalysis (SR).

shorter than that for the head-ambiguous SOR sentence; it falls between the ambiguous SOR and the unambiguous version. Overall, the results were consistent with Hirose and Chun's whole-sentence reading time experiment. We further found that the longer RT for the wholesentence reading for the ambiguous SOR sentences compared to unambiguous sentences that was found in Hirose and Chun can be attributed not only to the final matrix verb position, but also to the head noun position.

6. DISCUSSION An effect of head noun ambiguity was found both at the head noun and at the matrix verb in the comparison between the two SOR types, (6a) and (6b). We discuss the effect at the head noun first and the effect at the matrix verb second. 6.1. Ambiguity Effect at Head Noun The crucial finding of the present experiment is the difference at the head noun: ambiguous head nouns exhibited longer reading times than unambiguous head

84

Yuki Hirose and Atsu Inoue

nouns. It seems that the head-noun thematic ambiguity interfered with the decision process and delayed the decision. This suggests that it is not the parser's primary strategy to base its decision solely on structural preference. If it had made a structure-based decision and chosen one analysis immediately, the thematic ambiguity of the head noun would have had no effect on reading times. This result is also not consistent with any account which presumes that it is only after the structural preference strategy applies that the parser responds to a mismatch of head noun to verbal thematic role information. Sturt and Crocker's model would predict greater reanalysis cost and hence longer reading times for processing the unambiguous head noun, because it conflicts with the parser's structural preference (see section 4), but this is contrary to the experimental facts presented here. It appears, then, that no structure-based decision is made, or at least it is delayed until after thematic roles have been consulted. How could this be so? Let us make two assumptions: (i) it is the thematic processor that is primarily responsible for making decisions about the position and thematic role of the relative clause gap, and (ii) the thematic processor makes its decision in parallel with or prior to any decision by the structural processor. These two assumptions are essential to the explanations of the decision delay considered below. Before discussing the explanations two comments on these assumptions about the thematic processor are in order. The active use of thematic information in processing Japanese relative clauses may partly be attributed to the syntactic properties of Japanese. In Japanese relative clauses, there often is no syntactic cue available for determining where the gap occurs (as has been seen in [2] and [3]) and what its thematic role is. By contrast, in English, there are various syntactic cues that signal the position and the role of the gap in relative clauses, as illustrated in (8a-e) below. (8) a. b. c. d. e.

The boyi I looked for ei this morning The boyi who ei looked for me this morning The boyi whom I looked for ei this morning The foresti which we couldn't see ei The foresti where we couldn't see ei

The examples include stranded prepositions (8a) and (8c), and morphologically marked relative pronouns (8c)-(8e). In addition, the gap position is usually obvious because the overt argument is missing in that position, because English does not allow phonologically empty discourse-bound pronouns (whereas Japanese does). This observation about English and Japanese relative clauses suggests that it is natural for the Japanese parser to use any information available in deciding filler-gap relations in relative clauses. Among the available information, thematic information is likely to be the most informative. It is important to note that our proposal about the earlier (prior to structure-

Ambiguity of Reanalysis

85

based decision) use of thematic information in the resolution of the SR/SOR ambiguity is compatible with the modularity thesis. A standard view of the human sentence processing mechanism as a modular system gives precedence to structural factors over semantic (hence thematic) factors on first-pass parsing. On the revision pass, however, it would not be controversial to assume that thematic information has already become accessible (see Carlson and Tanenhaus, 1988). We believe that it is possible for the parser to make immediate use of such information in reanalysis to decide on the best thematic role of the head noun—and consequently the best position for the gap in the relative clause—without overstepping the bounds of a modular parsing mechanism. We now consider two possible explanations for the delay of a structure-based decision. The assumptions (i) and (ii) mean that an analysis that requires the head noun to carry a semantically incompatible or pragmatically implausible thematic role in the relative clause (via the gap bound to it) is rejected by the thematic processor regardless of any structural preferences. This predicts that in the case of a sentence with an unambiguous head noun, such as (6b), the SR analysis is rejected by the thematic processor because the head noun sisetu ('orphanage') is not a possible AGENT. In the ambiguous case, as in (6a), the thematic processor cannot make a decision on its own because the head noun bokusi ('priest') is suited to satisfy the thematic role of AGENT or of THEME. Unless it has some means of choosing between these two roles, only a structural preference could decide which analysis to proceed with. (We assume that it has a ranking of thematic roles, in the second explanation discussed below.) A period of indecisiveness might consume some time before the structural preference dictates. So the first explanation is the fact that the thematic processor lacks a preference between AGENT and THEME results directly in the longer reading times at the site of ambiguity. There is another way indecisiveness could cause the delay of decision at the site of ambiguity. The parser could be indecisive in choosing between the two analyses because there are two different preferences: an SR analysis preference and an SOR analysis preference, which come from various factors. We will consider what factors would yield each preference. Let us begin with the SR preference. Suppose there is a predetermined ranking for the thematic roles of the verb, and the thematic processor chooses the role that is ranked highest among the available roles. We will assume that the thematic role ranking is similar to the order of semantic roles in terms of salience discussed by Comrie (1981). If there is an AGENT, it is ranked highest; an EXPERIENCER and a CAUSE may be next; and then THEME might be ranked next. In (6a) the thematic processor will choose AGENT over THEME. This will result in the SR preference, because the assignment of AGENT role to the gap ousts the NP-ga. We now consider factor(s) that would favor the SOR analysis. One factor concerns the antecedent relation between the matrix subject and the embedded subordinate clause subject. In the SR analysis the subject gap is bound to the relative

86

Yuki Hirose and Atsu Inoue

head noun. In the SOR analysis the subject gap is controlled by the matrix subject NP (and the object gap is bound to the head noun).8 Hence, any preference that favors control of the subject gap by the matrix subject will interfere with the SR preference. What would motivate matrix subject control of the subject gap in the relative clause? We speculate that matrix subject control is a general tendency the parser often abides by in subordinate clause analysis. Consider the examples in (7). (7) a. roojin-ga sutego-o atikotikiite sagasita old man-NOM orphan-ACC by asking around looked for b. roojirii-ga [ ei sutego -o atikotikiite sagasita] to itta. old man-NOM orphan-ACC by asking around looked for COMP said 'The old man said he looked for the orphan by asking around.' c. roojini-ga [ ei sutego -o atikotikiite sagasita] toki naita. old man-NOM orphan-ACC by asking around looked for when cried 'The old man cried when he looked for the orphan by asking around.' d. ei [roojinj-ga sutego -o atikotikiite sagasita] old man-NOM orphan-ACC by asking around looked for to itta I toki naita. COMP said / when cried 'ei said / cried when the old man looked for the orphan by asking around.' The word string in (7a) may either be analyzed as a single clause, 'The old man looked for the orphan by asking around,' or as the initial portion of a sentence containing a complement clause as in (7b), or an adverbial toki-clause as in (7c). Native intuition suggests that when the presence of a subordinate clause becomes clear, the typical reanalysis is to revise the initial single-clause analysis into a structure in which the matrix subject controls the subject gap in the subordinate clause, as in (7b) or (7c), though there are other ways of reanalyzing the initial analysis (e.g., insert a phonologically empty discourse-bound pronoun in the matrix subject position and keep the initially analyzed single clause as an embedded complement or adverbial clause, as in [7d]). It is possible that this general tendency for subordinate clause reanalysis comes into effect at the relative clause head noun in (6a). Notice this preference for subject control of a subject gap appears to be a tendency to favor a parallel structure, in that the subjects of the two clauses refer to the same entity (though their thematic roles in each clause might differ, for example, between AGENT and EXPERIENCER). Significantly, in a relative clause such as (6a) the head noun is nonsubject in the matrix clause, and the gap bound to it is also nonsubject in the relative clause (e.g., bokusi 'priest' is the indirect

Ambiguity of Reanalysis

87

object in the matrix clause and the gap bound to it is the direct object in the relative clause). This parallelism seems to give some sense of ease in interpreting the sentence. To sum up, an SOR preference (at the ambiguous head noun in (6a)) could result from a general subordinate clause reanalysis strategy that favors matrix subject control of a subject gap or parallel interpretation of matrix and embedded clauses. There is another factor that might cause an SOR preference. It concerns how strongly the two NPs of a clause are connected when they are initially encountered. All principles and strategies discussed in section 3 favor breaking the initial connection between NP-ga and NP-o (e.g., roojin-ga and sutego-o, in [6]) in reanalysis rather than the connection between NP-o and the verb (e.g., sutego-o and sagasita, in [6]). (In either case the connection between NP-ga and the verb has to be broken.) This might seem reasonable because there is no grammatical dependency between the two NPs and, accordingly, it might be less costly to break the connection between them than to break the grammatical relation between NP-o and the verb. Fodor and Inoue (in preparation), for instance, define minimal revision in terms of minimizing the number of grammatical relations to be revised. However, we suspect that the parser assumes a certain relation between the two NPs (NP-ga and NP-o) on-line and has some tendency to maintain this relation in reanalysis. Assuming a relation between two grammatically nonrelated NPs may seem odd in parsing languages such as English, in which the verb usually precedes most of its argument (and adjunct) NPs, and hence NPs are attached to the verb as they are received on the first pass. In Japanese, however, the verb follows its NPs. Thus a full-attachment parser that works left to right is forced to attach an NP to the preceding NP. Connecting the two NPs to each other on the first pass might be helpful or even necessary for forming an expectation about the type of verb before the verb is received. Through such anticipation about the verb, a "pseudogrammatical relation" between the two NPs may be established. For example, the parser assumes one NP is the agent of an event, in which the other NP is the theme, though the exact nature of the event is not known yet. If such a pseudorelation is assumed by the Japanese parser, and it is costly to break that relation later, it could favor the SOR analysis in which both NPs end up in the matrix clause. To take stock of the factors that would cause the two incompatible preferences, the SR analysis will be preferred by a thematic processor that selects a gap position on the basis of a predetermined thematic role ranking; the preference for the SOR analysis could be a result of a general subordinate clause reanalysis strategy in which the matrix clause subject controls the subordinate clause subject, and also possibly a result of maintaining a grammatical relation between NP-ga and NP-o. Furthermore, we have considered two explanations for the ambiguity effect at the head noun. The first explanation assumed a lack of any basis for deciding

88

Yuki Hirose and Atsu Inoue

between the two possible thematic roles for the gap and argued that the thematic processor's indecisiveness results in longer reading times at the ambiguous head noun. The second explanation assumed a conflict between competing SR and SOR analysis preferences and attributed the longer reading times at the ambiguous head noun to the difficulty of resolving the conflict. It is not possible to determine, on the basis of the current experimental data, which one of the two explanations is the right one. The two explanations, however, make different predictions for unambiguous SR sentences: sentences with a relative head noun that can only be an AGENT but not a THEME, and hence clearly disambiguates the sentence to the SR analysis. The first explanation makes crucial use of the presence of the head noun thematic role ambiguity. Therefore, it will predict no difficulty for the unambiguous SR sentences at the head noun, since there will be no conflict of thematic roles and accordingly the thematic processor should be able to immediately assign the AGENT role to the gap. The second explanation, on the other hand, implies that whenever the thematic processor yields the SR analysis, regardless of whether it is the highest ranked analysis in the thematic role ranking or whether it is the only analysis available, this creates a conflict with the SOR preference and results in a period of tension in the parser while the conflict is resolved. So for unambiguous SR sentences, this explanation predicts that there will be no difference in difficulty at the noun position between an ambiguous SR/SOR head noun and an unambiguous SR head noun.9 6.2. Garden-Path Effect at the Matrix Verb The differences due to ambiguity found at the matrix verb and at the maru position remain to be explained. That is, the ambiguous SOR sentences exhibited longer reading times at both positions than the unambiguous SOR sentences. This result is most naturally interpreted as a garden path effect, in the sense that it reflects the parser's surprise on encountering an input that is not compatible with its current analysis. (However, note that it is not a garden path in quite the usual sense, because it is not a consequence of a misanalysis on the first pass but is a consequence of a wrong reanalysis choice.) For sentences with a thematically ambiguous head noun, it appears that the SR analysis is selected after the period of indecisiveness. When this is discovered to be an incorrect analysis at the matrix verb, reanalysis takes place and longer reading times result. For sentences with an unambiguous head noun, the SR analysis is excluded at the relative head noun and therefore no later garden path occurs. Alternatively, it might be suggested that the effect found at the matrix verb and at the maru position (longer RT for head-ambiguous SOR condition than in headunambiguous SOR condition) is the same indecisiveness effect as the one observed at the head noun, discussed in section 6.1. That is, it might be thought that the parser's indecision persists through the sentence end, carrying over the inde-

Ambiguity of Reanalysis

89

cisiveness from the head noun position, or carrying out parallel computations of both possible analyses in the ambiguity condition. However, we believe that this is not the case, because the same indecisiveness effect was not found at the adverb position intervening between the head noun position and the matrix verb position. The fact that the reading time difference disappears at the adverb and reemerges later suggests two different phenomena occurring. The severity of the SOR garden path is somewhat in doubt because the comparison between the ambiguous SOR sentences (e.g., [6a]) and the ambiguous SR sentences with the same (ambiguous) head noun (e.g., [6c]) failed to show a significant difference in reading times at either the matrix verb or the mam position, though in terms of numerical value the ambiguous SOR sentences showed longer reading times than the ambiguous SR sentences. The data for all three sentence versions at these positions is shown in Figure 2. The lack of a strong contrast between the SOR and SR ambiguous forms can be seen as due to the fact that the SR version is not as easy as the unambiguous version; it seems to show a weak garden-path effect also. The most natural interpretation of the intermediate status of the SR version is that on some trials it is the SOR analysis that is adopted by the parser, though the SR analysis is adopted more often. Thus each results in a garden path some of the time. This partially split preference is perhaps not too surprising in view of its being the outcome, as suggested above, of a temporary disagreement between preference principles. Some support for this interpretation comes from the off-line experimental data from Hirose and Chun (in press) on the resolution of the SR/SOR ambiguity. This study found very little bias towards the SR analysis (a nonsignificant difference), with a great deal of variation across subjects and materials. In short, our account of the data for all positions is that the reaction time (RT) difference at the head noun is a result of the parser's temporary indecision, which ceases at the subsequent position; the RT differences at the matrix verb and the maru positions are due to a weak garden path when the sentence ending enforces the SOR analysis, as a consequence of the SR analysis more often being chosen by the parser at the head noun. We still have to explain why the parser tends to favor the SR analysis after its period of indecision. The two views presented below are only speculative and the current data do not choose between them. One possibility is that the thematic processor's preference in accord with the thematic role ranking is the parser's ultimate choice. Earlier in this section we argued for the active use of thematic information and accordingly assumed the crucial role of the thematic processor in relative clause reanalysis in Japanese. In view of this it would not be unreasonable to suppose the eventual dominance of the thematic processor's preference over the preference stemming from nonthematic factors, such as the subject control preference and the pseudogrammatical relation factor, even though the thematic preference seems to experience an

90

Yuki Hirose and Atsu Inoue

initial period of being impeded by those factors. The alternative possibility is that the structure-based preference following from parsing strategies, such as the Minimal Expulsion or the Minimal Revisions Principle, comes into effect after the period of initial thematic indecision and influences the final decision. One question about structural preference as the final decision maker is why is its impact delayed and not influential earlier in the reanalysis process? Our speculation is that it is because this is reanalysis: the parser attempts something different in the reanalysis pass than in the first pass, where structural decisions take precedence over thematic factors. To conclude, our aim has been to investigate the ambiguity resolution process in relative clause reanalysis and to determine what kinds of factors differentiate easy and difficult relative clauses in Japanese. The experimental results presented here led us to draw the following conclusions. Thematic considerations play a prominent part in resolving ambiguity at the reanalysis stage, taking the place of (or at least taking precedence over) structural preference principles. It is the presence of thematic ambiguity in the head noun that makes SOR relative clauses difficult to process. When the thematic role of the head noun in the relative clause can be uniquely determined by the thematic processor, SOR sentences do not seem to present any great challenge to the human sentence processing routines.

ACKNOWLEDGMENTS This is part of ongoing research toward the doctoral dissertation of the first author. A preliminary version of this work was presented at the Tenth Annual CUNY conference on Human Sentence Processing (Hirose, 1997a) and will appear in Hirose (in press). The work is greatly indebted to Janet Fodor for her invaluable comments and suggestions. We also thank Dianne Bradley for her comments on an earlier draft of this work. Many thanks to Kazuhiko Kakehi, Naoomi Kuratani, Yukiko Miyazaki, Taira Shimazu, Yasuhiro Shirai, Mitsunobu Toda, Hillel Weintraub and others, to whom we are indebted for their assistance in conducting the research.

REFERENCES Carlson, G., and Tanenhaus, M. (1988). Thematic roles and language comprehension. In W. Wilkins (Ed.) Thematic relations. New York: Academic Press. Chomsky, N. (1986) Barriers. Cambridge, MA: MIT Press. Comrie, B. (1981). Language universals and linguistic typology. Chicago, IL. The University of Chicago Press.

Ambiguity of Reanalysis

91

Fodor, J. D., and Inoue, A. (in preparation). Motives for Reanalysis. Unpublished manuscript. Frazier, L., and Clifton, C. (in press). Sentence reanalysis and visibility. In F. Ferreira and J. D. Fodor (Eds.), Reanalysis in sentence processing. Boston, MA: Kluwer Academic Press. Frazier, L., and Fodor, J. D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6,1-34. Gibson, E. A. F. (1991). A computational theory of human linguistic processing: Memory limitations and processing breakdown. Doctoral dissertation, Carnegie Mellon University. Gorrell, P. (1987). Studies of human syntactic processing: Ranked-parallel versus serial models. Doctoral dissertation, University of Connecticut. Gorrell, P. (1995). Japanese trees and the garden path. In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing. Hillsdale, NJ: Lawrence Erlbaum. Hirose, Y. (1997a). Relative clause reanalysis and head noun ambiguity. Paper presented at the Tenth Annual CUNY Conference on Human Sentence Processing. Hirose, Y. (1997b). The relationship between structural preference and thematic information in reanalysis: Evidence from Japanese. Poster presented at AMLaP-97, Edinburgh, UK. Hirose, Y. (in press). What makes reanalysis difficult? In M. Ryan (Ed.) CUNYForum (vol. 20). New York: The City University of New York. Hirose, Y. and Chun, S. (in press). Attachment ambiguity in head final languages. In N. Akatsuka (Ed.), Japanese and Korean Linguistics (Vol. 7). CSLI, Stanford, CA: Stanford University. Hirose, Y., and Inoue, A. (in preparation). Two thematic roles of the relative head noun: Their effect on interpretation of relative clause constructions. The City University of New York. Huang, C.-T. J. (1984). On the distribution and reference of empty pronouns. Linguistic Inquiry, 15, 531-574. Inoue, A. (1991). A comparative study of parsing in English and Japanese. Doctoral dissertation, University of Connecticut. Inoue, A., and Fodor, J. D. (1995). Information-paced parsing of Japanese. In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing. Hillsdale, NJ: Lawrence Erlbaum. Mazuka, R. (1991). Processing of empty categories in Japanese. Journal of Psycholinguistic Research, 20, 215-232. Mazuka, R., and Itoh, K. (1995). Can Japanese speakers be led down the garden path? In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing. Hillsdale, NJ: Lawrence Erlbaum. Sturt, P., and Crocker, M. (1996). Monotonic syntactic processing: A cross-linguistic study of attachment and reanalysis. Language and Cognitive Processes, 11(5). 449-494. Yamashita, H. (1994). Processing of Japanese and Korean. Doctoral dissertation, Ohio State University. Yamashita, H., Stowe, L., and Nakayama, M. (1993). Processing of Japanese relative clause constructions. Japanese and Korean Linguistics, 2, 248-263. CSLI, Stanford University, Stanford, CA.

92

Yuki Hirose and Atsu Inoue

NOTES 1

For other types of relative clauses, in which a gap can be created in an adjunct position and hence no reanalysis of the preceding NP is necessary, or in which the relative clause structure is predicted by the mismatch between the argument structure of the verb and the preceding NPs, see Yamashita et al. (1993). 2 In fact, the SOR structure could be obtained without violating structural determinism by lowering the V0 node, not the VP node, even though Gorrell does not consider this as a possibility in his discussion. (See also Sturt and Crocker, 1996, footnote 29, for discussion on this point.) If lowering is applied at the V° node, the SOR structure is no longer an impossible structure for the parser to build, but then an independent account for why it is less preferred to SR will be needed. 3 In (5a) and (6a), in which both the subject and the object of the relative clause are empty, it may be possible to interpret the subject gap to be the relative clause gap, and the object gap to be an empty pronoun coindexed with the matrix object. However, such reading is not readily obtainable to all speakers and even for those speakers who can get such reading the degree of accessibility varies among individual sentences. We do not discuss this possibility of further ambiguity in this chapter, but we will discuss it in relation to the potential effect of the interaction between the thematic role of the head noun in the main clause and its role in the relative clause in general in Hirose and Inoue (in preparation). 4 Actually, Hirose and Chun's data may be compatible with some noncommittal parsing models since the data do not show where in the sentences the parsing difficulty occurred. For example, a parallel parsing model could account for the finding that ambiguous SOR sentences are read slower than the unambiguous ones if it is assumed that the reading time difference reflects the cost of maintaining multiple analyses. It is then predicted that the cost associated with the ambiguous sentences may be found throughout the ambiguous region (from the head noun position though the adverb position). A delay model, which does not compute any structure until the ambiguity is resolved might find some cost for the ambiguous sentences only at the disambiguation point. The experimental data in section 5 will show that these predictions are not borne out. 5 Sturt and Crocker refer to the example below, from Mazuka and Itoh (1995). Hirosi-ga aidoru kashu-o kakusita kamera-de totta. -NOM popular singer-ACC hid camera-with photographed 'Hirosi photographed the popular singer with the camera which he was hiding.' Mazuka and Itoh maintain that the above sentence is easy even though both the subject and the object NPs have to end up in the higher clause, because of the implausibility of "hiding" a popular singer. This implies that the parser does not even commit to the singleclause analysis. Thus, we do not consider this sentence an example of what we define as "SOR" sentences, because our SOR classification entails an initial single-clause misanalysis. Sturt and Crocker's analysis, however, entails that the parser first commits to the single-clause analysis and then responds to the oddness produced by the SR analysis (i.e., 'the camera that hid the popular singer'). 6 Reaction times (RTs) for the set of ambiguous head nouns and unambiguous head nouns used in the experiment were measured in isolation in a post hoc test, using an additional 13 subjects. In this test, the subjects saw the nouns on the computer screen, each

Ambiguity of Reanalysis

93

followed by another noun, either relevant or irrelevant in meaning. The subjects were asked to judge if the two words they saw were related in meaning. The RTs to the first nouns, not the performance of the judgment task, were of interest. The average RTs for the ambiguous nouns and the unambiguous nouns were not significantly different (668 ms and 673 ms, respectively, F < 1). 7 As predicted, the comparison did not yield a reliable difference at the head noun position (P5) or at the AdvP/PP position (P6). 8 We use the term control loosely to indicate an antecedent relation. It is unclear whether such a preference is purely syntactic and hence called subject control, or thematic and hence called AGENT control, since control relations might be at the thematic level. We take no stand on this here. Also, it is not important here whether the empty category in the subject position in the subordinate clause is PRO or pro, (Chomsky, 1986) and whether the relation to the matrix subject is controlled in the narrow sense, or "generalized control" (Huang, 1984) or merely pronominal coreference. 9 In Hirose (1997b), for the two SR structures shown below, no ambiguity effect was found at the head noun, (i.e., the head nouns that are ambiguous between AGENT and THEME of the relative clause verb and the head nouns that are unambiguously AGENT did not yield a significant difference in reading time). Although different accounts may be possible, the data seem to be more consistent with the second explanation outlined above: as long as the AGENT interpretation of the gap is available, there is conflict between the SR (AGENT) preference and the SOR preference, regardless of the presence of ambiguity. SR structure, head noun temporarily ambiguous between AGENT and THEME of 'look for': Makoto-ga zansinna hihyoo-o koogi-no ato sagasiteita kyooju-ni -NOM innovative review-ACC after class was looking for professor-DAT ikinari dekuwasita. suddenly bumped into 'Makoto suddenly bumped into the professor who was looking for the innovative review after class.' SR structure, head noun can only be AGENT of 'write': Makoto-ga zansinna hihyoo-o koogi-no ato kaiteita kyooju-ni -NOM innovative review-ACC after class was writing professor-DAT ikinari dekuwasita. suddenly bumped into 'Makoto suddenly bumped into the professor who was writing the innovative review after class.' Twelve sentence pairs were tested on thirty-two subjects. The reading times for the head nouns were 782 ms and 794 ms, respectively, F < 1 for both subject and item analyses.

This page intentionally left blank

THE PROCESSING OF EMPTY SUBJECTS IN ENGLISH AND JAPANESE TSUTOMU SAKAMOTO * MATTHEW WALENSKr *Faculty of Letters Kyushu University Fukuoka, Japan *Department of Linguistics University of California, San Diego La Jolla, California

1. INTRODUCTION The present chapter is concerned with the mechanism by which the parser identifies the correct antecedent of a gap. In particular, we will discuss the particular construction involving the case of a subject gap in control structures (PRO, within the generative literature). We take a crosslinguistic perspective, in keeping with the themes of this book, and compare the results of experiments from both English and Japanese. In this chapter, we pursue this idea and propose a "theta-checking strategy" that can explain the universality of antecedent identification in control structures, as well as differences between English and Japanese concerning the processing of these empty subjects. Essentially, this strategy claims that information is used by the parser as soon as it becomes available. This follows from universal characteristics of the parser, with the acknowledgement that the time course of information availability, as well as the salience of particular cues, differs from language to language. In sections 2 and 3 we will review the experimental literature that has been put forward for these constructions in English and Japanese, Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

95

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

96

Tsutomo Sakamoto and Matthew Walenski

respectively. Section 4 compares the results obtained from the two languages, and acknowledges criticisms of earlier work in this area. Section 5 examines a proposal from the linguistic literature regarding the identification of PRO, which is incorporated into the parsing strategy outlined in section 6. Section 7 discusses the types of parsers that could be used to explain these findings.

2. PROCESSING OF EMPTY SUBJECTS IN ENGLISH One type of gap that has been identified within the generative tradition (cf. Chomsky, 1982) is PRO, which appears as the subject of nonfinite clauses in English. However, one question that has yet to be answered, both within grammatical and parsing theories, is how the antecedent of PRO is identified. Within a generative framework, one analysis for these structures is that a given verb will assign control of PRO to one of its arguments (Jackendoff, 1972), provided that PRO is uniquely identified with one of these arguments (this has been termed "obligatory control").1 One verb class that has been identified in this respect assigns control to its subject (subject control), whereas another type of verb assigns control to its object (object control). In other cases, when the verb does not uniquely specify an argument to be the controller of PRO, the situation is less clear (this has been termed "nonobligatory control" by Williams, 1980, and will not further concern us here). Frazier, Clifton, and Randall (1983) (see also Clifton and Frazier, 1986) tested English control structures to determine the strategy used by the parser to identify the antecedent of the subordinate subject. Consider the following examples from Frazier et al.: (1) Recent Filler (Subject control) Sentence Everyone liked the woman who1 the little child2 started [PRO2 to sing those stupid French songs for tracei last Christmas]. (2) Distant Filler (Object control) Sentence Everyone liked the woman who1 the little child forced trace1 [PROi to sing those stupid French songs last Christmas]. In the recent filler sentences, (1), the verb start, a subject control verb, assigns its subject the little child to be the controller of PRO. Notice that of the two possible fillers, the woman and the little child, it is the little child that is closer (by number of intervening words) to the subject gap. Hence, the actual filler is also the more recent filler. For the distant filler sentences, (2) above, the verb force assigns its object the woman to be the controller of PRO. In these sentences, the actual filler is the more distant of the two possible fillers. Frazier et al. hypothesize

Processing Empty Subjects

97

TABLE 1 RESULTS OF FRAZIER ET AL.'S EXERIMENT 1 Sentence type Distant filler Recent Reaction time (ms) "Got it" (%)

1165 66

1071 78

filler

Difference

94 (p < .05) -12(p<.01)

that if temporal recency is a factor in the identification of the antecedent, so that a more recent filler is preferred over a more distant one, then sentences for which the recent filler is also the actual filler will produce faster reaction times (RTs) in a comprehension "got-it" task than sentences for which the recent filler is not the actual filler. The experimental results by Frazier et al. (given in Table 1) showed that distant filler sentences required more processing time and produced more errors than the recent filler sentences. Frazier et al. explained these findings with what they called the Most Recent Filler (MRF) strategy, which is stated as (p. 196): The MRF strategy: During language comprehension a detected gap is initially and quickly taken to be co-indexed with the most recent potential filler.

Frazier et al. imply that the MRF strategy applies only when the parser does not have reliable information about the correct filler for a gap. In the absence of lexical control information, the MRF strategy is claimed to apply. This strategy assigns the nearest potential filler to a gap. This initial choice by the parser is later overridden by lexical information supplied by the verb if necessary. It is this errorcorrecting procedure that causes the longer processing time in the distant filler sentences. Furthermore, Frazier et al. imply that the parser does not recognize a gap as a possible filler for another gap (this is termed the "Lexical Filler Only" hypothesis in Sakamoto, 1991,1996). This account has two important consequences: first, lexical control information is not immediately available, and second, that in the meantime the parser relies on a heuristic strategy based on temporal recency. Consistent with this finding, Nicol (1988) found, using an on-line cross-modal priming technique, that identification of the missing subject of a subordinate clause (PRO) was slow and inaccurate. Specifically, she found that immediately following the gap in such a sentence, the controller of the empty subject has not yet been identified.2 Though Nicol makes no specific claims concerning a recency strategy, her findings suggest that there is no immediate commitment on the part of the parser to identify an antecedent. So although control information may be delayed, there may be no strategy used by the parser in the meantime.3

98

Tsutomo Sakamoto and Matthew Walenski

3. PROCESSING OF EMPTY SUBJECTS IN JAPANESE Working in Japanese, Sakamoto (1991,1995a, 1995b, 1996) conducted a series of experiments to test the MRF hypothesis. Japanese offers two advantages over English. First, verbal control information is necessarily delayed, since Japanese is a verb-final language. Second, Japanese allows scrambling of arguments, so that either argument of a control verb may be put in a recent filler or distant filler position. Sakamoto tested both subject control (3a) and object control (3b) verbs using nominalized subordinate clauses. Subjects were instructed to listen to each sentence and respond by naming the person who is going to be in Tokyo. An example of each type of sentence is given in (3): (3) a. Subject control Tosio-ga ototoi Junko-ni [PRO Tookyoo ik]-o Tosio-NOM day before yesterday Junko-DAT Tokyo going-ACC tegami-de hakuzyoosita. letter-by confessed. 'The day before yesterday, Tosio confessed to Junko by mail that he would go to Tokyo.' b. Object control Tosio-ga kissaten-de Junko-ni [PRO Tookyoo ryokoo]-o Tosio-NOM cafe-at Junko-DAT Tokyo travel-ACC hakkirito meireisita. clearly ordered 'At the cafe, Tosio clearly ordered Junko to travel to Tokyo.' The results of experiment 1 (given in Table 2), in which experimental sentences have "subject-object" order, showed that object control sentences were easier to process than subject-control sentences (666 — 607 ms = 59 ms, p < .05). The TABLE 2 RESULTS OF SUBJECT-OBJECT ORDER EXPERIMENT IN JAPANESE Sentence type S-control O-control Reaction time (ms) Consistency (%)

666 88.9

607 90.2

Difference 59 (F( 1,22) = 6.23, p = 0.021) - 1.3 (F(l,22) = 0.491, p= 0.491)

Processing Empty Subjects

99

objects in this experiment were also the more recent fillers. Thus, the results of this experiment are compatible with the hypothesis that the MRF strategy applies to Japanese control structures. However, another possible explanation is that the parser prefers to assign control to an object initially. A second experiment by Sakamoto was designed to compare these two hypotheses. In this experiment, the order of the subject and object NPs was switched, so that in the object-control sentences the object was the distant filler, and in the subject-control sentences the subject was the recent filler. The materials are given in (4) below: (4) a. Subject control Junko-ni ototoi Tosio-ga [PRO Tookyoo iki]-o Junko-DAT day before yesterday Tosio-NOM Tokyo going-ACC tegami-de hakuzyoosita. letter-by confessed. 'The day before yesterday, Tosio confessed to Junko by mail that he would go to Tokyo.' b. Object control Junko-ni kissaten-de Tosio-ga [PRO Tookyoo ryokoo]-o Junko-DAT cafe-at Tosio-NOM Tokyo travel-ACC hakkirito meireisita. clearly ordered 'At the cafe, Tosio clearly ordered Junko to travel to Tokyo.' If the MRF strategy is correct, then the results should be the opposite of the previous experiment. If object control is preferred, then the results should come out the same as in the previous experiment. The results of experiment 2 (given in Table 3), in which experimental sentences have "object-subject" order, revealed that not the most recent lexical filler but the object was preferred as a controller, even when it was the distant lexical filler (749 - 648 ms = 101 ms,p < .05). TABLE 3 RESULTS OF OBJECT-SUBJECT ORDER EXPERIMENT IN JAPANESE Sentence type S-control O-control Reaction time (ms) Consistency (%)

749 84.4

648 90.8

Difference 101 (F(l,16) = 7.93, p = 0.0124) -6.4 (F(l, 16) = 3.01,p = 0.102)

100

Tsutomo Sakamoto and Matthew Walenski

The findings of the above two experiments indicate that the processor favors object controllers regardless of their surface positions in the sentence. The results of this work indicate that recency is not the deciding factor, but in fact that there is a preference for the parser to choose object control initially. On the other hand, one possible way to maintain the MRF strategy would be to admit the existence of a movement trace in the "object-subject" sentences, which could be recognized by the parser as a possible filler. However, this conflicts with Frazier et al.'s assumption that only lexical fillers may be recognized by the parser (Lexical Fillers Only hypothesis). See Sakamoto (1996) for a full discussion of the consequences of abandoning this hypothesis to maintain the MRF strategy for both English and Japanese.

4. COMPARISON OF ENGLISH AND JAPANESE Now, let us consider again the English and Japanese examples relevant to our discussion, beginning with Frazier et al.'s materials (repeated here as [5] and [6]): (5) Recent Filler (Subject control) Sentence Everyone liked the woman who1 the little child2 started [PRO2 to sing those stupid French songs for trace1 last Christmas]. (6) Distant Filler (Object control) Sentence Everyone liked the woman who1 the little child forced tracel [PRO1 to sing those stupid French songs last Christmas]. Grain and Fodor (1985) (see also Ford and Dalrymple, 1988; Fodor, 1988; and Nicol, 1988) noted that Frazier et al.'s distant filler sentences contain two adjacent gaps, whereas the recent filler sentences have only one gap following the verb. They claim that the recent filler sentences are faster because they are easier. That is, processing one gap is easier and takes less time than processing two gaps. A second possible explanation (Fodor, 1988) has to do with the differences in thematic function of the controller between the matrix and subordinate clauses. In a subject-control sentence (recent filler) like (5), the subject of the matrix clause is also the subject of the subordinate clause, so no revision of function is necessary. In an object-control sentence like (6), the matrix object becomes the subordinate subject; it is this revision of function that is hypothesized to lead to the observed difficulty. Another way to look at it is that the extracted element does not have the same function in the two types of sentences. For a recent filler (subject-control) sentence like (5), the extracted NP the woman forms a filler-gap dependency that is independent of the dependency created by PRO and its controller. However, for a distant filler (object control) sentence like (6), the extracted element is involved

Processing Empty Subjects

101

with both the filler-gap dependency created by its extraction, as well as with the control relationship necessitated by PRO, as in this case the extracted element is also the controller. Keeping in mind the three possible explanations mentioned above, let us consider the cases in Japanese. For the Japanese experiments, four sentence types were used. These consisted of a subject-control verb and an object-control verb in an SOV sentence frame, and these same verb types again in an OSV sentence frame. These are shown below as (7a/b) and (7c/d), respectively. (7) a/b. Taroo-ga Hanako-ni [PRO Tookyoo iki]-o yakusokusita/susumeta. c/d. Hanako-ni1 Taroo-ga trace1 [PRO Tookyoo iki]-o yakusokusita/susumeta. The Japanese comparison of "subject-object" and "object-subject" order reflects the English in that there is a difference in the number of gaps involved. However, this difference in Japanese is not confounded with respect to subject or object control as it is in English. Both SOV and OSV word-order types are given in the Japanese experimental items in both subject- and object-control constructions. Therefore, even if double-gap constructions are more difficult to parse, this difficulty will affect the amount of time it takes for the parser to recognize the controller in Japanese in both control constructions equally. What is left to be explained is why, for Japanese, both one gap (subject-object order) and two gap (object-subject order) constructions prefer object control rather than subject control when both a subject and object are present. We explain this by assuming that in both Japanese and English, object control is preferred as an initial choice by the parser. It therefore becomes plausible to assume that although the Japanese data picked up an object controller preference (comparing subject-control to object-control constructions), the Frazier et al. data for English picked up the difficulty of processing the double-gap constructions in objectcontrol structures (compared to the single-gap subject-control constructions), and did not directly reflect the processing of the control dependency. Nakayama (1996) reports on the interpretation of empty subjects in embedded clauses by Japanese children. The experimental results showed that empty subjects with matrix object interpretation are more accurately identified than those with matrix subject interpretation. This result may support the hypothesis that there is a universal tendency for object control. Note that there is no a priori reason why object control is preferred. Even in English, however, there are some observational facts that exhibit object preference. First, for verbs that allow both a subject and an object, the majority are object control. Verbs such as promise, that both allow an object and show subject control, are very rare. Also, the class of verbs that allow either subject or object control (such as expect), allow subject control only if the (optional) object is not present; otherwise object control is the rule. Second, there is some evidence from

102

Tsutomo Sakamoto and Matthew Walenski

language acquisition facts (C. Chomsky, 1969; Cohen, 1983, 1987) that also suggests a preference for object control in English. Frazier et al. claim that the "recency" preference is one instance of a more general "saliency" preference. If recency is the salient factor in English and object control is the salient factor in Japanese, then this saliency explanation could account for both the English and Japanese experimental results reported here. As Gorrell (1995) correctly mentions, this line of argument would be plausible from a broad cognitive viewpoint. However, it is important to explain where this saliency comes from (i.e., how this general cognitive saliency is realized in our linguistic knowledge). This issue will be considered in the following sections.

5. THEMATIC HIERARCHY If we admit that object control is universally preferred, the next question is why this is so. We will consider this issue in terms of thematic relations. Nishigauchi (1984) claims that a Goal role is chosen as the controller in the following examples: (8) a. Bill bought for Susan1 a large flashy car [PRO1 to drive], b. John1 received from Susan a book [PRO1 to read]. Nishigauchi claims that the indirect object denotes the Goal in (8a), since the Theme (a large flashy car) moves to the indirect object Susan. On the other hand, the subject is assumed to be the Goal in (8b) because the movement of the Theme (a book) is directed toward the subject (John). This difference in control behavior is explained by the difference of predicate (buy vs. receive), which governs the control relation. However, there are cases in which no Goal is specified in a sentence. Consider the following examples from Nishigauchi: (9) John1 owns a car [PRO1 to carry his own belongings in]. (10) They deprived Mary1 of the money [PRO1 to pay her rent with]. Nishigauchi argues that the subject of verbs such as own, retain, and so on is associated with the thematic relation of Location, which is specified as what the Theme (e.g., a car in [9]) belongs to. On the other hand, the object of verbs like deprive, cure, and so on is assumed to bear the thematic relation of Source, from which the Theme (e.g., the money in [10]) is transferred. To summarize, if Goal is specified in a sentence, it is always chosen as a controller. When only Location or Source is specified, then that thematic role serves

Processing Empty Subjects

103

as controller. Nishigauchi thus claims that there is a hierarchical relation among the thematic roles chosen as controller, shown below: (11) The Primary Location Hierarchy l.Goal >> 2. Location, Source Nishigauchi's proposal regarding the general tendency for Goal antecedents does have some limitations, as he himself notices. In English, the promise-type obligatory control verb does not demonstrate a Goal preference. For instance, in "John promised Mary [PRO to go to Tokyo]," the object Mary should be the antecedent of PRO, since she is the Goal of John's promising, but this is not the correct interpretation. Furthermore, in a gerund construction like "John admitted to Mary [PRO having eaten the cake]," the indirect object Mary is not the controller despite being the Goal to which John admitted something. In Japanese, all the subject-control sentences are counterexamples to the general thematic preference for Goal. In these sentences, as in a promise-type sentence in English, the subject is the controller despite the fact that the object is the Goal. Thus, if there is a Goal preference, it seems to be overridden by specific lexical features of subject-control verbs. Nishigauchi's claim is based on an analysis of English, which is a head-initial language. However, Nishigauchi's proposal can account for the experimental findings in Japanese. Verbs carry a lot of information (e.g., subcategorization information, argument structure information, thematic structure information, etc.) that is useful for the parser. Based on verb information, the parser can expect what kind of theta-role is to be assigned to a given NP. (See, among others, Carlson and Tanenhaus, 1988, and Pritchett, 1992, for the argument that parsing is performed through theta-role assignment.) Because of the head-final characteristics of Japanese, however, the Japanese parser cannot utilize verb information in the early stage of parsing. Thus, Nishigauchi's claim is in need of some modification to deal with cases in Japanese sentence processing. Instead of verb information, the parser relies on information carried by nouns (i.e., grammatical case information). If the parser ignores information available from the Case-marking particle or delays utilization of this information, the parser must wait until the final verb appears. With this "wait-and-see" strategy, the parser will make no mistakes in determining the controller. This means that there should be no difference in RTs between subject- and object-control sentences. This contradicts our experimental findings. Therefore, it is reasonable to assume that the parser uses information from the Case marker before the final verb appears. The assumption made here is that the parser uses available information when it has a chance to do so. Recall that in the Japanese examples used as experimental sentences, objects were marked with the dative particle, -ni (meaning roughly 'to' or 'toward'). This dative marker may facilitate the interpretation that the object denotes a Goal. Thus

104

Tsutomo Sakamoto and Matthew Walenski

it could be assumed that the parser recognizes the object as the Goal and prepares to parse the given sentence as an object-control sentence. This expectation by the parser is satisfied by an object-control verb but not by a subject-control verb. Then, why does the particle -ni facilitate a Goal interpretation of the NP it attaches to? A very suggestive study has been done by Sadakane and Koizumi (1995). They argue that there are several homophonous particles -ni that can be distinguished by syntactic tests. By these tests, four types of -ni are found: the dative case marker, the postposition ni, the ni of ni insertion, and a form of the copula. Although these four types of ni can be differentiated by these syntactic tests, they mention that there is an ambiguous case between a Case marker and a postposition. Consider the following example: (12) Hokuto-wa Mika-ni hanataba-o okutta Hokuto-Top Mika-NI bouquet-Acc sent 'Hokuto sent a bouquet to Mika.' Sadakane and Koizumi argue that "Hokuto caused a bouquet to go to Mika with the intention of transferring possession of it to her, with Mika being an affected goal (the affected reading), or it may simply mean that Hokuto caused a bouquet to go to Mika, with Mika being a 'mere' goal (the nonaffected reading)." The idea of "affectedness" indicates that the referent of an NP with a particle is affected in the action denoted by the verb (predicate/sentence). In the above example, the point is whether Mika is affected by the action of sending a bouquet. It must be noticed that the difference between the two readings is not binary but a matter of degree. What is important for our discussion is that the referent of the NP is an affected goal, since under that interpretation -ni would have to be interpreted as a Case marker, not as a postposition. The sentence-final verb determines whether the nimarked NP is a goal or not. However, the parser encounters the NP before the verb, so if the parser is to make assumptions about the theta-role of the NP before it gets to the verb, it must be able to interpret the function of -ni (Case marker or postposition in this example) before it gets to the verb. But, given that there are four types of particle -ni, how can the parser make the correct decision concerning the Case and theta-role of a m'-marked NP? One possible way to answer this question would be to appeal to frequency. That is, the particle -ni is most frequently used as a dative Case marker with a Goal theta-role. Sadakane and Koizumi mention that data from Japanese children show that Case markers are acquired earlier than postpositions. Thus, there may be present in speakers from an early age this association between the particle -ni, the dative Case marker, and the Goal theta-role. Until the end of the sentence, the parser does not have the information that determines the grammatical properties of the particle -ni. Without the grammatical information the parser could rely on its association until more information becomes available.

Processing Empty Subjects

105

Another explanation would assume that there is a schema (or Construction), which forces an expectation for a specific type of sentence. For example, the sequence, 'Taroo-ga Hanako-ni ringo-d" induces the appearance of a specific type of verb (e.g., ageta [gave]). The basis of this induction is that the string "NP-ga NP-ra NP-o" facilitates an interpretation of the sequence as "Nominative Dative Accusative." It is further assumed that this set of Cases induces an interpretation of the argument roles of the NPs as "Agent Goal Theme." This argument follows the idea of the so-called "Canonical Sentoid Strategy," which is formulated by Fodor, Bever, and Garrett (1974, p. 345) as follows: "whenever one encounters a surface sequence NP V (NP), assume that these items are, respectively, subject, verb, and object of a deep sentoid." As is well known, there has been a lot of discussion concerning this kind of perceptual strategy (cf. Kess, 1992; Pritchett, 1992). Although most of the discussion does not favor heuristic strategies, we believe that such strategies, insofar as they are based on our linguistic knowledge, may be used by the human parser. In other words, it would be more precise to use the term expectations, that are formed by the human parser based on linguistic information available at the time of processing. Returning to our experimental setting, we assume that the parser formed a top-down expectation for an objectcontrol verb, and that it was either the ra'-marked NP itself or the constructionspecific sequence of Case markers that forced the parser to expect a verb that allows an affected reading of the ra-marked NP.

6. THETA-CHECKING STRATEGY Returning to our experimental setting in Japanese, we claim that the parser is ready to assign a Goal theta role to an NP that is Case marked with the dative -ni. At this point, the parser commits itself to a decision about the thematic structure of the sentence. The first commitment is confirmed by late information provided by a final object-control verb and is not confirmed by a subject-control verb. The failure of this confirmation in subject-control sentences causes longer processing times than in object-control sentences. This strategy could be summarized as follows: (13) Theta-checking strategy: Assign a tentative theta-role using Case information, and check it using verb information. The parser in Japanese knows that the tentative decision can be wrong, because the grammar of the language makes explicit that other information, which could be incompatible with previous information, will come at the end of the sentence. However, the parser does not "wait and see" until the end of the sentence. This kind of wait-and-see, or delayed parser is made unrealistic when some aspects of

106

Tsutomo Sakamoto and Matthew Walenski

the cognitive architecture (e.g., the limitations of short-term memory) are considered. There is no clear answer concerning how long the parser has to (and/or can) wait until it gets the decisive information. The very basic parsing strategy proposed here is that the parser makes use of available information whenever it has a chance to do so. Namely, for both a headinitial language (e.g., English) and a head-final language (e.g., Japanese), the parser utilizes theta-role information, the only difference being the way in which the parser gets this information. In the former case, the parser relies on the thetarole information available from a verb. In the latter case, the parser uses theta-role information available from a noun via a Case-marking particle. Of course, both sources of information are available in both types of languages (though the distribution of the sources of information does vary). The scenario is very simple. When the parser gets a verb that assigns certain theta-roles, the parser expects to encounter NPs with those theta-roles. When the parser gets an NP with a certain theta-role (as indicated by Case markings or other strategies), it expects to encounter a verb that can assign that theta-role. When mismatches occur, the parser needs to take time (which can be experimentally observed) to sort out and correct the conflict. Consider the following examples (cf. Inoue and Fodor, 1995; Pritchett, 1992): (14) a. Todd gave [NP the boy ] [NP the dog] bit a bandage ?? : Garden Path b. Todd gave [NP the boyt [the dog bit e1] a bandage. (15) a. [ Np Bob -ga] [NP Mary -ni] [NP ringo -o] tabeta inu-o ageta. -NOM -DAT -ACCate dog-ACC gave 'Bob ate an apple to Mary gave a dog??' : Garden Path b. [Np Bob -go] [Np Mary -ni] [Np [s e1 ringo-o tabeta] inu1-o ageta. -NOM -DAT apple-ACC ate dog-ACC gave 'Bob gave Mary the dog that ate the apple.' In (14a), when the parser gets the verb gave, it expects that two NPs will follow; one for the Goal theta-role and another for the Theme. This expectation is tentatively satisfied at the point where the dog is attached to the parse tree. However, when the parser gets the second verb, bit, which requires an Agent and a Patient NP, the initial attachments cannot be maintained, as there aren't enough NPs to go around. Therefore, the parser must reanalyze the string as in (14b), where one of the NPs, the boy, in effect does double duty, as both the Goal of give and the Patient of bit. In (15a), when the parser gets three NPs in succession, it expects to encounter a three-place predicate such as give, which can take an Agent, Goal, and Theme argument, and would be consistent with the pattern of Case markings on the NPs. This expectation is not satisfied because tabeta 'ate' is a two-place predicate that cannot subcategorize for three NPs, but can only take an Agent and Theme argument. On encountering the following noun, inu 'dog' the parser must reanalyze

Processing Empty Subjects

107

(15a) to (15b) resulting in a (mild) garden-path effect (but see Mazuka and Itoh, 1995, for a different view). It is often mentioned that the garden-path effect in Japanese is not as strong as the one in English (e.g., Mazuka et al., 1989). This is explained in the proposed framework as follows. In English, the verb (or predicate) is practically the only source of information that determines the theta-role of an NP. In Japanese, there are two sources of information that are useful for indicating the theta-role of an NP—the Case-marker and the verb (predicate).4 The English parser has to determine the theta-role of an NP when it gets a verb (predicate), whereas the Japanese parser has a chance to do it twice. In other words, the English parser is more confident than the Japanese parser in the sense that the former has only one information source and the latter has two information sources, which might be incompatible with each other. Since the Japanese parser is less confident with its thetarole assignments, it is more ready to remedy its first incorrect decision upon encountering the second source of information. See Inoue and Fodor (1995) for this "confidence-remedy" argument in Japanese parsing. However, the point is that the parser in both types of language uses theta-role information as soon as it can. Yamashita (1995) reports an experimental result that "indicates that by receiving the information up to the verb, the subjects expected how the syntactic structure unfolds" (p. 344).

7. MODELS OF THE HUMAN PARSER Let us briefly consider what kind of parser can account for (or is compatible with) the findings of the experiments reported here. Here we restrict the discussion to issues of "parallel/serial/delayed," and "top-down/bottom-up" processing. These issues have been well discussed in the literature and are directly relevant to the basic design of the parser. If the parser is strictly parallel, it will process both the subject-control and the object-control options at the same time. We should expect to see no differences in RTs between these two types of sentence if this is the case. In a similar vein, if the parser delays its decision until it can get decisive information, the parser will not make mistakes. Neither the parallel nor the delayed (wait and see) processing model can account for why there is a difference in RTs for the different types of sentence. Therefore, the properties of the parser must either be explained by some weakened version of these models (ranked parallel, cf. Gorrell, 1987; or partially delayed, cf. Mazuka and Itoh's "Tentative Attachment Strategy"), or by some form of serial model. Now let us consider the second property of the parser (i.e., top-down or bottomup). It has been argued that the traditional type of top-down parsing system (Kimball, 1973; Woods, 1970) wrongly predicts enormous garden-path phenomena for a head-final language such as Japanese (Inoue and Fodor, 1995; Mazuka and

108

Tsutomo Sakamoto and Matthew Walenski

Itoh, 1995). Since the decisive information comes last in a head-final language, it is always dangerous to make a top-down prediction. For example, when the parser receives the sequence "Taroo-ga Hanako-o mita" ("Taroo saw Hanako"), it is not clear whether this sequence is a simple sentence or a relative clause such as "Taroo-ga [e1 Hanako-o mita] otoko1-o nagutta" (Taroo hit a man who saw Hanako), or a time adverbial clause such as "Taroo-ga Hanako-o mita toki" (When Taroo saw Hanako), or a nominalized clause such as "Taroo-ga Hanako-o mita koto" (the fact that Taroo saw Hanako). There are so many possible ways of continuing this sequence that any top-down predictions that the parser makes are likely to be wrong. On the other hand, assuming that the parser is strictly bottom-up also cannot account for the observed results. Since the bottom-up parser postulates a clause only after the head of the clause has been received, this type of parser has the same properties as a parser in a delay model. The bottom-up parser waits until it gets the head of a structure, just as the wait-and-see parser waits until it gets decisive information (often contained within the head of a particular structure). Therefore, it is reasonable to suppose that the type of parser that is most compatible with the reported results is partially bottom-up (i.e., data driven), and partially topdown (expectation making). The parser makes a top-down expectation through a bottom-up procedure (cf. Marcus, 1980). The parsing model most consistent with both of these considerations (parallel/serial/delayed and top-down/bottom-up) is the "committal" model of Inoue and Fodor (1995), which forms an expectation based on current information, and then checks this expectation as new, potentially decisive information, becomes available.

8. FINAL REMARKS Two apparently conflicting results regarding the processing of control structures in English and Japanese were examined. The English results (Frazier et al., 1983) suggest that subject-control structures are easier to process, whereas the Japanese data (Sakamoto, 1996) suggest that the parser prefers to assume that an object is a controller in the absence of other information. To resolve this conflict, it was noted that the English data were confounded, in that the object-control sentences were always accompanied by a double-gap construction, whereas the subject-control sentences had only a single gap. Therefore, it is reasonable to conclude that the English experiments were responsive to this factor, and not to factors involved in the identification of the antecedent of PRO. Since the Japanese materials were not confounded in this way, it is reasonable to conclude that the results from these experiments are universally applicable. To account for this result using parsing theory, two proposals were made. First, it is claimed that the

Processing Empty Subjects

109

parser's preference for object control stems from a grammatical preference for Goal control (cf. Nishigauchi, 1984). Because Goals are overtly marked in Japanese by Case information, we are led to the conclusion that Case information is used in parsing, and forms the basis for the parser's initial decisions about the incoming structure. The parser will make use of later information to modify these initial assumptions as necessary. The parsing model that is most consistent with these findings is the serial "committal" parser that makes top-down predictions based on available information (e.g., Case information on a noun) and makes necessary corrections in accord with incoming information (e.g., Theta-role information in a verb).

ACKNOWLEDGMENTS This work was supported in part by Grant-in-Aid for Scientific Research (c) No. 07801073 from the Japanese Ministry of Education to the first author, and in part by a Monbusho summer fellowship awarded to the second author.

REFERENCES Boland, J. E., Tanenhaus, M. K., and Garnsey, S. M. (1990). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language 29, 413 -432. Carlson, G. N., and Tanenhaus, M. K. (1988). Thematic roles and language comprehension. In W. Wilkins (ed.), Syntax and Semantics (vol. 21, pp. 263-288). San Diego, CA: Academic Press. Chomsky, C. (1969). The acquisition of syntax in children from 5 to 10. Cambridge, MA: MIT Press. Chomsky, N. (1982). Some concepts and consequences of the theory of government and binding. Cambridge, MA: MIT Press. Clifton, C., and Frazier, L. (1986). The use of syntactic information in filling gaps. Journal of Psycholinguistic Research, 15, 209-224. Cohen Sherman, J. (1983). The acquisition of control in complements.—The Role of Structural and Lexical Factors. Ph.D. Dissertation, Cornell University. Cohen Sherman, J. (1987). Evidence against a Minimum Distance principle in first language acquisition of anaphora. In B. Lust (Ed.), Studies in the acquisition of anaphora, Volume II.—Applying the constraints. Dordrecht: Reidel. Grain, S., and J. D. Fodor (1985). How can grammars help parsers? In D. R. Dowty, L. Karttunen and A. Zwicky (eds.) Natural language parsing. Psychological, computational and theoretical perspectives. Cambridge: Cambridge University Press. 95-128.

110

Tsutomo Sakamoto and Matthew Walenski

Fodor, J. D. (1988). On modularity in syntactic processing. Journal of Psycholinguistic Research, 17, 125-168. Fodor, J. D., Bever, T. G., and Garrett, M. (1974). The psychology of language. New York: McGraw-Hill. Ford, M., Dalrymple, M. (1988). A note on some psychological evidence and alternative grammars. Cognition, 29 (nl), 63-71. Frazier, L., Clifton, C, and Randall, J. (1983). Filling gaps: decision principles and structure in sentence comprehension. Cognition, 13, 187-222. Gorrell, P. (1987). Theoretical and methodological issues in natural language processing. Ph.D. dissertation. The University of Connecticut. Gorrell, P. (1995). Japanese trees and the garden path. In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing (p. 331-350). Hillsdale, NJ: Erlbaum. Inoue, A., and J. D. Fodor (1995). Information-paced parsing of Japanese. In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing (9-63). Hillsdale, NJ: Erlbaum. Jackendoff, R. (1972). Semantic interpretation in generative grammar. Cambridge, MA: MIT Press. Kess, J. F. (1992). Psycholinguistics: Psychology, linguistics, and the study of natural language. John Benjamins Publishing Company: Amsterdam, Netherlands. Kimball, J. (1973) Seven principles of surface structure parsing in natural language. Cognition, 2, 15-47. Marcus, M. (1980). A theory of syntactic recognition for natural language. Cambridge, MA: MIT Press. Mazuka, R., and Itoh, K. (1995). Can Japanese speakers be led down to the garden path? In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing (295-329). Hillsdale, NJ: Erlbaum. Mazuka, R., K. Itoh, S. Kiritani, S. Niwa, K. Ikejiri, and K. Naito (1989). Processing of Japanese garden path, center-embedded, and multiply-left-embedded sentences: reading time data from an eye movement study. Annual Bulletin of the Research Institute of Logopedics and Phonetics, 23, 187-212. Mazuka, R., and Nagai, N. (Eds.) (1995). Japanese sentence processing. Hillsdale, NJ: Erlbaum. Nakayama, M. (1996). Acquisition of Japanese empty categories. Tokyo, Japan: Kurosio Publishers. Nicol, J. (1988). Coreference processing during sentence comprehension. Ph.D. dissertation, MIT. Nishigauchi, T. (1984) Control and the thematic domain. Language, 60, 215-250. Pritchett, B. (1992). Grammatical competence and parsing performance. Chicago, IL: University of Chicago Press. Sadakane, K., and Koizumi, M. (1995). On the nature of the "dative" particle ni in Japanese. Linguistics 33, 5-33. Sakamoto, T. (1991). Processing empty subjects in Japanese.—Implications for the transparency hypothesis. Ph.D. dissertation. City University of New York. (Published as Sakamoto, 1996). Sakamoto, T. (1995a). Transparency between parser and grammar: On the processing of empty subjects in Japanese. In R. Mazuka and N. Nagai (Eds.), Japanese sentence processing (275-294). Hillsdale, NJ: Erlbaum.

Processing Empty Subjects

111

Sakamoto, T. (1995b). Koubun Kaiseki niokeru Toumeisei no Kasetu: Kuusyugo-o Fukumu Bun-no Syori ni kanshite. [Transparency hypothesis in parsing: concerning the processing of sentences with empty subjects.] Ninti Kagaku (Cognitive Science) 2 (No. 2), 77-90. Sakamoto, T. (1996). Processing empty subjects in Japanese.—Implications for the transparency hypothesis. Fukuoka: Kyushu University Press. Williams, E. (1980). Predication. Linguistic Inquiry, 11 (No.l), 203-238. Woods, W. (1970). Transition network grammars for natural language analysis. Communications of the Association for Computing Machinery, 13, 591-606. Yamashita, H. (1995). Verb argument information used in a prodrop language: An experimental study in Japanese. Journal of Psycholinguistic Research, 24 (5), 333-347.

NOTES 1 There have been many attempts to address this question within the generative tradition, a full discussion of which is beyond the scope of this work. See, among others, Williams (1980), and Nishigauchi (1984) (discussed in detail below). 2 Boland et al. (1990) argue that control information, which would in turn determine the identity of PRO, is available immediately. 3 See below for an argument that strategies based on case information may be used in this interval. 4 The lack of clear Case information in English may make it difficult for the parser to commit before verb information becomes available.

This page intentionally left blank

THE PRODUCTION OF AGREEMENT IN ENGLISH AND JAPANESE: ANIMACY EFFECTS (OR LACK THEREOF) JANET L. NICOL Departments of Psychology and Linguistics University of Arizona Tucson, Arizona

1. INTRODUCTION This chapter is about the implementation of agreement. Agreement is of interest to linguists and psycholinguists because the phenomenon, is widespread, and exhibits variation, both with respect to the type of agreement and the lexical categories that enter into agreement. In this chapter, I discuss three cases of agreement: subject—verb agreement of number in English, subject-verb agreement of animacy in Japanese, and subject-pronoun agreement of animacy in English. Agreement typically involves particular semantic or conceptual features, such as number, animacy, and gender (among others), yet the implementation of agreement could be based on syntactic aspects of a sentence, since agreement is often between elements in particular syntactic positions within a structure. This duality has led to different characterizations of agreement, as either syntactic or semantic in nature. The syntactic view holds that agreement features—though they may be semantically based—have been grammaticized: a feature like [+plural] is syntactic. In addition, the constituent that controls agreement is identifiable from its syntactic position as the head noun phrase (NP) within the preverbal NP, as shown by the examples in (1), in which key is the head noun, and the verb agrees in number with this head noun. Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

113

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

114

Janet L. Nicol

(1) a. The secretaries' key to the cabinets by the desk was stolen, b. The secretary's keys to the cabinet by the desk were stolen. But notice that the number of the verb could be specified according to the number of the true subject: sentence (la) is, after all, about a key, key is the subject— in the nontechnical sense—of the clause. Subject-verb agreement may simply mean that the number of the verb matches the number of the subject. Thus, the tight correspondence between a semantic, or logical, representation and its corresponding syntactic representation makes it difficult to know, in most instances, the level at which agreement is computed. Therefore, we need to consider cases in which there is a mismatch between the semantic and syntactic representations. Consider the sentences in (2). (2) a. The committee is getting together their fall schedules, b. The scissors are on the table. English has groups of nouns in which grammatical number diverges from conceptual or notional number (see Bock, 1995; Bock and Miller, 1991). Collective nouns (e.g., committee, team), as in (2a) are grammatically singular, but may have a plural reading (a reading that may be picked up by a later-occurring pronoun, as in 2a). Dual-object nouns (e.g., binoculars, scissors), as in (2b), are notionally singular but grammatically plural. In both cases, there are crosslinguistic (and even dialectal) differences: hair is singular in English, but plural in French; binoculars is plural in English but singular in Dutch; family is likely to elicit singular verb agreement in American English, but plural verb agreement in British English. In these cases, it is the syntactic status that is subject to crosslinguistic variation; presumably, the conceptual representations for these items are similar for speakers of different languages. In all, these examples suggest that when there is a mismatch, the verb agrees with syntactic, rather than semantic, aspects of the subject's head noun. The examples in (1) and (2) also suggest that some lexical items may be flagged: the lexical entries for words like committee may be specified as syntactically singular (in American English), words like pants may be specified as syntactically plural. By contrast, the head NP in the example in (Ib) is a derived plural based on the particular context: the situational context involves more than one key. The lexical entry for key is not flagged with a special feature; it receives a plural feature during the encoding of the utterance. However, in both (1) and (2), verb number could be specified through operations that consider only the syntactic specifications of the head noun: once the features are in place (whatever their source), verb number may be specified. But other examples suggest that verb agreement is sometimes semantic. Consider the following examples from Pollard and Sag (1994).

Animacy Effects

115

(3) a. The hash browns at table nine is getting angry, (p. 69) b. Eggs is my favorite breakfast, (p. 70) c. Doing phonology problems and drinking vodka makes me sick. (p. 70) (3a) is an example of reference transfer in which, arguably, the hash browns refers to the customer who ordered the hash browns. Here, the number of the verb agrees with the number of the referent, not with the syntactic subject. Pollard and Sag argue that agreement features "are not attributes of syntactic categories,. . . but rather are attributes of the indices" (p. 75), which are encoded on agreeing elements in the context of a given utterance. According to this view, the indices are context-dependent, so that, for example, a name like John may bear a feminine, third-person index if it so happens that John refers to a woman. This approach allows a lexical item to become further elaborated, in terms of agreement features, by taking into account attributes of the real-world referent referred to by that lexical item. Pollard and Sag argue that this analysis accounts for (3), in which a singular verb appears with a plural subject. In each case, the lexical representation in subject position is assigned an index that connects it to a real-world referent: The lexical item hash browns is augmented with an index that specifies that it is— in this context—human and singular.1 A similar analysis may be applied to pronominal forms that co-refer with the head. (4) a. The hash browns at table nine burned himself on the hot plate, b. Eggs is your favorite breakfast, isn't it? The reflexive in (4a) and the pronoun in (4b) both agree with the entity referred to by the expressions the hash browns and eggs, rather than to the expressions per se. The following examples also show that conceptually based agreement is possible: (5) a. Half the students are still working on the exam. b. A lot of kids watch too much television. c. The rest of the groceries are in the car. Although it may be argued that quantifying expressions such as a lot of ("a lotta") and a number o/are idiomatic, and have come to be synonymous with many, it is more difficult to make that argument about less frequently used quantifiers such as half and rest. Altogether, these examples suggest that when grammatical number and semantic number mismatch, verb agreement is computed in some instances on the basis of grammatical number, and in other instances on the basis of semantic number. However, these examples simply show that both options are possible; they do not

116

Janet L.Nicol

address the question of how agreement is typically implemented in the course of production. I turn now to the experimental evidence.

2. STUDIES OF AGREEMENT ERRORS A set of studies by Bock and Miller (1991) inspired a number of follow-up experiments designed to elicit agreement errors (e.g., Bock and Cutting, 1992; Bock and Eberhard, 1993; Nicol, 1995; Vigliocco, Butterworth, and Garrett, 1996; Vigliocco, Butterworth, and Semenza, 1995; see Bock, 1995, for a review of some of this research). Experiments typically involve asking participants to repeat a sentence fragment or "preamble" (e.g., The key to the cabinets . . .} presented either auditorily or visually, and to provide a sentence completion (e.g., a participant might say, "The key to the cabinets is in the box."}. In some of the preambles, head nouns and nonhead nouns match in number (e.g., The key to the cabinet, The keys to the cabinets}; in others, they mismatch (e.g., The key to the cabinets, The keys to the cabinet}. Participants do make errors (e.g., "The key to the cabinets are in the box"}, typically with sufficient frequency for differences between preamble types to emerge. Based on the distribution of errors across different preamble types, hypotheses have been generated about how, and over what types of representation, agreement is computed. In the following sections, I will review some of the studies that focus primarily on semantic effects on agreement. I will begin with results of subjectverb agreement. I will then discuss the results of two recent studies that extend the prior work in two directions: (a) animacy errors in verbs in Japanese; and (b) animacy and number errors in tag question pronouns in English. 2.1. Subject-Verb Agreement Errors in English Bock and Miller's initial study showed that more agreement errors occur when head nouns and nonhead nouns mismatch than when head nouns and nonhead nouns match. However, the pattern of errors is asymmetric: more agreement errors occurred after singular heads than plural ones (e.g., The key to the cabinets than The keys to the cabinet} (Bock and Miller, 1991). These findings suggest that agreement errors are not simply due to performance failures (e.g., memory limitations, lapses of vigilance), since these should affect both mismatch conditions. Bock and Eberhard (1993) have suggested that the source of the asymmetric pattern of errors is an asymmetry between the status of a singular and the status of a plural: the plural is marked, whereas the singular is the unmarked default. Because of the markedness of the plural, plural nonhead nouns may interfere with the agreement process, whereas singular nonhead nouns are unlikely to do so.

Animacy Effects

117

This asymmetry is present in natural discourse, as shown by analyses of English corpus data (Bock and Miller, 1991) and has been replicated in English (Bock and Cutting, 1992; Bock and Eberhard, 1993), Spanish (Anton-Mendez, 1996; Vigliocco et al., 1996), and Italian (Vigliocco et al., 1995).2 In addition, Bock and her colleagues have shown that error incidence is unaffected by semantic aspects of words and phrases. In one set of studies, conceptual or notional number was manipulated. In another set of studies, the animacy of the head and nonhead nouns was varied. I will consider each set in turn. Bock and Miller (1991) investigated whether or not notional number affected the incidence of agreement errors. They constructed two types of items: "multiple-token" items like The label on the bottles and "single-token" items like The key to the cabinets: In the former, the most plausible interpretation of the phrase involves multiple tokens of the label (rather than one label spanning multiple bottles); the latter is plausible on the single-token reading (one key to multiple cabinets). If, as Pollard and Sag (1994) assume, the features of subjects are encoded referentially, multiple-token items should cause more agreement errors than single-token items. However, results showed no effect of notional number. In a follow-up study, Bock and Eberhard (1993) investigated whether or not the notional number of nonhead nouns would cause agreement errors. Results showed that collective nonhead nouns (e.g., The condition of the fleet) did not trigger more agreement errors than singular nonhead nouns (e.g., The condition of the ship). These results clearly indicate that in English, the agreement process is not typically sensitive to meaning.3 The effect of animacy of nouns within a preamble was also investigated. Statistically, sentential subjects are more often animate than inanimate (Bates and MacWhinney, 1989; Bock, Loebell, and Morey, 1992; Bock and Miller, 1991). If subject-verb agreement were computed at a semantic, or conceptual representation, then aspects of word meaning such as animacy would be available to interact with the agreement process. Bock and Miller (1991) tested the possibility that there would be greater interference from a plural nonhead if that NP were also animate, since an animate NP (but not an inanimate one) might compete for subjecthood with the head NP. They compared preambles such as The speech of the authors (and its singular-singular control) with The author of the speeches (and its control). They found no increase in errors following preambles containing animate nonheads. This finding provides further evidence that the agreement process is not sensitive to meaning. In contrast, animacy does influence error rate when the preamble contains two subjects. Bock and Miller presented participants with main clause subjects followed by relative clause subjects (e.g., The songs that the composer . . .), and participants were required to produce an embedded verb (a verb that would be compatible with the composer as its subject), followed by a matrix verb (for which the songs would be the subject). In each experimental trial, either the matrix subject or the embedded subject was animate; the other NP was inanimate. As usual,

118

Janet L.Nicol

the number of each noun was systematically varied. Under these circumstances, embedded verbs (i.e., the verbs produced immediately after the fragment) sometimes erroneously agreed with animate main clause subjects. But the reverse error—the embedded verb agreeing with the inanimate main clause subject—did not arise. This suggests that when there were two subjects—the matrix subject and the embedded subject—and momentary confusion on the part of the speaker as to which NP was the subject of the immediately upcoming verb, the animate NP was selected over the inanimate one. (It is important to point out, however, that error rates were generally low: in the main, participants did not ignore structural relations). The results of the two animacy studies together show that animacy may be taken into account during the process of designating the subject for a verb but does not otherwise appear to interact with the agreement process. This view is supported by naturally occurring errors such as the following (Stemberger, 1982): That's supposed to hang on to you (instead of You're supposed to hang on to that.) What has gone wrong in this utterance is that the two NPs have ended up with the wrong roles: the second person pronoun was meant to be the subject; the demonstrative pronoun was meant to be the oblique object. But once the error has occurred, subject-verb agreement appears to proceed in normal fashion, blind to the intentions of the speaker. What Bock and Miller (1991) showed was that animacy does not interact with number agreement. But perhaps animacy does interact with animacy agreement. Verbs in English do not agree with heads in terms of animacy, but for a limited set of cases, verbs in Japanese do. We turn to these cases now. 2.2. Subject-Verb Animacy Agreement in Japanese Japanese has no number agreement between subjects and verbs; however, with a small set of predicates, there is agreement of animacy. The predicate iru ('exist') agrees with animate subjects, and the predicate aru ('exist') agrees with inanimate subjects. Similarly, inai ('not exist') agrees with animate subjects, and nai ('not exist') agrees with inanimate subjects.4 This contrast is illustrated in the examples below. Note that Japanese typically marks case, so that each noun in the examples below appears with a case marker. In addition, Japanese is a head-final language, so that in examples (8) and (9), the first two NPs modify the head, which is marked with the nominative case marker -ga below (from Suzuki, 1997). (6)

Tora-ga iru/*aru. tiger-NOM exist 'There is a tiger.'

(7)

Saku-ga *iru/aru. fence-NOM exist 'There is a fence.'

Animacy Effects

(8)

Saku-no mae-no tora-ga inaku not ta/*naku nat ta. fence-GEN front-GEN tiger-NOM not-exist-become-past 'The tiger in front of the fence is gone.'

(9)

Tom-no mae-no saku-ga nakunat ta/*inaku nat ta. tiger-GEN front-GEN fence-NOM not-exist-become-past 'The fence in front of the tiger is gone.'

(10)

Tora-tosaku-ga *inaku nat ta/*naku nat ta. tiger-and fence-NOM not-exist-become-past 'The tiger and the fence are gone.'

119

As these examples illustrate, the form of the predicate depends on the animacy of the head. Note that in (10), the subject NP contains a coordination of an animate noun and an inanimate noun; this coordinate NP, with conflicting animacy features, is ungrammatical with either an animate predicate or an inanimate predicate. In a study reported by Suzuki (1997) and Suzuki and Nicol (1996), the animacy of NPs within a complex subject was manipulated, and productions were elicited by having participants read a computer-displayed preamble, say it aloud, and complete the sentence by providing one of the predicates described above.5 Preambles consisted of a complex subject NP in which the head was preceded by a modifying phrase containing an NP that either matched or mismatched in animacy with the head. For variety, there were three types of preamble, containing either (a) a single phrasal modifier; (b) a double modifier such as the examples in (8) and (9); or (c) a clausal modifier. In addition, the animacy of the head and nonhead was varied systematically (for the double modifier cases, the intermediate noun—which designated a location—was always inanimate). Examples of the experimental stimuli are shown in Table 1. A goal of the study was to explore whether (a) an animacy mismatch elicits errors and (b) if so, whether there would be an asymmetry in the mismatch effect. One possible source of asymmetry could be the correlation between animacy and subjecthood, which has been demonstrated for Japanese as well as for English. For example, it has been shown that when Japanese speakers are given asyntactic word strings in which case markers are absent and asked to identify a subject, they choose animate nouns as subjects much more frequently than inanimate nouns (Harrington, 1987; Sasaki, 1991). In English the animacy of nonhead nouns did not affect number agreement (Bock and Miller, 1991), presumably because the number agreement process is blind to animacy information. It is possible, however, that an agreement process that attends to animacy will be unable to ignore the fact that an animate noun is a likely subject head. If the correlation between animacy and head-of-subject is to interact at all with a subject-verb agreement process, it should be manifest here: animate (mismatching nonhead) NPs should interfere more than inanimate ones. The alternative is that one set of operations attends to semantic role designation for which the animacy of a NP is clearly

120

Janet L. Nicol TABLE 1 EXPERIMENTAL STIMULI USED IN JAPANESE ANIMACY AGREEMENT STUDY Sentence codesa

Type 1—Single modifier A A (animate nonhead, animate head)

AI (animate nonhead, inanimate head)

11 (inanimate nonhead, inanimate head)

AI (animate nonhead, inanimate head)

Type 2—Double modifier A A (animate nonhead, animate head)

Type 3—Clausal modifier A A (animate nonhead, animate head)

IA (inanimate nonhead, animate head)

II (inanimate nonhead, inanimate head)

AI (animate nonhead, inanimate head)

Example

Zimuin-no neko office clerk-GEN cat 'The cat of the office clerk's' Heya-no nezumi room-GEN mouse 'The mouse in the room' Heya-no terebi room-GEN TV 'The TV in the room' Zimuin-no huku office clerk-GEN clothes 'The clothes of the office clerk's' Tora-no mae-no tiger-GEN front-GEN 'The deer in front of the tiger' Iwa-no mae-no rock-GEN front-GEN 'The deer in front of the rock' Iwa-no mae-no rock-GEN front-GEN 'The fence in front of the rock' Tora-no mae-no tiger-GEN front-GEN 'The fence in front of the tiger'

sika deer sika deer saku fence saku fence

Kirin-ga kanda koara giraffe-NOM bit koala The koala that the giraffe bit' Booru-ga attata inosisi ball-NOM hit against wild boar 'The wild boar that the ball hit against' Booru-ga attata teeburu ball-NOM hit against table 'The table that the ball hit against' Kirin-ga kanda ringo giraffe-NOM bit apple 'The apple that the giraffe bit'

a

A = animate; I = inanimate; heads are underlined. From Suzuki, 1997.

relevant, and a different set of operations is dedicated to agreement (whatever the type of agreement). This latter view predicts that, all else being equal, there will be no asymmetry in the distribution of errors.

Animacy Effects

121

2.2.1 RESULTS Table 2 displays the numbers of agreement errors (and proportions of errors per total number of responses), miscellaneous errors, and correct responses. There were more agreement errors when there was a head-nonhead animacy mismatch. In addition, there is no asymmetry: animate nonheads created no more interference than inanimate nonheads. (Statistical analysis reveals a mismatch effect, but no interaction.) Two other aspects of the data are worth noting. First, there was a greater number of errors for preambles containing animate heads than for those containing inanimate heads. There is no obvious explanation for this difference, though it is possible that the animate forms (iru, inakunattd) are morphologically derived from the inanimate forms. Hence, the inanimate forms could simply be more readily accessed and would have the effect of raising the number of errors for both match and mismatch cases. Whatever the explanation for effect of head animacy, this effect does not undermine the fact that both head types show a mismatch effect, and the mismatch effect in both cases is equivalent: there is absolutely no asymmetry. A second notable finding is that there is a mismatch effect despite the fact that the head noun and verb are contiguous. This replicates a finding reported by Vigliocco and Nicol (1997), that a plural nonhead triggers agreement errors even when participants are asked to form questions (e.g., "Were the key to the cabinets shiny?"}. The mismatch effect observed in this experiment suggests that when a head and nonhead differ with respect to an agreement feature, interference may result, even though the mismatching NP does not intervene between the head and the verb. It is important to note, however, that the error rate is very low (overall, there were only 2% agreement errors, in contrast to the 5-15% observed in other elicited production experiments involving agreement). There are three possible explanations for this. One is that this represents a base rate for error occurrence when the head and verb are adjacent. Second, it may be that the completion task was so constrained that participants were able to concentrate on the preamble more fully than an open-ended completion task; hence, they made fewer errors. These two TABLE 2 RESULTS OF STUDY OF SUBJECT-VERB ANIMACY AGREEMENT IN JAPANESEa

a

Preamble type (N = 1792)

Agreement errors

Error rate (%)

Other errors

Animate head, animate nonhead Animate head, inanimate nonhead Inanimate head, inanimate nonhead Inanimate head, animate nonhead

37 54 16 39

2.1 3.0 0.9 2.3

54 67 54 59

From Suzuki, 1997.

Correct responses 1701 1671 1722 1694

122

Janet L.Nicol

explanations are at odds, however, with the error rate found in Vigliocco and Nicol's (1997) question experiment (about 6.5%), in which stimuli were also presented visually, the head NP and verb were also adjacent, and the completions similarly constrained. A third possibility—and from my perspective the most plausible—centers around the difference between number agreement and other types of agreement. As discussed in the introduction, number marking is "contextual" in the sense that an item is marked as plural if it denotes a plural entity in the discourse: for most nouns, the stored lexical item (or lemma) is presumably singular, and is affixed with a plural marker only if the occasion calls for it. Suppose that the likelihood of an error is directly related to the dissociability of the agreement feature from its lexical host. In most cases, a plural feature is entirely dissociable, because if it is removed, it leaves behind a singular. But in general, animacy features are not dissociable from their hosts; animacy is an inherent property. Hence few errors arise. I will return to this point in the discussion of the next experiment. In sum, it was predicted that if the probability that a verb agrees with a noun correlates with the animacy of the nouns (Bates and Mac Whinney, 1989), animate nonhead nouns should cause more agreement errors than inanimate nonhead nouns. No such animacy effect was found. Hence, even when the agreement process is one which is sensitive to animacy, animate nonheads have no greater interference effect than inanimate ones. It is possible that for both Japanese and English, subject-verb agreement processes are blind to features that are not directly involved in the agreement process, and blind to factors that may matter to antecedent operations, such as how subjects are so designated in the first place. If subject-verb agreement across languages is predominantly syntactic, then this process would not necessarily take other kinds of information into account. But other types of agreement, such as pronounantecedent agreement, may consider conceptual aspects of a representation, and if so, then these may show an animacy effect. 2.3. Pronoun Agreement Errors A number of studies in English have shown that pronominal agreement may be more likely than subject-verb agreement to involve a conceptual representation of a subject. In a sense, this is to be expected, because although some pronominal forms must agree with the head of the subject, they also corefer with it, and coreference is a semantic notion. A study by Cutting, Nicol, and Bock (1995) tested preambles such as the following. (11) a. The gang with the dangerous rival... b. The gang with the dangerous rival armed... c. The gang with the dangerous rival vanished ...

Animacy Effects

123

There were three groups of participants. One group was presented with preambles such as (1 la) and asked to repeat the preamble and produce an ending. The focus here was on the form of the verb that was produced. Another group was given preambles like (IIb) and was asked to repeat the preamble and complete the sentence with a reflexive. A third group was presented with preambles such as (11c) and asked to repeat the sentence and add a tag question containing a pronoun. Obviously, these were not meant to elicit errors: A collective term like gang may be interpreted as either singular or plural, so plural forms (such as are, themselves, they) may be grammatical. Our focus was on the number of plural forms associated with each type of agreeing element—verbs, reflexives, and pronouns. The results showed a clear contrast between the number of plural verbs (36% of all inflected verbs produced) and the number of plural preforms (71% of the reflexives and 74% of the pronouns). This suggests that verb agreement processes are more sensitive to grammatical number than conceptual number, but just the opposite for pronoun agreement. This difference also emerged for preambles that allow a multiple token reading. As discussed earlier, the most plausible interpretation of the phrase The label on the bottles is one in which there is one label per bottle (different tokens of the same type); hence, multiple labels. According to Bock, Eberhard, and Cutting (1992), although verb agreement errors occur with equal frequency for both multiple-token preambles and single-token preambles (e.g., The key to the cabinets), plural pronouns (in tag questions) occur with greater frequency in the multiple-token cases. In sum, these studies showed that pronominal agreement is more likely to implicate semantics than is verb agreement. In a recent paper (Nicol, 1998), I described the results of an experiment which involved the elicitation of tag questions. This study used preambles such as those in Table 3, along with another set of items that were part of a different test. All experimental sentences contained a singular head NP, which was either animate or inanimate. The nonhead was either (a) singular or plural, and (b) animate or TABLE 3 STIMULI USED IN ENGLISH TAG EXPERIMENT Typea

Examples

AA A A-plural AI A I-plural II I I-plural IA I A-plural

The bachelor with the unpleasant neighbor disappeared. The bachelor with the unpleasant neighbors disappeared. The bachelor with the overdue bill disappeared. The bachelor with the overdue bills disappeared. The delivery with the overdue bill disappeared. The delivery neighbors with the overdue bills disappeared The delivery for the unpleasant neighbor disappeared. The delivery for the unpleasant disappeared.

a

A = animate; I = inanimate; heads are underlined.

124

Janet L.Nicol TABLE 4 ENGLISH TAG EXPERIMENTa

Animacy Animacy Number Preamble type errors errors (%) errors N = 224 (per condition) AA A A-plural AI A I-plural II I I-plural IA I A-plural

03 05 03 03 04 03 03 03

1.3 2.2 1.3 1.3 1.8 1.3 1.3 1.3

03 15 03 11 02 27 01 13

Number errors (%)

Other errors

Correct

1.3 6.7 1.3 4.9 0.9 12.1 0.4 5.8

32 26 28 36 10 33 25 30

186 178 190 174 208 161 195 178

a All responses (56 participants, 4 items per condition). Percentages of animacy and number errors (as a function of total number of errors per condition) are also shown.

inanimate. (Among the other preambles, which are not the focus of this chapter, there were also variations in gender.) Note that because both number and animacy of the nonhead were manipulated, it is possible to look for animacy effects in both the animacy errors and in the number errors. This study was conducted in order to address three questions: (a) Does an animacy-mismatching nonhead trigger agreement errors, as in Japanese verb agreement? (b) If so, do animate nonheads cause more interference than inanimate nonheads? (c) In preambles in which there is an animacy mismatch and a number mismatch, do more number errors arise when the plural nonhead is animate than when it is inanimate? The results of this study are displayed in Table 4. 2.3.1. ANIMACY ERRORS Let us first consider animacy errors in the number match conditions only (since these provide the clearest examples): this is an error such as, "The bachelor with the overdue bill disappeared, didn't it?" and "The delivery for the unpleasant neighbor disappeared, didn't she?" As can be seen in Table 4, the incidence of animacy errors across conditions was uniformly very low, and mismatches in animacy elicited no increase whatsoever in the number of such errors. Since there is no mismatch effect, there is obviously no mismatch asymmetry. These results contrast with the results from the Japanese study, which showed an animacy mismatch effect. The reason for this is not yet clear; however, there are a number of differences in the materials (and languages) that may be at the root of the discrepancy. First, in both languages, inherent animacy may be "overridden" by reference transfer, as in examples like (12a) (from Pollard and Sag, 1994). (12) a. The volcano who just left the room was Bill's kid.

Animacy Effects

125

Suzuki (1997) reports that in Japanese, an animate object represented by a statue, for example, would take the inanimate form of the verb, as in (13). (13)

Tora-ga am. tiger-NOM exist 'There is a (statue) tiger.'

If languages differ with respect to how readily such transfer of reference occurs, then animacy may be more easily dissociated from a lexical item in Japanese than in English. This could account for the fact that the Japanese mismatch cases produced some errors, but nowhere near the high proportions typically reported for number erors. A second possibility is that in the Japanese study, a variety of preamble types were tested. If only the single phrase modifiers from that study are considered—these are analogous to the preambles tested here—the mismatch effect is no longer significant (the error rate is 1.2% for match conditions and 1.75% for mismatch conditions). Hence, it is the more complex preambles that are primarily responsible for the effect. Perhaps if the tag experiment had used structurally more complex preambles, a mismatch effect may have emerged. 2.3.2. NUMBER ERRORS Although errors such as "The bachelor with the overdue bills disappeared, didn't they?" are ambiguous (since, they could be either animate or inanimate), it is still possible to evaluate whether or not an animate plural nonhead causes relatively greater interference than an inanimate plural nonhead. There are a number of different comparisons that may be made to evaluate this. Consider first the difference between the two animacy-mismatched pairs (e.g., conditions containing bachelor-bill vs. bachelor-bills and delivery-neighbor vs. deliveryneighbors); this is the contrast discussed in Bock and Miller (1991) for verb agreement. There is a slight numerical difference in the relative number of plural pronouns produced: For animate-inanimates (e.g., bachelor-bill[s]), a plural nonhead elicits eight errors more than the singular nonhead baseline condition; for inanimate-animates (e.g., delivery-neighbor[s]), a plural nonhead elicits 12 errors (over the baseline condition). The difference between these is not statistically significant. Now consider a comparison of conditions that use the same animate nonhead, paired with an animate versus inanimate head (e.g., bachelor-neighbor[s} vs. delivery-neighbor[s}). The difference between the number mismatch versus number match (baseline) conditions is identical (12 in both cases). Hence, an animate plural nonhead has the same interference effect with both types of head. This is not the case, however, with the inanimate plural nonhead conditions. When conditions in which the nonhead is inanimate are compared (e.g., delivery-bill[s] vs. bachelor-bill[s]), there are relatively more errors when the head is also inanimate. For the inanimate-inanimate conditions, the net difference (plural nonhead

126

Janet L.Nicol

minus singular nonhead) was 25; but for the inanimate-animate conditions, the net difference was 8. Hence, there is a conflict in the pattern of errors: In the animate nonhead conditions, the animacy of the head NP does not appear to affect error incidence, but in the inanimate nonhead conditions, the animacy of the head does appear to affect error rate. The latter effect suggests the type of competition alluded to by Bock and Miller (1991): a head NP that is inanimate is not as "good" (or typical) as one that is animate and is therefore more vulnerable to interference from a plural nonhead. But surely, if this were so, then there should be (at minimum) a corresponding increase in errors when the nonhead is animate, since in this condition, the vulnerable inanimate head would be paired with an animate NP. However, this was not the case. Therefore, there must be another explanation. One possible reason for the relatively larger number mismatch effect with inanimate-inanimate preambles is that it is this condition that is most likely to give rise to a multiple-token interpretation. Recall that Bock, Eberhard, and Cutting (1992) found that multiple-token preambles elicited a greater number of plural pronouns than single-token preambles, which also contained a singular head and plural nonhead. If this is the source of some portion of the errors observed for the inanimate head-inanimate plural nonhead condition, then the error sum for this condition is actually inflated. If this is the case, then the conflict noted above is resolved. This possibility is currently being explored. Overall, the results suggest that in the process of selecting a pronoun that agrees with the head, an animate nonhead has no more power to interfere than does an inanimate nonhead.

3. GENERAL DISCUSSION What we have seen, then, is three sets of studies that show that although a nonhead may interfere with the agreement process between a head NP and another element (a verb or pronoun), an animate nonhead does not induce more errors than an inanimate nonhead. The subject-verb number agreement experiments in English show that the incidence of number errors is unaffected by the animacy (and other features) of the nonhead. The subject-verb animacy agreement experiment in Japanese also shows that an animate nonhead provides no greater interference than an inanimate one. Finally, the subject-pronoun experiment in English showed that an animate nonhead triggers no additional errors in the form of the pronoun—either in animacy errors or number errors. These results add to the body of evidence that supports a model of language production in which there are discrete stages of processing (e.g., Bock and Levelt, 1994; Garrett, 1988; Levelt, 1989). In all, the studies described here support a

Animacy Effects

127

multi-stage model of production in which, for example, the assignment of functional roles (such as head, modifier, object) is isolated from the process by which agreement takes place. Logically, functional-role assignment occurs first; once a NP has been selected as the subject, agreement processes come into play. This stands in contrast to some of the claims about comprehension: It has been argued that the representation of lexical items during language comprehension includes information about, for instance, how good an agent some noun is (based on how frequently it appears as an agent), and that such information is used, in combination with syntactic information, to compute the structure and meaning of a sentence (MacDonald, Pearlmutter, and Seidenberg, 1994). To the extent that this holds true of sentence comprehension, it appears that sentence production is entirely different: some operations may take into account attributes of nouns that are related to frequency of use; others seem to be blind to such attributes. In addition, the agreement processes themselves seem to be differentially sensitive to different kinds of information: Subject-pronominal agreement takes account of conceptual features of the subject; subject-verb agreement attends primarily to syntactic properties of the subject NP.6 This differentiation is also at odds with a system in which all aspects of a representation are available and active. Hence, although input processes may turn out to be best characterized as highly interactive, (at least some) output processes appear to be modular.

ACKNOWLEDGMENTS The author gratefully acknowledges support for this research from grant NIDCD DC-01409 (as part of the National Center for Neurogenic Communication Disorders, University of Arizona), and to the Cognitive Science Program, University of Arizona. Thanks to Toshiyuki Suzuki for making available to me the experimental details and results of the Japanese study. I would also like to thank the following students for assistance with some of the research reported in this chapter: Jason Barker, George Figgs, Sean Hendricks, Meghan O'Donnell, Patience Robinson, and Stephanie Venne.

REFERENCES Anton-Mendez, I. (1996). Clitics and attraction errors: An experimental study of language production. Unpublished ms., University of Arizona, Tucson. Bates, E., and MacWhinney, B. (1989). Functionalism and the Competition Model. In B. MacWhinney and E. Bates (Eds.), The crosslinguistic study of sentence processing (pp. 3-73). Cambridge, UK: Cambridge University Press. Bock, K. (1995). Producing agreement. Current Directions in Psychological Science, 4, 56-61.

128

Janet L.Nicol

Bock, K., and Cutting, J. C. (1992). Regulating mental energy: Performance units in language production. Journal of Memory and Language, 31, 99-127. Bock, K., and Eberhard, K. M. (1993). Meaning, sound and syntax in English number agreement. Language and Cognitive Processes, 8, 57-99. Bock, K., Eberhard, K., and Cutting, J. C. (1992). Controlling number agreement on verbs and anaphors. Presented at the 33rd Annual Meeting of the Psychonomic Society, St. Louis, Missouri. Bock, K., and Levelt, W. (1994). Language production: Grammatical encoding. In M. A. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 945-984). San Diego, CA: Academic Press. Bock, K., Loebell, H., and Morey, R. (1992). From conceptual roles to structural relations: Bridging the syntactic cleft. Psychological Review, 99, 150-171. Bock, K., and Miller, C. A. (1991). Broken agreement. Cognitive Psychology, 23, 45-93. Cutting, J. C., Nicol, J. L., and Bock, J. K. (1995, November). Controlling pronoun number across clause boundaries. Presented at the 36th Annual Meeting of the Psychonomic Society, Los Angeles, CA. Franck, J., and Butterworth, B. (1997, March). Syntactic distance as a key factor in the construction of subject-verb agreement in French. Poster presentation at the tenth Annual Conference on Human Sentence Processing, Santa Monica, CA. Garrett, M. F. (1988). Processes in language production. In F. J. Newmeyer (Ed.), Linguistics: The Cambridge Survey: Vol. 3. Language: Psychological and biological aspects (pp. 69-96). Cambridge, UK: Cambridge University Press. Harrington, M. (1987). Processing transfer: Language-specific processing strategies as a source of interlanguage variation. Applied Psycholinguistics, 8, 351-377. Levelt, W. J. M. (1989). Speaking: From intention to articulation. Cambridge, MA: MIT Press. MacDonald, M., Pearlmutter, N., and Seidenberg, M. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 4, 676-703. Nicol, J. L. (1995). Effects of clausal structure on subject-verb agreement errors. Journal of Psycholinguistic Research, 24, 507-516. Nicol, J. L. (1998). Pronominal feature distinctions in English. Unpublished manuscript. Pollard, C., and Sag, I. A. (1994). Head-driven phrase structure grammar. Chicago: University of Chicago Press. Sasaki, Y. (1991). English and Japanese interlanguage comprehension strategies: An analysis based on the competition model. Applied Psycholinguistics, 12, 47-73. Stemberger, J. P. (1982). Syntactic errors in speech. Journal of Psycholinguistic Research, 11,313-345. Suzuki, T. (1997). Subject-verb agreement of animacy in Japanese. Unpublished manuscript, University of Arizona, Tucson. Suzuki, T., and Nicol, J. L. (1996, March). Broken agreement in animacy. Poster presentation. Ninth Annual Conference on Human Sentence Processing, New York, NY. Vigliocco, G., Butterworth, B., and Garrett, M. F. (1996). Subject-verb agreement in Spanish and English: Differences in the role of conceptual constraints. Cognition, 61, 261-298. Vigliocco, G., Butterworth, B., and Semenza, C. (1995). Constructing subject-verb agreement in speech: The roles of semantic and morphological factors. Journal of Memory and Language, 34, 186-215.

Animacy Effects

129

Vigliocco, G., and Nicol, J. (1997). The role of syntactic tree structure in the construction of subject-verb agreement. Unpublished manuscript.

NOTES 1

A counterargument might be that in (3a), the hash browns is like a proper name and hence is not a real plural any more than a name like Freckles or Daddy Long-Legs is plural. Such terms refer to a singular entity and are syntactically singular. But examples like (3b) and (3c) remain problematic for a purely syntactic account. A different analysis of (3b) is that the equative construction ("An X is a Y" imposes its own constraint on agreement: the two terms of the equation should have the same number, and so the term that represents an exemplar (eggs) takes on a singular "collective" interpretation (this may be aided typicality—eggs may be more typical than an egg; note the relative awkwardness of Steaks is my favorite dinner; it is unusual to eat more than one steak for dinner). A counteranalysis for (3c) is that it is elliptical: that is, the real head noun is implied, but missing: (3c) really says "The act of doing phonology problems and drinking vodka makes me sick." But it's not easy to formalize this notion: it would need to be specified how speakers know which heads may be omitted and which ones may not be (one cannot say, for example, The book is giving a speech to mean The author of the book is giving a speech) and, correspondingly, how listeners know what to fill in. Ultimately, unless examples like (3c) can be assigned an analysis in which the conjoined gerunds can, together, be treated as singular, these present a case in which agreement may be computed via the semantics of the subject. 2 The asymmetry appears to be less robust in languages like Italian and French (Vigliocco et al., 1995; Franck et al., 1997). 3 However, in Italian and Spanish, Vigliocco et al. (1996) found more agreement errors after multiple-token than singular-token items. This raises the possibility that the features of verbs may be specified referentially in languages other than English. 4 Categorically, iru and aru are verbs, and inai and nai are adjectives. 5 Participants attended two sessions: in one, they were asked to provide the "exist" verb (either the animate or inanimate form) and in the other, the "not exist" verb. Half the participants were assigned to one order; half to the other order. 6 One might wonder how agreement is computed in languages with heavily inflected verbs, because the distinction between a verbal inflection and a pronoun is not at all clear. After all, if a verbal affix encodes number, gender, and person, and a pronoun encodes number, gender, and person, they may have a different syntactic status, yet, in the minds of speakers, they may serve a similar function. This is an area that warrants further exploration.

This page intentionally left blank

CONTEXT-INDEPENDENT SENTENCE PROCESSING SUSAN BORSKY* LEWIS P. SHAPIRO* *Department of Communicative Disorders San Diego State University San Diego, California Department of Psychology and Center for Complex Systems Florida Atlantic University Boca Raton, Florida

1. INTRODUCTION The search for invariant features of human language processing motivates many different approaches to psycholinguistic research, and results should, in theory, converge on a single account of the organization and operation of the language comprehension system. However, no single account of language, or even of sentence processing, has received anything near universal acceptance. Indeed, conflicting data from different experimental paradigms have only intensified debate on at least one basic issue : "Is initial sentence processing context independent?" The investigation of the role of context has been, and continues to be, an important focus of psycholinguistic research because different accounts of sentence processing make different predictions regarding context effects. In modular accounts, a sentence is constructed on-line by rapid, automatic subprocesses (e.g., phonological, lexical, syntactic, or discourse analysis) that use only domain-specific information for initial processing (Fodor, 1983; Forster, 1976, 1979). Although it is clear that all kinds of information affect sentence interpretation at some point, such initially context-independent processing could optimize speed and resources by reducing the cognitive and memory demands of Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

131

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

132

Susan Borsky and Lewis P. Shapiro

including multiple sources of information that may not be needed. Of course the initial analysis might be incorrect, but the savings in resources for the many correct interpretations is presumably worth the cost of occasional reanalysis. Modular accounts differ in detail, but generally propose some "simplest parse" heuristic for automatic syntax-driven initial parsing and a mechanism for using context to aid possible reanalysis (e.g., the garden-path model, Frazier, 1987; Frazier and Clifton, 1995; Frazier and Rayner, 1982). The experimental predictions of modularity for real-time sentence processing are initial context independence and later context integration. By contrast, an interactive account does not view different levels of sentence analysis as different processing domains, but as multiple sources of information that can all influence or constrain initial interpretation; ambiguity resolution is attributed to the result of competition among possible interpretations or the satisfaction of multiple lexical constraints (MacDonald, Pearlmutter, and Seidenberg, 1994). There is no need for reanalysis or "second-pass parsing" in interactive accounts because interpretations can be continuously revised. One theoretical problem for such an account is the cognitive and memory demands of including so many sources of information in on-line processing. The consequence of interactive constraint satisfaction is, in effect, continuous integration of context, and the experimental predictions of this theory are immediate context effects on the real-time processing of an unfolding sentence. In this chapter we will report the outcome of a number of investigations of context effects in normal populations, and the kind of data that support different processing accounts. We will then present the results of some recent experiments investigating context integration, and conclude with a discussion of the importance of the experimental task for interpreting data.

2. INVESTIGATING CONTEXT EFFECTS All human languages, by their creative nature, allow multiple continuations and interpretations of an ongoing sentence or discourse; it is the investigation of how different kinds of information affect operations like lexical activation and ambiguity resolution that provide the mainstay of psycholinguistic research. Because sentence context and real-world inferences clearly influence ultimate interpretations, knowledge about the time course of these contextual effects is particularly informative concerning real-time sentence processing and the human language comprehension system. 2.1. Contextual Constraints One measure of contextual influence or constraint is the effect of encountering a word that is anomalous, that is, implausible or uninterpretable in a sentence.

Context-Independent Sentence Processing

133

One such study (Marslen-Wilson, Brown, and Tyler, 1988) compared wordmonitoring times for different types of anomalies in sentences like (1-4) in which the monitoring target was the word beach. Lead in: The nurses walk to their work each morning. (1) Plausible:

They pass the beach on the way to the hospital.

(2) Pragmatic Anomaly: They measure the beach on the way to the hospital. (3) Semantic Anomaly:

They chew the beach on the way to the hospital.

(4) Syntactic Anomaly:

They yawn the beach on the way to the hospital.

The results were a continuum of monitoring times; they were fastest in the plausible context and slowest in the syntactic anomaly, but significantly delayed in all anomalous conditions. This pattern was interpreted as evidence that syntactic and semantic entailments, and even real-world inferences associated with a verb, are used immediately. Later, we will contrast these results with another study using similar materials, but a different task (Lewis, 1996; Lewis, Shapiro, Afton, and Tuller, 1996). Plausibility effects have also been reported for makes-sense word-by-word reading experiments in which subjects indicated the point at which a sentence stopped making sense (Tanenhaus, Boland, Garnsey, and Carlson, 1989). A plausibility-animacy interaction was found for gap filling using a makes-sense, self-paced word-by-word reading paradigm (Stowe, 1989). Also, the results of an eye movement study indicated contextual constraints on structural ambiguity resolution; the animacy of a noun phrase (NP) appeared to facilitate its interpretation as the subject, rather than the object of a verb (Trueswell, Tanenhaus, and Garnsey, 1994; but not Ferreira and Clifton, 1986; Frazier and Rayner, 1982). Contextual constraints on lexical activation have also been reported; results of a naming experiment indicated activation only for the context-appropriate meaning of an ambiguous noun (Simpson and Kreuger, 1991). However, this effect could be attributed to the test method; subjects read a sentence aloud, then named a visually presented probe that was always in sentence-final position where context integration might be expected (Nicol and Swinney, 1989). Contextappropriate versus context-inappropriate probes also produced different patterns of brain activity (evoked potentials) when presented in slowed speech (Van Petten and Kutas, 1987). A series of Italian cross-model lexical decision studies reported selective priming for targets related to a specific feature of the primary meaning of an ambiguous noun when that feature was made salient by the context (Tabossi, Colombo, and Job, 1987; Tabossi and Zardon, 1993). This finding of selective activation did not extend to contexts biased toward the secondary meaning, or to any other types of probes or contexts; the context-independent results of Tabossi and colleagues will be discussed in section 2.2. Note that the studies that reported contextual constraints included several kinds

134

Susan Borsky and Lewis P. Shapiro

of reading measures (including eye movements), as well as word-monitoring and make-sense judgments (except Tabossi and Zardon, 1993). Any conclusions about on-line processing (including interactive contextual constraints) that are based on immediate context effects need to consider whether the experimental task was sensitive to early occurring processes. We will address this issue in more detail later when we compare the theoretical implications and experimental outcomes of different types of tasks. 2.2. Context Independence There is also a considerable body of work that supports a modular account of sentence processing by demonstrating initial context independence for a number of linguistic operations. After briefly describing some previous studies, we will focus on the results of some very recent research. Much support for modularity comes from data that show exhaustive activation of multiple meanings for ambiguous nouns, even when biased by context or frequency. This phenomenon has been demonstrated with cross-modal lexical priming (Love, and Swinney, 1996; Nicol, Swinney, Love, and Hald, under review; Onifer and Swinney, 1981; Swinney, 1979), evoked potentials for speech at normal speeds (Conseco-Gonzales, Hickok, Zurif, Prather, and Stern, 1994), eyetracking (Rayner, Pacht, and Duffy, 1994), and cross-modal naming (Seidenberg, Tanenhaus, Leiman, and Bienkowski, 1982; Tanenhaus, Leiman, and Seidenberg, 1979). Exhaustive activation of all possible thematic grids when a verb is encountered, regardless of context, has been shown using a cross-modal lexical decision interference task (Shapiro, 1997; Shapiro, Zurif, and Grimshaw, 1987, 1989). Context independence for gap filling has been demonstrated by cross-modal lexical priming (Hickok, Conseco-Gonzales, Zurif, and Grimshaw, 1992; Nicol and Swinney, 1989; Swinney, Ford, Frauenfelder, and Bresnan, 1988; Lewis, 1996; Lewis et al., 1996), and evoked potentials (Garnsey, Tanenhaus, and Chapman, 1989; Swinney and Osterhout, 1990). A series of cross-modal lexical priming studies investigated the effect of different types of biasing sentence contexts on the activation of nouns with more than one meaning (Tabossi, 1988; Tabossi et al., 1987; Tabossi and Zardon, 1993). In sentences biased toward the secondary or less common meaning, probes related to both meanings were facilitated (i.e., priming was context independent). However, in sentences like (a) that made salient a specific feature of the primary meaning of a word (e.g., "port"), a probe related to that feature of the biased meaning (e.g., SAFE) showed a selective priming effect; probes related to a feature of the unbiased secondary meaning (e.g., RED for port wine) did not show priming. (a) The violent hurricane did not damage the ships that were in the port, one of the best equipped along the coast. (b) The man had to be at five o'clock at the port for a very important meeting.

Context-Independent Sentence Processing

135

By contrast, priming for both meanings was produced by sentences like (b) that were plausible only for the primary meaning, but did not prime a feature of that meaning. Moreover, when probes were directly associated with word meanings (e.g., SEA, LIQUEUR for "port") rather than salient features (e.g., SAFE, RED), even feature-priming contexts like (a) produced priming for both meanings. Although these experiments raise some interesting questions about the role of frequency, the results are consistent with other findings of context independent access to multiple meanings of ambiguous nouns. The specific circumstances that produced a context effect suggest that the source of selective facilitation was priming by prior context for the feature-related probe, and not selective activation for the biased meaning. Both modular and interactive accounts generally agree that an analysis of sentence structure does not wait for disambiguating information; a "working hypothesis" is computed on-line, as a sentence unfolds in time. A description of this analysis as a domain-specific syntax-driven process came initially from eyemovement studies and led to the development of the garden-path model of sentence processing, (e.g., Ferreira and Clifton, 1986; Frazier and Rayner, 1982; Rayner, Carlson, and Frazier, 1983). For example, eye movements were recording while subjects read temporarily ambiguous sentences like (5) and (6), in which "answer" could be attached as simple direct object (NP) or as a new subject in a more complex embedded clause structure (CP). (5) Simpler NP Attachment:

John knew the answer [to the question].

(6) More Complex CP Attachment: John knew the answer [was correct]. At the bracketed disambiguation region, the results were that both the average reading times and the probability of regressive eye movements were greater for sentences like (6), where simple attachment proved incorrect. These results were interpreted as evidence of an automatic "simplest parse" attachment at the earliest point licensed by the syntax, and subsequent reanalysis when later arriving information required a different syntactic structure. The precise description of a "simplest parse" varies among theories; there are even some indications that the first pass parse may be determined by individual pre-existing preferences (as opposed to online contextual constraints) (Shapiro, Nagel, and Levine, 1993). Notwithstanding minor differences, the context independence of many linguistic operations has been well established by previous research; a number of new studies that we will describe in greater detail serve to extend and elaborate a primarily modular account. 2.2.1. LEXICAL ACCESS AND GAP FILLING Because many reports of contextual constraints come from tasks such as word monitoring, word-by-word reading, or makes-sense judgments, and many contrasting reports of context-independent processing come from cross-modal lexical

136

Susan Borsky and Lewis P. Shapiro

priming (CMLP), this study (Lewis, 1996; Lewis et al., 1996) intentionally used materials similar to the word-monitoring study described earlier (Marslen-Wilson et al., 1988) but used CMLP to assess lexical activation. Consider examples (7-9) and their similarity to the Plausible, Semantic, and Syntactic Anomaly Conditions in (1-4) from the word-monitoring experiments. (7) Plausible:

Norman charged [1] the interest [2] monthly.

(8) Implausible:

Norman performed [1] the interest [2] monthly.

(9) Nonsyntactic:

Norman arrived [1] the interest [2] monthly.

The procedure for CMLP was as follows: While subjects listened to the auditory sentences for comprehension, a visually presented letter string (probe) appeared briefly, and subjects pressed a "word" or "non word" button as quickly as possible; response times (RT) were recorded. In experimental sentences1 such as (7-9) the probe was either a word related to a meaning of "interest" (e.g., CONCERN) or an unrelated control word (e.g., LUGGAGE); in separate trials, each probe appeared at two probe positions, [1] at the offset of the verb and [2] at the offset of a potential object. Priming was defined as the reaction time (RT) difference between the related and control probes at each position. The results showed no significant priming at [1], but significant priming at [2] in all three conditions. In other words, there was no context effect on the lexical activation of the potential target. Recall that Marslen-Wilson et al. (1988) found faster wordmonitoring times for the plausible object condition and significantly slower times for all of the anomalous conditions, and interpreted these results as evidence of interactive context-dependent processing; this conclusion was not borne out by the CMLP data. The interpretation of priming as evidence of lexical activation cannot be alternatively explained as a context effect because the related CMLP probes were always associated with a meaning that was not biased by the plausible sentence target (e.g., CONCERN is not related to the biased "money" meaning of "interest"). A recent criticism of CMLP results as an artifact of continuation priming rather than lexical access (McKoon and Ratcliff, 1994) also cannot explain these data because no priming was found at point [1] immediately before the target (see also Nicol, et al., 1997). Given the just described finding that a preceding verb does not constrain lexical access for a potential object, lexical reactivation at an object gap was investigated. The same sentence sets illustrated in (7-9) were converted into wh-questions like (10-12) that licensed a gap after the verb in the plausible and implausible sentences (but not after the intransitive verbs in the non syntactic condition). (10)

Which interest did Norman [1] charge

(11)

Which interest did Norman [1] perform

(12)

Which interest did Norman [1] arrive

[2] monthly? [2] monthly? [2] monthly?

Context-Independent Sentence Processing

137

Sentences of this form require a postverb object position (gap) to be "filled" by activation of an object occurring elsewhere in the sentence. Note that intransitive verbs like arrive in the nonsyntactic condition have no object position, hence no gap. To test for verb constraints on gap filling, the same CMLP procedure was used to measure priming at pregap [1] and gap [2] positions. The result was priming at the gap for plausible and implausible sentences, but not for the nonsyntactic (no gap) sentences; lexical reactivation was not constrained by context; it occurred automatically when a gap was licensed by the syntax. 2.2.2. SYNTACTIC CATEGORIZATION Some words have meanings that depend on syntactic category. For example drop off can be a base verb with the literal meaning "drop" plus the preposition "off"; alternatively, drop off can be a verb-particle that has a figurative interpretation. In light of previous findings of exhaustive activation for multiple meanings of ambiguous nouns, an experiment was designed to investigate whether sentence context constrains the syntactic categorization of a lexical item, and consequently its interpretation (Shapiro and Ortega, 1997). Observe the meaning of "drop off" in (13-15), noting the split verb-particle construction in (15). (13)

Base Verb, Literal Context: Carol was resting on a narrow ledge after rock climbing over a river. Suddenly, she lost her footing. Carol dropped [1] off [2] the narrow ridge [3] into the water.

(14) Verb-particle, Figurative Context: Tim had an important meeting and realized that all his shirts were dirty. He decided to go to the dry cleaners. Tim dropped [1] off [2] his laundry [3] on his way to work. (15)

Split Verb-particle, Figurative Context: Tim had an important meeting and realized that all his shirts were dirty. He decided to go to the dry-cleaners. Tim dropped [1] his laundry [2] off [3] on his way to work.

The CMLP procedure described earlier was used to assess the activation of the different interpretations at points [1], [2], and [3]. In this case, the probes were related to either the literal (base verb) or the figurative (verb-particle) interpretation or were similar but unrelated controls. The results were priming for the literal meaning right after the base verb in every context, and priming for the figurative sense after the potential particle in every context; there was no priming for either sense after the intervening direct object in (15). These response patterns support a context-independent, modular account in which an interpretation is activated when the processor encounters a lexical item that licenses that activation. The figurative sense was not simply activated more slowly; in the split particle case it was primed only when a possible particle was encountered, regardless of elapsed time or sentence bias.

138

Susan Borsky and Lewis P. Shapiro

2.2.3. TYPICALITY STRUCTURE Most studies of lexical activation look at context effects on the selection of alternative interpretations of a sentence element; now the effect of sentence context on the activation of different exemplars of a category has also been investigated (Raczaszek, 1995; Raczaszek, Shapiro, Tuller, and Kelso, 1998). According to prototype theory (Rosch, 1975), objects in the world come in natural groupings; a prototypical member of a category is one that has many features in common with other members, and few features in common with nonmembers of the category. Thus, an exemplar of the category FRUIT may be typical (e.g., APPLE) or atypical (e.g., LEMON). There is a well-established typicality effect whereby encountering a category such as FRUIT will result in greater activation or facilitation of typical exemplars than atypical ones. However, categories and their internal typicality structure are also flexible (Barsalou, 1982), which raises at least two issues for psycholinguists: (a) "Do categories have invariant typicality structures that are initially independent of sentence context?" and (b) "When in the time course of ongoing sentence processing can sentence context influence internal category structure to modify the typicality effect?" The typicality effect was assessed by a cross-modal lexical decision (CMLD) interference task in which the procedure was the same as described for CMLP, but the critical comparison was the comparative response times for probes related to typical or atypical exemplars in different contexts. Consider the following examples: (16) Neutral Context:

(17) Atypical-biased Context:

Response times for the two probe types were compared at the three indicated probe positions: 0 ms (at the offset of the category name), and 450 ms and 750 ms later. If categories are indeed formed "on the fly"; that is, are formed on the basis of contextual constraints, then the typicality effect (faster response times for the typical exemplar) should reverse in the atypical-biased contexts. On the other hand, if categories have some invariant or pre-existing structure that is initially insensitive to sentence context, the typical probe should produce faster RTs regardless of context—at least in the temporal vicinity of the category word. Although, based on other on-line studies, context would be expected to exert its influence at later probe positions. At 0 ms, RTs showed similar facilitation for the typical probe (e.g., APPLE) in both contexts; that is, even when biased (e.g., by "sour fruit") toward the atypical exemplar (e.g., LEMON). At 450 ms, the typicality effect was still significant in

Context-Independent Sentence Processing

139

the neutral context, but in the atypical-biased context, RTs for atypical and typical probes were equally fast (i.e., both were facilitated).2 Finally, by 750 ms there were no significant effects. An additional experiment investigated the nature of the change in typicality structure that was seen at 450 ms by comparing lexical decision times for an atypical probe at five points within the original 0-750-ms range. The RTs at 0, 150, and 300 ms were similar, but decreased sharply by 450 ms after the category word and the variability of the RTs peaked at 300 ms, that is, before the new typicality structure was evident. This kind of nonlinear pattern does not indicate a smooth rise in activation for the atypical exemplar, but a phase transition between two relatively stable structures. The implications for sentence processing are that an atypical-biased context (a) does not initially affect typicality structure, even though a category is accessed by the lexical activation of a category word, (b) produces a period of unstable contextual adaptation by 300 ms, and (c) finally (by 450 ms) results in a new, contextually adapted, typicality structure in which the previously atypical exemplar is now typical. 2.2.4. STRUCTURAL CONTEXT EFFECTS: VP-ELLIPSES Information about the on-line processing of long-distance dependencies can tell us something about how a sentence processor assigns reference. To investigate these long-distance dependencies, a recent series of CMLP experiments is testing gap filling in complex VP-ellipses (Shapiro and Hestvik, 1995). Consider the following examples: (18)

Coordinated VP-Ellipsis: The policeman defended himself and the fireman [1] did [2] too, according to someone [3] who was there.

(19) Subordinated VP-Ellipsis: The policeman defended himself because the fireman [1] did [2] too, according to someone [3] who was there. Notice that both can have two interpretations: (a) known as the "strict" reading in which the fireman defended the policeman or (b) known as the "sloppy" reading in which the fireman defended himself (the fireman). Although the two readings of (18) and (19) seem to be available both on- and off-line, a study is in progress to investigate the effect of reflexive verbs such as "perjure" that do not (at least ultimately) allow the strict reading (i.e., the fireman can only perjure himself, not the policeman). The on-line activation of both readings would indicate a contextindependent gap-filling operation; activation only for the sloppy reading would be an example of a lexical constraint, although it is unclear whether the constraint would be semantic (associated with the specific meaning of the verb) or syntactic (the consequence of encountering a verb class "reflexives"). The study that will be reported here investigated the time course of gap filling

140

Susan Borsky and Lewis P. Shapiro

by looking at priming for the "strict" interpretation in coordinated (e.g., and) and subordinated (e.g., because) VP-Ellipses in sentences like (18-19). Interestingly, the linguistic differences between coordinated and subordinated structures seem to entail on-line processing differences as well. In the coordinated case (18) gap filling was observed right at the elided NP (after "did"). By contrast, in the subordinated structure (19) evidence of gap fiLing was observed not at the gap, but a few syllables later (after "someone"). These results suggest that in the coordinated case, gap filling is automatic, driven by the syntactic nature of the traceantecedent relation. However, in the subordinated case, in order to interpret the second clause a causal relation must be computed between the two clauses (X happened because Y happened); the results suggest that using this extra-syntactic information from semantics or discourse to compute causal relations delays the gap-filling operation. 2.2.5. STRUCTURAL CONTEXT EFFECTS: WH-QUESTIONS The distinction between two types of wh-questions also seems to have processing implications, according to the results of a CMLP study of gap filling (Shapiro, Thompson, and Borsky, 1998). Consider examples (20-21): (20) The soldier is pushing the unruly student violently into the street. Who is the soldier pushing violently into the street? (21) The soldier is pushing the unruly student violently into the street. Which student is the soldier [1] pushing [2] violently [3] into the street? Intuitively (even without linguistic details), the two questions require different sorts of reference. Although both types can, and often do, refer to prior discourse, who (and what) questions do not require it; for example "Who [in the heck] is the soldier pushing?" Which-NP questions do require explicit reference to prior discourse; that is, "Which student [of all the students in the group under consideration] is the soldier pushing?" (contrast the incongruity of "Which student [in the heck] is the soldier pushing?"). The CMLP results for who/what-questions like (20) showed immediate gap filling after the verb, but which-NP questions like (21) showed gap filling that was delayed by a few syllables. These results suggest that when gap filling is driven solely by syntactic considerations (e.g., in coordinated VP-ellipses and who/what questions) processing is automatic and immediate effects are observed. But, when extrasyntactic information is involved (e.g., in subordinated VP-ellipses and which-NP questions) gap filling is delayed. Is this a context effect? Yes, in the sense that the sentence processor appears to be sensitive to the structural differences entailed by a lexical item (e.g., because, or which), yet these findings are still congruent with a syntax-driven account in which linguistic operations like gap filling are immediate and automatic when licensed by the syntax.

Context-Independent Sentence Processing

141

2.2.6. PROSODIC EFFECTS ON ATTACHMENT A linguistic role for sentence prosody has been largely ignored in the psycholinguistic literature until recently. Several new studies suggest that the timing and intonation pattern of an utterance can provide important information for positing syntactic gaps and resolving attachment ambiguities. A CMLP study (Nagel, Shapiro, and Nawy, 1994) used sentences like (22-23), in which the structural interpretation of the ongoing sentence depended on whether or not a gap was posited at the first syntactically licensed position. Observe the position of the gap in the following sentences. (22) Which doctori did the supervisor call # to get help for crisis? (23) Which doctori did the supervisor call daughter?

i

i

during the

# to get help for his youngest

The lexical content of the auditory sentences was identical up to (and beyond) the first potential gap, but there were differences in the naturally occurring prosody of the sentence reader.3 Evidence for activation of the potential filler was assessed by priming at point #, after the main verb, where a syntax-driven account would predict the positing of a gap, regardless of prosody. Results showed priming for the potential filler (interpreted as positing a gap) only in the gap condition (23) even though the only difference at that point was appropriate prosody. A second study (Nagel, Shapiro, Tuller, and Nawy, 1996) tested prosodic influences on the resolution of attachment ambiguities in sentences like the following: (24)

The company owner promised the wage increase to # the workers.

(25) The company owner promised the wage increase would # be substantial. The garden-path theory and other "simplest parse" accounts predict initial attachment of "the wage increase" as a direct object in both cases; in (25) the need for reanalysis becomes apparent at the disambiguating word "would" and should produce greater CMLD times at that point, regardless of prosody. To test for prosodic effects, two new sentences were formed by switching the initial segments, up to and including the main verb (the lexical content is identical to that point). There was a significant effect of prosody for CMLD times at the offset of the disambiguating word; it took subjects longer to execute a lexical decision when prosody was incongruent with the subsequent material. These results suggest that appropriate prosody may help listeners avoid incorrect analyses. 2.2.7. PROSODIC EFFECTS ON SYNTACTIC BOUNDARIES Another study investigated the effect of prosody on the interpretation of a sentence in which the overall structure was clear, but the location of a syntactic

142

Susan Borsky and Lewis P. Shapiro

boundary was ambiguous (Kelso, Case, Holroyd, Horvath, Raczaszek, Tuller, and Ding, 1994; Raczaszek, 1995). The two interpretations of the sentence are indicated by the bracketing in (28-29). (28) [Pat or Kate] [and Bob] will come. (PK)B (29)

[Pat] [or Kate and Bob] will come. P(KB)

When the sentence is partitioned as indicated in (28), it has the meaning "Either Pat or Kate will come, and in any event Bob will come, too"; when partitioned as indicated in (29), the identical sentence has the alternate meaning "Pat will come, or else Kate and Bob will both come." The two interpretations do not differ in overall sentence structure, only in the location of the boundary between constituents of the compound subject. In natural speech, this boundary appears to be cued by at least two acoustic features: longer duration for the last stressed vowel before the boundary, and longer pause duration at the boundary itself. The manipulation of these two acoustic parameters alone reliably biased interpretation in the expected direction, confirming the salience of these cues. However, the most interesting results of this acoustic/prosodic manipulation were the perceptual patterns that emerged when an acoustic continua between the two percepts was presented sequentially; the patterns were very similar to the results of corresponding manipulations on basic perceptual categorization in both auditory (Case, Tuller, Ding, and Kelso, 1995) and visual modalities (Hock, Kelso, and Schoner, 1993). This similarity suggests that at least some elements of sentence prosody may be so closely tied to the acoustic signal that they are not separable in the same sense as other linguistic domains like semantics or syntax. There is not yet enough evidence that acoustic cues for prosody are reliably produced and consistently used to justify a primarily prosody-driven model of sentence processing, or a special role for prosody, but recent results are providing an exciting new direction for ongoing research.

2.2.8. SUMMARY The study of lexical activation and gap filling (Lewis, 1996; Lewis et al., 1996) found that verb information does not appear to constrain the initial activation of potential arguments, or fillers for syntactically licensed gaps. These findings strongly suggest that at some level of analysis, access to the properties of verbs and access to the meanings of potential arguments are independent. The study of verb-particle constructions (Shapiro and Ortega, 1998) showed that initial syntactic categorization and the activation of the figurative meaning of a verb in a verbparticle construction does not depend on sentence context, but on encountering a syntactically appropriate lexical item (i.e., a possible particle). The study of typicality structure (Raczaszek, 1995; Raczaszek et al., 1998) demonstrated the initial

Context-Independent Sentence Processing

143

context impenetrability of internal category structure and also provided a fascinating moving picture of a context-integration process. The finding of extrasyntactic context effects that seem to relate to inferences about causal relationships in subordinated VP-ellipsis (Shapiro andHestvik, 1995) is still very new, and part of an ongoing study. Nevertheless, this line of research may expand our understanding of automatic processing domains. The finding of delayed gap filling in which-NP questions (but not who/what questions) (Shapiro et al., 1998) is also very new, and continued investigation can increase our understanding of sentence processing that involves discourse reference. The investigation of prosodic context effects is also an exciting new area, but the generalizability of a prosody-driven account of spoken language and in particular the availability of effective prosodic cues in speech await further research. The suggestion that prosody may operate at a very early perceptual level also has implications for the organization of basic perceptual systems. In contrast to most studies that report immediate context effects, each of these studies used cross-modal lexical decision tasks that do not involve any judgments, implicit or explicit, about how a lexical item fits into the sentence context. Further support for modular processing comes from experiments that do find context effects, but downstream, presumably as lexical items that were encountered earlier are integrated into an ongoing sentence analysis. Context integration may also be required for certain kinds of tasks. The following section will describe some studies that investigated context integration. 2.3. Context Integration The continuing debate about context effects on sentence processing has never been about the ultimate influence of context, but about the timing of this influence and the nature of the language comprehension process. Much of the data cited as evidence of interactive initial processing are interpreted by proponents of modularity as context integration. The following experiments investigated the kind of information that is integrated during sentence processing, and under what circumstances. 2.3.1. THE EFFECT OF PREPOSED ARGUMENTS ON VERB INTEGRATION This study used a self-paced word-by-word reading task to test the effect of preposed arguments on verb integration during ongoing sentence analysis (Borsky, Lewis, and Shapiro, 1995). The previous finding that all possible argument structures for a verb appear to be momentarily activated at the verb, regardless of context (Shapiro et al., 1987, 1989; Rubin, Newhoff, Peach, and Shapiro, 1996), suggests that this information is an integral part of a verb's representation. If the appearance of potential arguments before the verb could facilitate selection of the

144

Susan Borsky and Lewis P. Shapiro

appropriate argument structure, then preposed arguments should result in faster verb integration during ongoing sentence processing; and such results would support an argument-driven rather than a verb-driven account of sentence processing. The following sentences were constructed with either optional (e.g., return) or obligatory (e.g., give) three-place verbs and one, two, or three preposed arguments. (30)

One preposed argument: It was the book that was returned to Mary by John four days ago.

(31)

Two preposed arguments: It was to Mary that the book was returned by John four days ago.

(32)

Three preposed arguments: It was from John to Mary that the book was returned four days ago.

The results for subject-controlled reading times at each word were as follows: the number of preposed arguments did facilitate the integration of the verb, but only for optional three-place verbs, and verb integration was sensitive only to the number of preposed arguments, and not to the particular argument (first, second, third; e.g., agent, theme, goal). These results provided some support to an argumentdriven account of sentence processing, but the interaction between the number of preposed arguments and verb type (optional or obligatory three-place verbs) suggests a model that combines elements of both verb and argument-driven accounts. The distinction between the effect of the number of arguments and their role (which may be more closely related to semantic properties than thematic grids) suggests different domains, even during early stages of verb integration. 2.3.2. PHONEME IDENTIFICATION DURING SENTENCE PROCESSING Sentence processing research has focused mainly on the effect of sentence context on lexical and syntactic operations, but for spoken language, the acoustic signal itself presents potential ambiguities for phoneme categorization and word identification. The acoustic boundaries between phonemes can be influenced by context, including sentence meaning (Connine, 1987). To explore this effect in an ongoing sentence, 10 target stimuli from a 10-step acoustic continuum that could be heard as "goat" or "coat" were embedded in biased sentences (Borsky, Tuller, and Shapiro, 1998). (33)

GOAT-biased: The laughing dairyman hurried to milk the [GOAT/COAT] in the drafty barn.

(34)

COAT-biased: The expert tailor tried to shorten the [GOAT/COAT] in the cluttered attic.

Context-Independent Sentence Processing

145

Identifications and response times at the offset of the target stimulus were obtained while the sentence continued uninterrupted. Results showed a biasing effect of context on the identification of ambiguous midrange target stimuli, and processing differences were indicated by an acoustic-context interaction for response times. More importantly, there was a further interaction between the acoustic value and the congruency of the identification with sentence context; identifications that fit the sentence were faster than those that didn't fit (i.e., an anomaly effect), but again, only for the potentially ambiguous midrange stimuli. Although there is considerable debate about the timing and mechanism of sentence context effects, many kinds of contextual manipulations appear to affect the identification of speech sounds. Because identification and congruency judgments require conscious reflection and an explicit decision response, we can only say that context does influence the processing of ambiguous speech segments in an ongoing sentence, but not necessarily immediately, as interactionists would claim. The outcome of a recent experiment (Borsky et al., 1998) using the same materials but a CMLD interference task suggests that acoustic information influences phonological processing before sentence context. 2.3.3. A COMPARISON OF WORD-BY-WORD READING AND CMLP Reading tasks were described earlier as integrative tasks that are less sensitive to automatic processes than CMLP. However, a better test of task differences is to present the same materials for different behavioral measures and compare the results (Lewis, 1996; Lewis et al., 1996). The same plausible, implausible, and nonsyntactic sentences (7-9; 10-12) from the CMLP studies described earlier in section 2.2.1 were presented as a noncumulative word-by-word moving window display, and subject-controlled reading times were recorded. The results showed a sensitivity to the syntactic anomalies that was not found for the CMLP task. This finding is consistent with our claim that interference tasks like CMLP and CMLD tap processes that occur earlier than those tapped by reading tasks, which are more appropriate for studying integration.

3. THE IMPORTANCE OF EXPERIMENTAL TASKS FOR INTERPRETING DATA Psycholinguists seem to agree that "on-line" measures are essential for investigating real-time sentence processing. The underlying assumption of on-line tasks is that presenting a stimulus, or requiring a response at a certain point in time is a reliable measure of immediate processing effects. Let's look again at

146

Susan Borsky and Lewis P. Shapiro

some experimental tasks and the kind of data they produce. Reading tasks, makessense judgments, and word-monitoring experiments typically find context effects, which interactionists cite as evidence of contextual constraints. Proponents of modularity frequently refute such interpretations by attributing the results to context integration, arguing that integration is not simply a function of time, but a response to any task that involves conscious reflection or judgment about the relation between a stimulus and its context, or even an implicit need to resolve ambiguities as correctly as possible at the end of a phrase or sentence. By contrast, tasks such as CMLD and CMLP typically find context independence. Proponents of modularity claim that these paradigms are less subject to integration effects because they do not require conscious reflection about the stimulus-context relation and are therefore sensitive to immediate automatic processing. However, proponents of interactive contextual constraints attribute context "insensitivity" to insufficient context. Just to confuse the issue further, the outcome of some other measures (e.g., evoked potentials and eye tracking) are somewhat mixed, but the experimental paradigms used with these measures vary considerably. Nevertheless, the strongest case in favor of context-independent linguistic operations as a valid interpretation of many cross-modal lexical decision experiments comes from finding context effects, but only "downstream" (e.g., Raczaszek, 1995; Raczaszek et al., 1998; Swinney, 1979), or from getting different results for similar materials (contrast Marslen-Wilson et al., 1988, and Lewis, 1996; Lewis et al., 1996). Even identical sentences produced different results: word-by-word reading showed context effects, but CMLP did not. Much remains to be learned about on-line sentence processing (for example, about the role of prosody), and new paradigms are a welcome addition to the investigative arsenal (e.g., compressed speech stimuli: Fodor, Ni, Grain, and Shankweiler, 1996; the auditory moving window: Ferreira, Henderson, Anes, Weeks, and McFarlane, 1996). Technological advances in physiological measures and functional brain imaging are providing an increasingly accessible and important source of data. No single experimental paradigm is best, and every method has its limitations, but the interpretation of all data must critically and cautiously take into account the theoretical and empirical characteristics of the experimental task.

4. CONCLUSIONS We began with the question "Is initial sentence processing context independent?" The preponderance of the evidence says "yes," sentence processing is, at least for some operations, context independent. "Is initial sentence processing independent of all contextual constraints and under all circumstances?" The answer is just as clearly "no." All kinds of contextual influences from sentence,

Context-Independent Sentence Processing

147

discourse, or real-world knowledge and experience undoubtedly affect interpretation at some point. The debate has always been about the time course of context effects and their implication for a larger account of the organization and implementation of our sentence-processing system. We have already emphasized that task selection is crucial for collecting data that inform us about rapid automatic processes. The speed and efficiency of language comprehension demands that a theoretical model be equally rapid and efficient. Certainly, the assignment of many common linguistic operations to domain-specific modules that can operate automatically on limited information according to admittedly arbitrary rules has the capacity for optimizing speed and resources—but only to the extent that errors can be held to an acceptable level and corrected quickly. Modular accounts address those issues by continually refining the criteria for automatic processing. For example, criteria for nonautomatic processing can actually strengthen a modular model, and some of those criteria may involve some of the findings described in this chapter, such as, syntactic markers for causal inference or discourse reference to slow or "short-circuit" automatic processes (Shapiro and Hestvik, 1995; Shapiro, Thompson, and Borsky, 1998); prosodic cues that may sometimes override automatic parsing to avoid acoustically incongruent analyses (Nagel et al., 1994, 1996); and for tasks that require explicit or implicit congruency judgments to trigger immediate context integration. For this reason experimental findings of contextual constraints in specific circumstances do not seriously challenge current modular accounts. Scientific knowledge cannot advance on theory alone, and the most compelling evidence for context independence comes not from theoretical arguments, but from sound research. Examples abound for results that show initial context independence and context effects a little downstream; context effects that disappear with tasks that are more sensitive to early occurring processes (e.g., CMLD; CMLP) may need to be interpreted differently. The addition of multiple probe points to an increasing number of studies is helping to identify contextindependent processes, clarify the boundaries of processing domains, and extend our knowledge of context integration. Certainly, more cross-language experiments will also help psycholinguists draw closer to the goal of discovering, describing, and ultimately of understanding the universal principles of human language.

REFERENCES Barsalou, L. W. (1982). Context-independent and context-dependent information in concepts. Memory and Cognition, 10, 82-93. Borsky, S., Lewis, J. R., and Shapiro, L. P. (1995). The effect of preposed arguments on

148

Susan Borsky and Lewis P. Shapiro

verb integration times during ongoing sentential analysis. Poster presented at the Eighth Annual CUNY Conference on Human Sentence Processing. Tucson, AZ. Borsky, S., Tuller, B., and Shapiro, L. P. (1998). "How to milk a coat:" The effects of semantic and acoustic information on phoneme categorization. The Journal of the Acoustical Society of America, 103(5), 1-7. Borsky, S., Shapiro, L. P., Tuller, B., Wolf, K., and Langford, C. (1998). When do you milk a coat? The time course of acoustic and semantic information in sentences. Poster presentation at the Eleventh Annual CUNY Conference on Human Sentence Processing, New Brunswick, NJ. Case, P., Tuller, B., Ding, M., and Kelso, J. A. S. (1995). Evaluation of a dynamical model of speech perception. Perception and Psychophysics, 57(7), 977-988. Connine, C. M. (1987). Constraints on interactive processes in auditory word recognition: The role of sentence context. Journal of Memory and Language. 26, 527-538. Conseco-Gonzales, E., Hickok, G., Zurif, E., Prather, P., and Stern, C. (1994). Electrophysiological analysis of lexical access. Unpublished manuscript. Ferreira, E, and Clifton, C. (1986). The independence of syntactic processing. Journal of Memory and Language, 25, 348-368. Ferreira, E, Henderson, J. M., Anes, M. D., Weeks, Jr., P. A., and McFarlane, D. K. (1996). Effects of lexical frequency and syntactic complexity in spoken-language comprehension: Evidence from the auditory moving-window technique. Journal of Experimental Psychology, 22, 324-335. Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MIT Press. Fodor, J., Ni, W. Grain, S., and Shankweiler, D. (1995). Tasks and timing in the perception of linguistic anomaly. Unpublished manuscript, Haskins Laboratories, CUNY Graduate Center, University of Maryland, University of Connecticut. Forster, K. I. (1976). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimental Psychology 33, 465-495. Forster, K. I. (1979). Levels of structure of the language processor. In W. Cooper and E. Walker, (Eds.), Sentence processing. Hillsdale, NJ: L. Erlbaum Associates. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 559-586). Hillsdale, NJ: Erlbaum. Frazier, L., and Clifton, C. (1995). Construal. Cambridge, MA: MIT Press. Frazier, L., and Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Frazier, L., and Rayner, K. (1987). Resolution of syntactic category ambiguities: Eye movements in parsing lexically ambiguous sentences. Journal of Memory and language, 26, 505-526. Garnsey, S. M., Tanenhaus, M. K., and Chapman, R. M. (1989). Evoked potentials and the study of sentence comprehension. Journal of Psycholinguistic Research, 18, 51-60. Hickok, G., Conseco-Gonzales, E., and Grimshaw, J. (1992). Modularity in locating whgaps. Journal of Psycholinguistic Research, 21, 545-561. Hock, H., Kelso, J. A. S., and Schoner, G. (1993). Bistability, hysteresis and loss of tern-

Context-Independent Sentence Processing

149

poral stability in the perceptual organization of apparent motion. Journal of Experimental Psychology: Human Perception and Performance, 19, 63-80. Kelso, J. A. S., Case, P., Holroyd, T., Horvath, E., Raczaszek, J., Tuller, B., and Ding, M. (1994). Multistability and metastability in perceptual and brain dynamics. In P. Kruse, M. Stadler (Eds.), Ambiguity in mind and nature. Berlin: Springer-Verlag. Lewis, J. R. (1996). Local contextual impenetrability of lexical access and gap-filling: A comparison of outcomes from cross-modal priming and word-by-word reading tasks. Unpublished doctoral dissertation, Florida Atlantic University, Boca Raton, FL. Lewis, J. R., Shapiro, L. P., Afton, R., and Tuller, B. (1996). The magnitude of cross-modal lexical priming as a Junction of the verb-object relationship in a sentence. Poster session presented at the Ninth Annual CUNY conference on human sentence processing, New York, NY. Love, T., and Swinney, D. (1996). Coreference processing and levels of analysis in objectrelative constructions: Demonstration of antecedent reactivation with the cross-modal priming paradigm. Journal of Psycholinguistic Research, 25, 5-24. MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). Lexical nature of syntactic ambiguity resolution. Psychological Review, 101(4), 676-703. Marslen-Wilson, W., Brown, C. M., and Tyler, L. K. (1988). Lexical representations in spoken language comprehension. Language and Cognitive Processes, 3,1-16. McKoon, G., and Ratcliff, R. (1994). Sentential context and on-line lexical decision. Journal of Experimental Psychology: Learning, Memory and Cognition, 20, 1239-1243. Nagel, H. N., Shapiro, L. P., and Nawy, R. (1994). Prosody and the processing of filler-gap sentences. Journal of Psycholinguistic Research, 23, 473-485. Nagel, H. N., Shapiro, L. P., Tuller, B., and Nawy, R. (1996). Prosodic influences on the resolution of temporary ambiguity during on-line sentence processing. Journal of Psycholinguistic Research, 25, 319-344. Nicol, J., and Swinney, D. (1989). The role of structure in coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18, 5-19. Nicol, J., Swinney, D., Love, T, and Hald, L. (1997). Examination of sentence processing with continuous vs. interrupted presentation paradigms. Center for Human Information Processing Technical Report #97-2. (University of California, San Diego). Onifer, W., and Swinney, D. (1981). Accessing logical ambiguities during sentence comprehension: Effects of frequency-of-meaning and contextual bias. Memory & Cognition, 9, 225-236. Raczaszek, J. (1995). Contextual influences on real-time processing of linguistic stimuli: Traditional and dynamical approaches. Unpublished doctoral dissertation, Florida Atlantic University, Boca Raton. Raczaszek, J., Shapiro, L. P., Tuller, B., and Kelso, J. A. S. (1998). Category names in context: On-line adaptation. Unpublished manuscript. Rayner, K., Carlson, M., and Frazier, L. (1983). The interaction of syntax and semantics during sentence processing. Journal of Verbal learning and Verbal Behavior, 22, 358-374. Rayner, K., Pacht, J. M., and Duffy, S. A. (1994). Effects of prior encounter and global discourse bias on the processing of lexically ambiguous words: Evidence from eye fixations. Journal of Memory and Language, 33, 527-544.

150

Susan Borsky and Lewis P. Shapiro

Rosch, E. (1975). Cognitive representations of semantic categories. Journal of Experimental Psychology: General, 104, 192-233. Rubin, S. S., Newhoff, M, Peach, R. K., and Shapiro, L. P. (1996). Electrophysiological indices of lexical processing: The effects of verb complexity and age. Journal of Speech and Hearing Research, 1071-1080. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., and Bienkowski, M. (1982). Automatic access of the meaning of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Shapiro, L. P. (1997). Tutorial: An introduction to syntax. Journal of Speech, Language, and Hearing Research, 40, ?-?). Shapiro, L. P., and Hestvik, A. (1995). On-line comprehension of VP-Ellipsis: Syntactic reconstruction and semantic influence. Journal of psycholinguistic Research, 24, 517-532. Shapiro, L. P., Nagel, H. N., and Levine, B. A. (1993). Preferences for a verb's complements and their use in sentence processing. Journal of Memory & Language, 32,96114. Shapiro, L. P., and Ortega, K. (1998). The time-course of comprehension of verb-particle constructions. Unpublished manuscript, San Diego State University. Shapiro, L. P., Thompson, C. K., and Borsky, S. (1998). On-line processing of who/what and which-NP questions in normal listeners: Evidence for differential time-course of (re)activation. Unpublished manuscript, San Diego State University. Shapiro, L. P., Zurif, E. B., and Grimshaw, J. (1987). Sentence processing and the mental representation of verbs. Cognition, 27, 219-246. Shapiro, L. P., Zurif, E. B., and Grimshaw, J. (1989). Verb processing during sentence comprehension: Contextual impenetrability. Journal of Psycholinguistic Research, 18, 223-243. Simpson, G. B., and Krueger, M. A. (1991). Selective access of homograph meanings in sentence context. Journal of Memory and Language, 30, 627-643. Stowe, L. A. (1989). Thematic structures and sentence comprehension. In G. N. Carlson and M. K. Tanenhaus (Eds.)., Linguistic structure in language processing (pp. 319357). Boston, MA: Kluwer Academic Publishers. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 18, 645-659. Swinney, D. A., Ford, M., Frauenfelder, U., and Bresnan, J. (1988). On the temporal course of gap filling and antecedent assignment during sentence comprehension. In B. Grosz, R. Kaplan, M. Macken, and I. Sag (Eds.), Language structure and processing. Stanford, CA: CSLI. Swinney, D. A., and Osterhout, L. (1990). Inference generation during auditory language comprehension. In A. Graesser and G. Bowers (Eds.), The Psychology of Learning and Motivation, 25, 17-33. San Diego: Academic Press. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27, 324-340. Tabossi, P., Colombo, L., and Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tabossi, P., and Zardon, F. (1993). Processing ambiguous words in context. Journal of Memory and Language, 32, 359-372.

Context-Independent Sentence Processing

151

Tanenhaus, M., Boland, J., Garasey, S., and Carlson, G. (1989). Lexical structure in parsing long-distance dependencies. Journal of Psycholinguistic Research, 18, 37-50. Tanenhaus, M. K., Leiman, J. M., and Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18, 427-440. Trueswell, J. C., Tanenhaus, M. K., and Garnsey, S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33, 285-318. Van Petten, C., and Kutas, M. (1987). Ambiguous words in context: An event-related potential analysis of the time-course of meaning activation. Journal of Memory and Language, 26, 188-208.

NOTES 1 There were also additional filler sentences with nonword probes and varied probe positions. 2 Just to be sure that the atypical probes were also being activated at the category word, an additional experiment was conducted using only the atypical contexts and the 0-ms probe position. The results were priming for both typical and atypical probes indicating that both exemplars of the category were activated by the category word. 3 An acoustic analysis showed both greater duration and greater F0 pitch declination for main verbs in the "gap" sentences than in the "no-gap" (later gap) sentences.

This page intentionally left blank

THE INFLUENCE OF CANONICAL WORD ORDER ON STRUCTURAL PROCESSING TRACY E. LOVE DAVID A. SWINNEY Department of Psychology University of California, San Diego La Jolla, California

1. INTRODUCTION Language comprehension is something that humans perform rapidly, effortlessly, and typically without conscious thought. However, underlying this reflexive behavior there exists a multitude of processes at each of a number of different levels of analysis (spanning from basic phoneme discrimination to speech segmentation, to lexical access, to structural analysis, to discourse-level processing, etc.), processes that must be understood in detail if we are to have an accurate model of comprehension. The focus of this chapter is on one of the more subtle and complex areas in this array, structural processing. In particular, we will examine the role of underlying or canonical word order (or, more specifically, grammatical role order) on structural processing—something that is best done from a crosslinguistic perspective. Different languages have putatively different canonical word and grammatical role orders, allowing independent examination of the effects of such underlying facts about language on standard comprehension routines. In examining this issue, we will focus specifically on the structural processes involved in processing long-distance dependencies as found in fillergap constructions, and we will examine the moment-by-moment processes involved in the comprehension of such structures via on-line methodologies. Finally, as theory of experimental methodology is intimately tied to theory of Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

153

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

154

Tracy E. Love and David A. Swinney

mental process, we also examine methodological issues relevant to the on-line examination of structural processing.

2. BASIC ISSUES 2.1. Canonical Order A fundamental issue in the study of language is the distinction between the "underlying" word order (Subject-Verb-Object), hereafter referred to as canonical order, of a language and the many different surface forms of these elements that may appear in a language. Although there is often disagreement about the canonical order for any specific language, in general there is agreement that some languages have fundamentally different canonical orders (e.g., SVO vs. SOV): some of these languages are strongly ordered (again, underlyingly) and others are not as strongly ordered. "Strength" of the role of canonical order is commonly a function of issues such as degree of reliance of word order versus case marking for indication of the grammatical role of words in sentences—in short, the degree to which word "scrambling" is allowable and practiced in language use (see, e.g., Greenberg, 1963; Haegemen, 1991; Taraldsen, 1991; Travis, 1991). Due to this intralanguage difference in canonical order, crosslinguistic experimentation particularly lends itself to answering questions about the fundamental role, if any, of such canonical order on ongoing processing involved in comprehension. 2.2. Structural Processing and Discontinuous Dependencies Discontinuous dependencies are a common structural property of language. They occur when two related elements are separated in the surface form of a sentence. In most languages, discontinuous dependencies exist in many types and forms. In English, one important type involves what are called antecedent-gap or filler-gap dependencies—conditions in which "gaps" in canonical word order are created in the surface form of a sentence by 'movement' of a word to a different position in the sentences (the moved element is termed the "filler" for the gap). In much of what follows we will examine processing of such antecedent-gap relationships as found in the object-relative constructions such as the following: "The police stopped the boy that the couple accused

of the crime.'

English is deemed to be a canonical SVO language. In the object-relative construction in English, however, the object precedes (rather than follows) the verb, which is contrary to this canonical order (the underlying form of this sentence can actually be broken down into two SVO sentences: "The police stopped the

Canonical Word Order

155

boy" and "The couple accused the boy."). In this example, the object of the verb accused is "the boy" which actually appears (has been moved to) a position in front of the verb. Such "movement" from underlying canonical position is often treated in formal linguistic theory as though it has left a phonologically empty "trace" marking the "canonical" position (the "gap") from which it has been "moved" (Chomsky, 1981).1 A long-standing and fundamental problem for models of language processing concerns discovering how the comprehension device links the "moved" components of these discontinuous (and long-distance) dependencies to allow for interpretation, and whether the underlying canonical order of a language plays any role in the structural processes involved in such linkage. That is, when there is a gap found (in violation of an expected element in a canonically specified position) does the processing system have a need to recover the moved word at THAT point in processing? Or, are such linkages only arranged after the entire sentence has been initially processed? In this regard, we note again that English is a Subject-Verb-Object (SVO) ordered language in which the fundamental (or underlying, canonical) order of sentential-conceptual constituents is Subject, followed by Verb, followed by Object. Other languages have different underlying orders, and surface movement may not change the expected configurations in the way they do in English (For example, German (SOV), Hebrew (VSO), Japanese (SOV), all have different canonical orders from English, and also different degrees of adherence to such dominant standard orders). That being the case, one can use on-line examination techniques to determine whether and/or when the displacement of words from their underlying (language-specific) positions has an (on-line) effect on processing. This may allow examination of questions such as whether there is a Universal Canonical Order to grammatical relationships, whether these are languagespecific, and whether such relationships hold more strongly for certain languages than for others. (For example, Bulgarian has a relatively free word-order system, and it may be the case that the canonical word order will not have on-line immediate effects on processing.) 2.3. The Interaction of Research Methodology and Language-Processing Theory 2.3.1. OVERVIEW In the following we present an overview and discussion of the experimental paradigm(s) that we will use in examining the issues of the effect of canonical order on structural processing (and visa versa). Over the years, numerous experimental paradigms have been employed as a means of aiding researchers in understanding the processes involved in language comprehension. These methods can be broadly divided into two groups—off-line

156

Tracy E. Love and David A. Swinney

and on-line approaches. Off-line methods are particularly useful in determining the overall comprehension abilities (or disabilities) of various populations. These methods are typically untimed, and encourage the incorporation of world knowledge in the subjects' response, as in sentence-picture matching tasks, paraphrase tasks, and sentence-recall tasks. In contrast, on-line methodologies are concerned with detailing information as it unfolds during ongoing sentence processing. Such methods attempt to capture moment-by-moment operations of (typically, unconscious) processing that will allow us to understand the details of the comprehension process. A number of on-line methodologies are currently in use to those ends. Although many of these methodologies each have revealed important properties of language processing, cross-modal lexical priming (CMLP; Swinney, Onifer, Prather, and Hirshkowitz, 1979) has proven to be a particularly illuminative and sensitive measure of moment-by-moment sentence processing. CMLP comes in many varieties, but all involve the following conditions and properties. First, the sentential material under study is presented auditorily to subjects, who are told that their major job is to understand the sentence(s) or discourse they hear (subjects are standardly tested for comprehension throughout the experiment—to keep attention to the task of comprehension). Second, subjects are told they have another task to perform: at some point while they are listening to the sentence(s) a visual item will appear on a screen in front of them and they will have to make a decision about that visual item. This visual item may be a letter string (to which subjects may be required to make a lexical decision, or a classifying decision, or a "naming" response) or, the visual item may be a picture (again to which some type of classifying response is made, such as "edible/nonedible"). Work with the CMLP technique has shown that most two-choice classification responses work quite well in obtaining basic effects with this task. There are several aspects of this technique that need to be mentioned: First, presentation of the auditory sentence always continued throughout and beyond presentation of the visual item (and on to the end of the sentence). That is, the sentence is NEVER ENDED with the visual probe; this prevents the probe from being integrated into the ongoing sentential material (provided, of course, that the sentential material is presented normally—see discussion by Nicol, Swinney, Love, and Hald, 1997; Swinney, Nicol, Love, and Hald, 1998). Second, this 'secondary' task never requires the subject to make metalinguistic judgments about the sentential material they hear (such as, "Was this word in the sentence?"). This typically precludes metalinguistic examination of the auditory sentence, examination that involves the engagement of conscious (and hence largely nonautomatic) comprehension processes (see Swinney, Shapiro, and Love, 1998; Shapiro and Swinney, 1998, for more details). Third, at least up to the point of the visual target presentation, processing of the sentence is uninterrupted and "normal." In this regard, the task differs considerably from many other on-line tasks that ask the subject to evaluate each word in a sentence as it appears, or to hold a target in

Canonical Word Order

157

mind while the sentence is being processed. Thus, this task is one of the least intrusive behavioral techniques we have for the on-line examination of the normal comprehension process. A planned relation exists between the tasks the subject performs in CMLP (auditory sentence comprehension and visual target classification). On experimental trials, the visual target is associatively/semantically related to a critical word in the sentence. Following the principle of automatic semantic priming, occurrence of an auditory word (the prime) just prior to processing of another item (the visual target word) that is associatively (and/or semantically) related to that prior item results in speeded processing and classification of the target—a result that is generally known as priming (see, e.g., Meyer, Schvaneveldt, et al., 1975; Neely, 1991). The CMLP task uses the fact that priming occurs between associatively related words to provide an indication of WHEN critical words in the sentence are active during processing. Consider an example in which subjects are presented auditorily with the following object-relative sentence: The policeman saw *1 the boy *2 who the crowd *3 at the party *4 accused *5 of the crime If a visual probe that was related to the noun 'boy' (e.g., the letter string: GIRL) was presented at each of the asterisk-numbered (*) positions (in each instance, to different subjects in different experimental conditions), one might expect, ceritus paribus, the following effects on reaction time to make a lexical decision to that letter string:2 First, no priming effects would be expected at test position *1, as there have been no words related to 'GIRL' heard by the subject up to that point. Second, at test position *2, one might (correctly) expect that reaction time to GIRL would be speeded (primed) due to the subject just having heard 'boy'. Then, at position *3 (and certainly at position *4), one might predict that there would be no more priming effect of the word 'boy' from the sentence on lexical decision times to the letter string 'GIRL,' because sufficient time had passed so that 'boy' would have been fully processed and stored, and no longer active in immediate sentence processing to be able to exert a priming effect.3 On the other hand, if one believes a filler-driven account of linking antecedent fillers to gaps, then, continued activation of the filler ('boy') will occur at points *3 and *4 in the sentence (in that the filler is being kept active while the processor is looking for a gap site to fill). However, if one believes a verb-driven account of linking fillers and gaps (that is, the filler is searched for only once a gap—as marked by a verb requiring a direct object—is discovered), one would predict that reactivation of the filler will only take place at test point *5, when the verb requiring a direct object has been processed and a search of the appropriate antecedent filler is undertaken. In short, via the use of CMLP, we are exploiting the fact of priming to provide a basis for an existence proof about the time course of mental (re)activation of some 'key' word in the sentence—in this case the antecedent filler for a structural gap.

158

Tracy E. Love and David A. S winney

2.3.2. EFFICACY OF THE CMLP PARADIGM IN THE EXAMINATION OF STRUCTURAL PROCESSING Much debate, and rightfully so, centers around the efficacy and sensitivity of the experimental techniques employed to appropriately reflect sentence processing. Relevantly, the CMLP technique has recently been at the center of such a debate, and although it has passed muster on all counts, it is important that the issues raised in this debate be kept in the public eye, as they are relevant to the evaluation of all such on-line techniques. This particular debate has centered on the ability of the CMLP task to appropriately reflect the reactivation of antecedents in filler-gap constructions. This controversy focuses on claims by McKoon and Ratcliff (1994; hereafter M&R) of a potential confound in a filler-gap study by Ford, Frauenfelder, Bresnan and Swinney (see: Swinney, Nicol, Ford, Frauenfelder, Bresnan, 1987, see also Nicol and Swinney, 1989), which had demonstrated automatic reactivation of an antecedent filler at a gap site, only for the structurally correct antecedent. (More on this result will be presented, below, next section). McKoon and Ratcliff (1994) argued that the study was confounded in that the experimental (semantically related) visual target words used in the study constituted a "better fit" with the sentence than the unrelated, control words, and that this fit factor, and not the reactivation of antecedent fillers, constituted the source of priming found at the gap site. M&R then employed a reading task (a variant of Rapid Serial Visual Presentation—RSVP) to demonstrate this confound of fit could have an registerable effect. In their task, sentences containing no gaps were presented visually, one word at a time, and subjects read each word, making a lexical decision to a visual target (which was marked as being the target by being presented in a slightly different location on the screen from the sentential words that preceded it). Note that, except for location (five letter spaces to the right), the visual target was presented as a continuation of the rest of the word-byword visually presented sentence. With this task they found priming for target words they judged to be a better fit with the sentence compared to those that were not. Thus, given that they found priming in their reading task with better-fit targets, they claim that this effect accounted for the original results of Swinney et al. (1987) with the CMLP task. However, this interpretation is highly problematic. First, the task they use to demonstrate that better fit of targets can cause priming is exceedingly different from the CMLP task used by Swinney and colleagues. Most importantly in this regard, the visual sentence continuation task presents the target as a continuation of the sentence; due to this, it is naturally integrated into the sentence (something that does not happen with CMLP); hence it is no surprise that lexical decision times to the task are effected by fit with the sentence—that is, what an integration task measures. CMLP, when used standardly, does not allow such integration; a well-known property of the technique is that targets are essentially never reported as being perceived as 'integrated' into the sentence by subjects, likely because of

Canonical Word Order

159

the different modalities in which the sentence and targets occur. In addition, we note that the reading task used by M&R has very different properties from the auditory comprehension used in CMLP: Reading obviously has many properties that make for important theoretical and empirical differences than those studied in auditory comprehension. As a means of providing a direct test of the claim that CMLP is susceptible to the "fit" or integration of the target into the sentence, as M&R claim, Nicol et al., 1997 (see also, Swinney et al., 1998) undertook a controlled study directly comparing CMLP and RSVP tasks in good fit and bad fit conditions. Essentially, it was a test of whether this potential confound of 'fit' of the targets with the sentence could have caused priming effects in CMLP studies that supported claims of reactivation of antecedent fillers at a gap site. (This experiment also provided a test of whether it mattered whether the control condition constituted matched sentences or matched targets—an issue that McKoon, Ratcliff, and Ward, 1994, had raised as a potential problem in an earlier paper). In this study, subjects were presented with sentences such as, The woman instilled fear in her daughter Apple Agony The woman pushed fear... Apple Agony in both the RSVP and CMLP methodologies. In the examples, the word agony constitutes a better fit after the verb "instilled" but not after the verb 'pushed,' whereas APPLE constitutes a better fit after 'pushed' but not after 'instilled.' The ratings of good or bad fit were gathered by pretesting all material (see Nicol et al., 1997, for specifics). The results of the studies were quite straightforward: The word-by-word visual task (that used by M&R) showed a significant priming effect for target words that constituted a better fit compared to those target words that were a worse fit. This, then directly replicated the findings of M&R—with their reading task. However, no effect at all (not even a trend) of better versus worse fit target words was found in the CMLP paradigm (again, using exactly the same materials and same number of subjects). There was no effect of fit at all intruding into CMLP results. Thus, the potential confound of better fit of experimental versus control target words with the sentence cannot account for priming found at gaps with the CMLP paradigm. CMLP does not lend itself to integrating the probe target into the ongoing sentence that is being studied. This is a critical lesson for on-line examination and on-line methodologies. It is paramount that tasks not be susceptible to integration effects (such as found in word-by-word reading combined with visual-target presentations) unless integration is the factor of interest. In studies of filler-gap processing (activation-reactivation studies) therefore,

160

Tracy E. Love and David A. Swinney

CMLP is a behavioral task of choice (and sentence continuation tasks are not). Furthermore, this examination of M&R's claims found that there was also no difference between use of the matched sentence or matched probe designs—either type of control is equally efficacious. Finally, we note that in work by Love and Swinney (1996), the gap-filling effects that concerned M&R were replicated in a CMLP task that was specifically matched for good or bad fit of targets to the sentence, and thus much of this debate is academic: evidence for antecedent reactivation in gaps exists independent of the good or bad fit of targets with the sentence from CMLP experimentation. The CMLP task is a sensitive measure of specific details of sentence processing and, critically, it is not susceptible to integration of extraneous information into the ongoing sentence. It is, in fact, one of the more sensitive behavioral measures we have of ongoing sentence processing.

3. PROCESSING OF LONG-DISTANCE DEPENDENCIES IN ENGLISH We now turn to an examination of filler-gap dependency processing in English, the language in which the majority of the work has been done so far. (Recall, again, that English is an SVO language.) A multitude of studies (see, e.g., Nagel, Shapiro, and Nawy, 1994; Nicol, 1988; Osterhout and Swinney, 1993; Swinney, Nicol, Ford, Fruenfelder, and Bresnan, 1987; Zurif, Swinney, Prather, and Love, 1994, among others) have demonstrated the following effect in English using the CMLP task (or variant thereof): (a) prior to the verb, there is no evidence of activation of a moved constituent and (b) reactivation is demonstrated at the immediate offset of the verb at the gap. These studies have also demonstrated that this process is unaffected by issues of plausibility and that the process is driven by structural knowledge (other candidate NP antecedents occurring in positions that are structurally precluded from being an antecedent filler for a moved direct object are NOT activated). These results taken together show that in a language that has a strict SVO underlying canonical word order, the processing system is actively attempting to recover the object of the verb on-line at the gap. (We note in passing that similar reactivation effects have been shown for pronouns, and reflexives).4 To give a detailed example of the methodology and interpretation underlying these findings, we turn to a presentation of a recent study in this area: Love and Swinney (1996) investigated whether or not the process of automatic reactivation of filler-antecedents found in gaps in these prior studies involved a search for an antecedent through a deep or superficial representation of the sentence. In order to examine this issue, lexical ambiguities were used as filler antecedents, in order to provide a method of disentangling the 'level of representation' used in a search for an antecedent filler. The reason for use of ambiguities in this study is that all

Canonical Word Order

161

meanings of lexical ambiguities are initially activated when the (surface form of the) word is heard (e.g., Swinney, 1979; Tanenhaus, Leiman, and Seidenberg, 1979). Thus, if all meanings of the antecedent filler are found to be reactivated at the gap site, then one could conclude that the search for an antecedent filler occurs over a surface form (acoustic memory) representation of the sentence; however, if only the contextually appropriate meaning of the antecedent-filler ambiguity is reactivated at a gap, then the search for the antecedent must be over a deeper representation of the sentence—one in which the appropriate interpretation of the ambiguity has been uniquely determined and stored in the structurally appropriate representation for the sentence up to that point. In addition, this study replicates many facets of the prior published results on gap filling just summarized above. Briefly, subjects heard sentences that were strongly biased5 towards one interpretation of a lexical ambiguity such as: Jeff was concerned about Savings and Loan Institutions, so he went to the bank *1 which his family *2 always used *3 and asked about the safety it provided with respect to CD investments. Three separate probe points were tested (*1, *2, *3). The results were very clear: at the offset of the lexical ambiguity (* 1), there was evidence for exhaustive access (i.e., there was priming for both the 'money' and the 'river' meaning of bank). As stated earlier, this replicates the well-reported finding that lexical access occurs automatically and is not guided by the context of the sentence (among others, Swinney, 1979; Tanenhaus, Leiman, and Seidenberg, 1979). Next, at the baseline probe point (*2) there was no activation of either meaning evident. Finally, and most importantly, at the gap (*3), there was evidence for the reactivation of the antecedent—but only for the contextually relevant meaning ('money'). Moreover, a significant interaction in the level of activation between probe points 2 and 3 (for the contextually relevant meaning only) demonstrated there was reactivation of the antecedent at the gap (i.e., the system was actively recovering and reactivating the filler in the underlying canonical SVO order position). Furthermore, this reactivation involved a search through an underlying or deep memorial representation of the antecedent (since only the contextually relevant meaning was reactivated). As stated earlier, this finding of an automatic linking of a gap to its structurally defined antecedent has been shown in many other experiments using CMLP (Nagel, Shapiro, and Nawy, 1994; Nicol, 1988; Osterhout and Swinney, 1993; Zurif, Swinney, Prather, and Love, 1994, among others) and across other methodologies.6 Thus, there is considerable evidence that, at least in English, the comprehension device prefers to have the direct object activated immediately following the verb during ongoing processing. This fits with a view of comprehension as driven by the need to actively recover the canonical SVO order online during comprehension.

162

Tracy E. Love and David A. Swinney

Of direct interest to both processing and Universal Grammar models is whether this pattern of structurally driven reactivation as a means of recovering the canonical word order is language specific and/or tied to canonical orders found in each language. We can explore these issues by studying whether or not these findings are evident and predictable in languages possessing different fundamental word orders than English. What follows is a brief review of the only currently existing on-line evidence from two such languages: Bulgarian and Spanish, each of which brings a different perspective on canonical order. Unlike English, which possesses a strict adherence to its underlying word order, Bulgarian allows for extensive scrambling (i.e., it has a more relaxed word-order system; Stamenov and Andonova, this volume). Spanish, on the other hand, finds itself somewhere in between English and Bulgarian in degree of allowable scrambling (strict canonical order), but there is considerable debate in the literature as to whether its underlyingly canonical word order is VOS or SVO (see, e.g., Basilico, Pinar, and AntonMendez, 1995). These languages bring a unique perspective to this field of investigation and can allow us to explore these issues of underlying canonical word order driving structural reactivation.

4. A CROSSLINGUISTIC PERSPECTIVE Stamenov and Andonova (this volume) investigated the time course of antecedent reactivation via a CMLP task in Bulgarian, which, as mentioned earlier, has a much more relaxed word order than English. They studied object-relative constructions such as: Za obyad v restoranta predlagaha teleshki drob, koyto Stefan mnogo obichashe ot maluk. 'For lunch, in the restaurant (they) offered veal, which Stefan loved very much since his childhood.' They tested at multiple sites in the sentence (at the offset of the lexical ambiguity, a baseline probe point and at the gap) and found—quite unlike the findings for English—no evidence of reactivation of the antecedent drob at the gap. The authors entertain the idea that they can attribute this finding, in part, to the fact that Bulgarian allows for a more lax word-order representation in object-relative constructions (i.e., Bulgarian is considered to have a relatively free word-order system). Given that this language has a lax canonical word order, there may be no expectation built up in the system to find specific grammatical role objects in structurally defined positions (on-line), and hence no reactivation of such items in any particular position takes place in a first-pass analysis of the sentence.7 However, in a similar priming study in Spanish, Basilico, Pinar, and AntonMendez (1995) examined for reactivation priming for a verb in V-S-O and V-O-S Spanish constructions, at a point between the subject and object in those sentences

Canonical Word Order

163

(recall that Spanish is argued to be either an SVO or a VOS language, underlyingly). They reasoned that if the underlying canonical order is really SVO, then one should find reactivation of the verb between the subject and object in the VSO constructions only (as compared to the VOS construction, where it might be expected after the subject). And this is precisely what they report finding, lending credence to the belief that the underlying word order in Spanish might be S-V-O, and that the underlying word order is strict enough for the comprehension device to expect and require grammatical objects in their canonical position.

5. CONCLUSIONS AND DISCUSSION Although there are as yet few pieces of evidence from crosslinguistic work to contrast with the extensive English findings, we can see that work from two languages with far less stringent word orders than English presents us with mixed results. In one, Spanish, evidence suggests that even though one is allowed a freer surface word order, it appears that the comprehension device is attempting to actively recover conceptual information in an underlying SVO order during ongoing comprehension (as is found in English; we note that similar evidence has been hinted at in early work in German; Clahsen, personal communication, March 1997). However, in Bulgarian, no such evidence of the processor expecting or utilizing canonical word order is found. This fits with some general beliefs that strictness of word orders may be a continuum. Some languages are truly freer— even at an underlying level. And, this may well percolate to the workings of the comprehension device, as seen in these data. Clearly, in order to examine this hypothesis further we need many more studies that investigate languages with a variety of canonical word orders that fit all along this continuum of a lax to strict adherence to word order. Now, however, we have techniques sensitive enough to allow researchers to temporally map out such delicate sentence processes as they are occurring, and we can use crosslinguistic phenomena such as canonical word order to assist the understanding of processes by which long-distance dependencies are linked. Furthermore, we can investigate whether or not this linkage is language specific or in fact performed in a universal fashion, across similar language structures in all languages, thereby adding to our evidence concerning questions of universality of structural processing. As already seen, the influence of canonical word order on structural processing may vary (e.g., the Bulgarian evidence). However, even this evidence is only preliminary. A major goal of this work is to determine if there are natural categories or groupings into which languages may fall with regard to structural processing issues (which may, in turn, effect linguistic theory at some level), thus, only further crosslinguistic examination of such processes, with sensitive on-line techniques, will bring us closer to answers on these issues.

164

Tracy E. Love and David A. Swinney

ACKNOWLEDGMENTS The authors gratefully acknowledge the support of a grant from the National Institutes of Health (DC00094) for the writing of this chapter and for much of the research reported herein.

REFERENCES Basilico, D., Pinar, P., and Anton-Menedez, I. (1995, March). Canonical word order and the processing of verbal traces in Spanish. The Proceedings of the Eighth Annual CUNY Conference on Human Sentence Processing, Tucson, Arizona. Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris. Garnsey, S., Tanenhaus, M., and Chapman, R. (1989). Evoked potentials and the study of sentence comprehension. Journal of Psycholinguistic Research, 18 (nl), 51-60. Greenberg, J. (1963). Some universals of grammar with particular reference to the order of meaningful elements. In J. Greenberg (Ed.), Universals of language (pp. 73-113). Cambridge, MA: MIT Press. Haegemen, L. (1991). Introduction to government and binding theory. Cambridge, UK: Blackwell. Love, T., Nicol, 1, Swinney, D., and Nafie, K. (1997). The processing of explicit anaphors . in Broca 's and Wernicke 's aphasics. Center for Human Information Processing Technical Report, #97-2. Love, T., and Swinney, D. (1996). Co-reference processing and levels of analysis in objectrelative constructions: Demonstration of antecedent reactivation with the cross modal priming paradigm. Journal of Psycholinguistic Research, 25 (1), 5-24. McKoon, G., and Ratcliff, R. (1994). Sentential context and on-line lexical decision tasks. Journal of Experimental Psychology: Language, Memory and Cognition, 20(5), 1239-1243. McKoon, G., Ratcliff, R., and Ward, G. (1994). Testing theories of language processing: an empirical investigation of the on-line lexical decision task. Journal of Experimental Psychology: Language, Memory and Cognition, 20(5), 1219-1228. Meyer, D. E., Schvaneveldt, R. W. et al. (1975). Loci of contextual effects on visual word recognition. In. V. P. M. A. Rabbit and S. Domic (Eds.), Attention and performance New York: Academic Press. Nagel, H. N., Shapiro, L. P., and Nawy, R. (1994). Prosody and the processing of filler-gap sentences. Journal of Psycholinguistic Research, 23 (6), 473-486. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In G. W. H. D. Bresner (Ed.), Basic processes in reading: Visual Word Recognition. Hillsdale, NJ: Lawrence Erlbaum Associates. Nicol, J. (1988). Co-reference processing during sentence comprehension. Unpublished doctoral dissertation, M.I.T., Cambridge, MA. Nicol, J., and Swinney, D. (1989). The role of structure & coreference assignment during sentence comprehension. Journal of Psycholinguistic Research, 18 (nl), 5-19.

Canonical Word Order

165

Nicol, J., D. Swinney, Love, T., and Hald, L. (1997). Examination of sentence processing with continuous vs. interrupted presentation paradigms. Center for Human Information Processing, Technical report #97-3, La Jolla, U.C.S.D. Osterhout, L., and Swinney, D. (1993). On the temporal course of gap-filling during comprehension of verbal passives. Journal of Psycholinguistic Research, 22,(2), 273-286. Shapiro, L., and Swinney, D. (1998). On-line examination of language performance in normal and neurologically-impaired adults. American Journal of Speech-Language Pathology, 7, 49-60. Stamenov, M., and Andonova, A. (1997). Lexical access and co-reference processing in Bulgarian. In D. Hillert (Ed.), Cross linguistic studies of language processing. San Diego: Academic Press. Swinney, D., Nicol, J., Ford, M., Fruenfelder, U., and Bresnan, J. (1987, November). The time course of co-indexation during sentence comprehension. Paper presented at the Psychonomic Society Meeting, Seattle, WA. Swinney, D. (1979). "Lexical access during sentence comprehension: (Re) consideration of context effects." Journal of Verbal Learning and Verbal Behavior 18, 645-659. Swinney, D., J. Nicol, Love, T., and Hald, L. (1998). Methodological issues in the on-line study of language processing. In Schwartz (Ed.) Childhood language disorders. Hillsdale, NJ: Erlbaum. Swinney, D., Onifer, W., et al. (1979). Semantic facilitation across sensory modalities in the processing of individual words and sentences. Memory and Cognition 7(3), 159-165. Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27, 324 -340. Tanenhaus, M. K., Leiman, J. M., and Seidenberg, M. S. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18,427-440. Taraldsen, K. (1991). A directionality parameter for subject-object linking. In Freiden (Ed.), Principles and parameters in comparative grammar (pp. 219-268). Cambridge, MA: MIT Press. Travis, L. (1991). Parameters of phrase structure and verb second phenomena. In Freiden (Ed.), Principles and parameters in comparative grammar (pp. 339-364). Cambridge, MA: MIT Press. Zurif, E. Swinney, D. Prather, P., and Love, T. (1994). Functional localization in the brain with respect to syntactic processing. Journal of Psycholinguistic Research, 23, (6), 487-498.

NOTES 1 While this chapter is concerned only with providing a processing account of aspects of language (and, hence, we do not intend or attempt to support one linguistic account of these constructions over another), we have adopted much of the terminology of the government

166

Tracy E. Love and David A. Swinney

and binding theoretic herein because it descriptively captures the phenomena we examine, and empirical data we provide, more coherently than most other accounts. We note, however, that that this does not at all mean it will ultimately represent a better universal grammar account of these language phenomena. 2 Note that all effects are evaluated in comparison to lexical decision reaction time to a control letter string presented at each of these test points; a control letter string is a word that is associatively/semantically unrelated to the key word in the sentence, but which is matched to the 'experimental' (related) letter string on the basis of a priori reaction time (lexical decisions taken on the words presented in isolation). 3 The priming that is standardly found to classification of a visual target immediately following occurrence of a semantically or associatively related word in an auditory sentence typically lasts between 100-700 msec, ceterus paribus. 4 In a study by Nicol (1988), subjects were presented with sentences containing three unrelated noun phrases (NPs) such as: NP1

NP2

NP3

The boxer said that the skier thought that the doctor from the team had blamed HIM/ HIMSELF for the recent injury. At the offset of the pronoun or reflexive, a word either related to the 1st NP (fighter), 2nd NP (snow), or 3rd NP (nurse), or matched control words were presented. In the case above, the structurally correct antecedent for the reflexive HIMSELF is the doctor, whereas the structurally correct antecedent for the pronoun HIM cannot be doctor. The results showed that the structurally defined antecedent was in fact activated after the overt anaphors (skier for HIM and doctor for HIMSELF). These findings have recently been replicated (Love, Nicol, Swinney, and Nafie, 1997) and support the view that antecedent re-activation is automatic and structurally driven. 5 Of the many pretests run on these materials, three separate pretests were run to ensure (a) Tabossi (1988) criteria were met; (b) a strong biasing context; and (c) there were no goodness-of-fit confounds. (Please see Swinney et al., 1998, for details.) 6 One such example is by Garnsey, Tanenhaus, and Chapman (1989) where subjects were presented with the following sentences: a. b.

The businessman knew which customer the secretary called at home. The businessman knew which article the secretary called at home.

They obtained electrophysiological measures (evoked response potentials) during comprehension of these sentences and found a larger N400 at the verb ('called') for (b) than for (a). N400 potentials are believed to be, among other things, a response to semantic incongruity. If a large N400 was found in (b) at the verb, it likely is because the incongruent filler (the word article) was reactivated at the gap following the verb. Thus, it seems likely, based on these data that the object filler was activated at the site of the gap during processing (and, such activation was independent of plausibility constraints). 7 An interesting aside is that although they did not find reactivation at the gap, they did replicate the finding of contextually independent exhaustive access of lexical ambiguities when testing at the offset of the lexical ambiguity.

LEXICAL ACCESS AND COREFERENCE PROCESSING IN BULGARIAN MAXIM I. STAMENOV* ELENA ANDONOVA * Institute of the Bulgarian Language Bulgarian Academy of Sciences Sofia, Bulgaria Cognitive Science Department New Bulgarian University Sofia, Bulgaria

1. INTRODUCTION The aim of the present study is to use the Cross-Modal Lexical Priming task (CMLP) for providing detailed evidence about the time course of antecedent reactivation during sentence processing in Bulgarian as compared crosslinguistically with the available data for the English language. This chapter is based on the methodology developed and applied to English language material by David Swinney and his co-workers (cf. Love and Swinney, 1996). The aim of the paper is to reveal the nature of the representation that is examined when a referenceseeking element is linked to its antecedent during processing of complex sentences with object-relative clause constructions. In our study, following the procedure of Love and Swinney (1996), subjects heard Bulgarian sentences that contained a lexical ambiguity placed in a strongly biasing context. This ambiguous word was the "moved" or "fronted" object of the verb in an object-relative construction. A CMLP naming task was applied in order to determine whether one or more meanings of the polysemous ambiguity are activated at three temporally distinct points during the sentence realization: Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

167

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

168

Maxim I. Stamenov and Elena Andonova

(a) immediately after the lexical ambiguity (Position 1); (b) a later point that was 700 ms prior to the offset of the verb (Position 2); and (c) immediately after this verb at the gap in this filler-gap construction (Position 3). Our data provide partial confirmation of results achieved in studies with English language object-relative constructions. At Position 1, all meanings of the ambiguity were activated compared with controls. At Position 2, both meanings were still activated. At Position 3 none of the meanings were activated (unlike the results of Love and Swinney, where only the context-relevant meaning of the ambiguity was reactivated). Interpretation of the difference in the experimental results in crosslinguistic analysis of English and Bulgarian language data is offered. One possibility is related to the specificity of the relatively free word order in object-relative clauses in Bulgarian (unlike English); another possibility is the influence of gendernumber agreement between the relative pronoun and the object noun. A third (and most generally formulated) possibility would be related to language-specific interface(s) between structural, semantic (thematic role), and word-order constraints. Studies with CMLP imposing further restrictions and variation in the test materials in order to control language-specific components will provide more detailed evidence for the time course of processing object-relative clause constructions.

2. BACKGROUND In the literature on sentence processing (cf. Tanenhaus and Trues well, 1995; Hillert, 1997) there are currently three main standpoints. The first and most radical choice maintains that syntactic parsing is a separate and autonomous component in sentence processing. Syntactic processing is modular and serial (possibly involving two separate stages of early and late syntactic parsing; cf. e.g., Friederici and Mecklinger, 1996; Urbach, 1997). The second approach is lexicalist (i.e., it views syntactic parsing, including ambiguity resolution, as a "lexical process"; MacDonald, Pearlmutter, and Seidenberg, 1994:700). The third view attempts to find a compromise between the two alternatives of lexical versus syntactic specificity of sentence processing. The two of them are seen as both autonomous (in some respects) and interacting (i.e., capable of sharing information on-line) (cf. Boland and Cutler, 1996). At present, the overwhelming majority of experiments in the study of sentence processing have been made by using the English language. Thus we may expect to find that at least some of the specificities of sentence processing uncovered and reported in the available literature are not universal properties of sentence processing but ones specific to English. From this point of view, the crosslinguistic study of the way of processing of sentence structure can make a contribution to resolving the issue about the nature of the rules/constraints involved in lexicalist

Bulgarian: Lexical Access and Processing

169

versus syntax-driven sentence-parsing models. Especially intriguing and challenging for both approaches would be data displaying language-specific properties of lexical versus syntactic aspects of sentence processing (i.e., that one and the same function is performed in different languages by lexical vs. syntactic components contributing to sentence comprehension).

3. SOME SPECIFIC FEATURES OF THE BULGARIAN LANGUAGE Some of the peculiarities of the Bulgarian language are related to the fact that, genetically, it belongs to the group of Slavonic languages but from a typological point of view it is closer to analytic languages with its lack of explicit case markers. Instead, syntactic relationships within a sentence are expressed predominantly by means of prepositional constructions and word order, among others. Unlike English, in Bulgarian there is an elaborate verbal conjugation system with inflectional endings that are almost unambiguous markers for person, number, and tense of the verb. Verb forms agree with the subject in person and number and, in some past tense forms, even in gender. Similarly to other languages with a complex system of verbal inflections (e.g., Italian or Spanish), Bulgarian permits subject ellipsis (pro-drop). The range of permissible word orders for the transitive construction in the Bulgarian sentence covers all logical possibilities. In addition to features of the synthetic and analytic language types, it also has some agglutinative characteristics typical of languages belonging to the Balkan Sprachbund (e.g., the postpositional agglutinative definite article). These and other features of Bulgarian make it an important testing ground for linguistic performance models originally related to typologically different languages. Data on it can reveal different language-specific patterns of distribution and coding of lexical, syntactic, and semantic information. Having in mind these peculiarities of the Bulgarian language, it is possible to expect a close replication of results based on data from the English language as far as there exists syntactic parallelism in the sentence structure and there is no additional interference in Bulgarian from other factors contributing to sentence comprehension on-line.

4. THE PROBLEM OF GAP FILLING AND CROSS-MODAL LEXICAL PRIMING The problem of gap filling entered psycholinguistics in the context of the study of syntactic parsing as a modular stage and against the background of the Government and Binding (GB) approach of generative grammar (cf. Chomsky, 1981). In

170

Maxim I. Stamenov and Elena Andonova

GB framework, empty categories are treated as a type of referential dependency— a type of unexpressed (implicit) pronoun with a filler as its antecedent. These filler-gap dependencies are found in object-relative constructions of the type There is the coachi the boy saw i where coach is the filler, is the structural position gap, and the coreferencing is marked by the subscript i. The function of coreferencing and structural gap filling is to reconstruct the canonical ordering of the subject-verb-object (SVO) order in the object-relative construction. The reconstruction is posited as a necessary step in the comprehension of the corresponding complex sentence. Love and Swinney (1996) claim that sentence interpretation requires (among other things) the recovery of the underlying syntactic relationships represented by the canonical word order of the component clauses of a complex sentence.1 The procedure required is to get access to an underlying deep (D-) structure canonical representation and find the structural gap in it to allocate as soon as possible its antecedent in the main clause selected among a closed set of alternatives. The term D-structure has accumulated in the last 30 years quite a complicated set of usages. The differentiation between surface syntax and its "deep" structural correlates has been maintained, in one way or another, since the inauguration of the Standard Theory in generative grammar. In this context, D-structure is considered to be an autonomous and specifically syntactic module of language processing consisting of phrase-structure rules and lexical insertion. Surface (S-) structure is an outcome of transformations operating upon D-structure that form derived phrase markers (Chomsky, 1965). The Extended Standard Theory led to a four-level conception of the language module, including Logical Form (LF), D-structure, S-structure, and Phonetic Form (PF). The Principles and Parameters approach discarded the standard treatment of transformations and introduced instead movement transformations. In the Minimalist Program, the S-structure turned into the sole point of interaction among the three fundamental levels of LF, PF, and D-structure (Chomsky, 1995). While D-structure contains information about the syntactic structure derivation and optimizes the generation of the S-structure on-line, LF is supposed to constitute an interface between overt syntax and conceptual-intentional properties of language. The point of this short excursion into the story of the term of deep structure is the following. In the context of sentence processing there are cases where for its proper performance the mind requires access to information mentioned previously, (e.g., in anaphora resolution or coreferencing). The problem is what type of mental representation is consulted in order to find the appropriate information. Theoretically, there are at least two possibilities, in this respect. The parser gets access either to the surface structure of the previous clause(s) or to some deep structure representational format. In the first case, the inspection of the surface syntactic structure for the sake of finding an appropriate coreferent for gap filling would require automatic access to all of the meanings of an ambiguous word (i.e.,

Bulgarian: Lexical Access and Processing

171

a repetition of the access as enacted in the previous clause[s]). In the second case, the filler is "saved" in a different format during the on-line processing of coreference after its initial occurrence. In this format only the contextually dependent meaning of an ambiguity will be available at the site where the gap is licensed by the verb. Still, the claim about access to some deep structure format supporting coreferencing resolution is not without its own problems (providing we accept its existence). Here we also have three possibilities, in principle. The first is that the reference-seeking element triggers a search through a representational format of a (strictly GB) syntactic D-structure type of the sentence, resulting in the reactivation of potential referents. The second would maintain that the representational format accessed is "deeper" (more central) compared to surface structure, but this format is based on semantic principles (e.g., theta roles grid), as Tanenhaus and Trueswell (1995) find possible to maintain. The third possibility would be once again in compliance with generative grammar following the latest Minimalist Program (cf. e.g., Chomsky 1995:132) and maintaining that antecedent reactivation might require not just access to D-structure, but also to Logical Form (LF) or to a coordinated interfacing of S-, D-structure, and LF in on-line processing in looking for a "common solution." In any case, Love and Swinney (1996) explicitly claim that with the use of CMLP they test the availability of structural dependencies as found in objectrelative constructions. This is accomplished by employing ambiguous words as the antecedent in these constructions. It is hypothesized that if the representation that is examined during coreference processing "is a deep or conceptual one, then only the single context-relevant meaning of an ambiguity [ . . . ] will be reactivated at the structural gap following the verb" (p. 8). This is based on the well-established effect that all meanings of an ambiguity become activated when it is first encountered during on-line processing (cf. Swinney, 1979). In English the gap appearance for the direct object of transitive verbs in objectrelative constructions is fixed immediately after the verb. If the appearance of an object is licensed by the verb and an object-like NP does not appear in the surface structure at that point, this automatically triggers the procedure of "looking for an antecedent." In such cases, one can try to verify the exact time of the triggering by the technique of CMLP, which is rather sensitive to time course factors of on-line sentence processing (cf. Swinney et al., 1987). CMLP is considered to be one of the best developed and validated experimental techniques in psycholinguistics (along with ERP studies) for testing some of the specific features of the time-locked (serial) course of sentence processing. It is maintained that with this technique for study of sentence processing we can investigate what types of processes and representations must be available to the parser on-line for its proper performance. With CMLP, we can test the hypothesis that the parser has access to some format or representation of the structure of the

172

Maxim I. Stamenov and Elena Andonova

sentence that is not identical to its surface structure (available from the echoic or iconic memory, i.e., from the PF). From the PF one can extract directly only S-structure and access the lexically specific representations of single words. By checking access at the offset of the filler, at the gap, and in some position in-between the two, we aim at establishing experimentally the type of representation that is maintained in the working memory for the sake of binding the filler and the gap.

4.1. Method 4.1.1. PREFACE The cross-modal lexical priming (CMLP) task has only been used in the last 10 years (cf. Swinney et al., 1987), predominantly with English language material. This study is one of the first on-line experimental designs used with Bulgarian language material and definitely the first one to employ CMLP. In this respect, it confirms the applicability of the technique in research on Bulgarian. For our study, we chose to use the matched-probes configuration (here a matched probe is understood as the visually presented associate to the antecedent noun phrase). We used a delicate on-line technique, and it was necessary to control for the variability of stimuli processing within the subject group. As previous research has shown, it is important for the related probes to be strongly associated with the corresponding ambiguous noun, and it is necessary to ensure that the control probes have no associative relationship to the sentence. This was accomplished by means of pretests and appropriate controls on subject population.

4.1.2. SUBJECTS Subjects were drawn from a homogeneous population, all native speakers of Bulgarian, with no learning or other disabilities, of a relatively narrow age group (16-26). Homogeneity was ensured through a screening procedure employed for each of the pretests and main experimental conditions described below. Subjects were high school students from Sofia and undergraduate university students in New Bulgarian University, Sofia, who participated voluntarily and were paid a small amount of money (200 Iv) after passing through the screening procedure and participating.

4.1.3. PRETEST FOR AMBIGUITY BIAS One hundred ninety-four lexically ambiguous nouns with a distinct difference between first and second meaning were initially listed by the experimenters on the

Bulgarian: Lexical Access and Processing

173

basis of a dictionary search. Upon examination, some ambiguities revealed a feature that is strongly related to the specific nature of grammatical markers in Bulgarian. Gender marking may vary and serve as an additional disambiguation cue, differentiating between the two senses of the same word. Therefore, we eliminated the possible influence of this cue by screening nouns for same gender of both senses. Stimuli were presented auditorily, and 30 subjects were asked to provide the first word they thought of. After this, they were asked to provide an associate to another meaning of the word. After tallying numbers of first and second choices for each meaning of the nouns, lexical ambiguities were selected with clear preferences for a primary interpretation. The primary meaning had to have been rated in first place (rather than second, etc.) 75% of the time; additionally, the second meaning had to have been rated in second place at least 75% of the time. As a result, the final list of lexical ambiguities included 30 items. 4.1.4. PRETEST—A PRIORI EQUATED REACTION TIMES FOR RELATED AND CONTROL PROBES One of the three most frequent associates for each meaning from the pretest for ambiguity bias was then used in the lists of probes for the primary and secondary meanings of the ambiguity. In an isolated lexical-naming task words were included which were equated to the related associates on the basis of length, form class, and animacy, as well as frequency. Frequency measures were obtained by subjective estimations of judges due to the lack of a reliable general-purpose frequency dictionary for Bulgarian (the available word-frequency counts are for specific purposes only). This naming task was run on another group of subjects from the same population, a total of 30 students. Thus, matched control words were selected for the related associates for the two meanings of the ambiguity. Overall, the mean reaction time for the related probes was 553 ms and for the matched control probes was 560 ms. The difference was not significant [t = 0.229, p = 0.96]. 4.1.5. CREATION AND PRETEST OF OBJECT-RELATIVE CONSTRUCTION SENTENTIAL MATERIALS Each of the 30 ambiguous words was placed in a sentence that is strongly biased toward the word's primary, dominant meaning. This was achieved by the use of specific contextual bias in the initial part of the sentence preceding the occurrence of the ambiguity. Sentences were constructed in such a way that the ambiguous word played the part of the filler for the gap (where the gap is supposed to be an implicit structural entity created by movement of the direct object from its canonical position following the verb). Biasing contexts were created by using lexical items strongly related semantically to one of the ambiguity's meanings at

174

Maxim I. Stamenov and Elena Andonova

the expense of the other, secondary meaning. These items were placed prior to the ambiguity. For the sake of clarity, let us give an example of the sentential materials used here: (1) Za obyad v restoranta predlagaha teleshki drobi, koyto Stefan mnogo2 obichashe3 ot malak. For lunch, in the restaurant (they) offered veal [drob] which Stefan loved very much ever since his childhood. The ambiguous word in this sentence is drob whose primary meaning is 'liver' and whose secondary meaning is '(mathematical) fraction.' It is important to note here that some of the peculiarities of the Bulgarian language become obvious even in this example. Thus, because a verb inflectional system is in place and subjectverb agreement in number and gender is mandatory, it is not uncommon for the subject to be dropped, if it is implicitly understood. Another point to be made here about Bulgarian is the fact that word-order restrictions are very lax and few; thus the subject often comes after, and not prior to, the verb. Thirdly, unlike English, the structure of the Bulgarian complex sentence of this kind requires the explicit appearance of one of the relative pronouns koyto, koyato, koeto, koito 'which,' which are marked for gender and number. The Bulgarian relative pronoun occurring mandatorily after the ambiguous word is thus in agreement with the gap filler NP. For example, in (1) koyto 'which-MASC-SING' agrees in both gender and number with the preceding teleshki drob. In this particular case, the subject of the subordinate clause, Stefan, also happens to be marked for the masculine gender (and for singular number) as a proper noun. However, proper nouns of a different gender could appear in the same position, and thus a discrepancy between the gender markers of the subject NP and object NP would emerge. In such cases, a higher degree of disambiguation would result, and this may come as soon as the discrepancy becomes obvious (i.e., much earlier than the gap position and even before the onset of the verb). During the auditory presentation of the sentential materials there were three positions at which the visual probe words occurred. Probe Position 1 followed the ambiguity immediately (i.e., it was at the offset of the ambiguity). Probe Position 2 occurred after Position 1 and 700 msec before the offset of the verb in the object-relative construction (this point is placed at its approximate position in the example above). Probe Position 3 was at the offset of the verb in the objectrelative construction where the gap is postulated. In both studies discussed below, all three positions remained the same for all sentences. The probes used with example (1) (in both studies) were as follows. The sentence contains the ambiguous noun drob with a primary meaning 'liver' and a secondary meaning '(mathematical) fraction': Primary meaning: related = meso 'meat' Primary meaning: control = darvo 'tree'

Bulgarian: Lexical Access and Processing

175

Secondary meaning: related = matematika 'mathematics' Secondary meaning: control = poliklinika 'polyclinic'

4.1.6. MATERIALS AND DESIGN FOR THE EXPERIMENT The 30 sentential materials were used in combination with the same number of filler sentences that were designed to be of approximately the same structure. The filler sentences were matched with visual word probes that were unrelated in meaning to anything in the sentence. These filler probes were approximately identical in length and the same in form class with the probes in the experimental materials. In addition, five practice sentences matched with practice probes were created with the same structure and constraints as the fillers. The 30 experimental sentences and the 30 filler sentences together with 5 practice sentences that appeared in the beginning of the script were recorded by a male speaker of Bulgarian, at a normal speaking rate; the recording was subsequently entered digitally into a specialized software program for psycholinguistic studies, PsyScope. The 60 sentences were assigned random positions in a script such that no more than two fillers or two experimental sentences appeared in a row. Probe positions were assigned to the sentences. In order to avoid subject bias due to repetition of sentential materials, the whole experiment was divided into two studies. Study 1 was conducted in order to verify the occurrence of an effect that has been established very well for English but that has not been attempted in Bulgarian until now. The effect constitutes the activation of more than one interpretation of a lexical ambiguity even if it follows the occurrence of a strongly biasing context. The activation of two interpretations, or of access to both meanings of an ambiguous word is measured by means of lexical priming at the point of the offset of the ambiguity in a contextually biasing sentence. Thus subjects heard the script with 65 sentences and the experimental and control probes for both primary and secondary meanings of the ambiguity were presented visually at the point immediately following that word. This is Position 1 in our experiment. There were four visual probe conditions in Study 1 concerning Position 1. The four conditions were created by assigning the four probe types, as illustrated earlier, to four separate lists so that all four types were equally represented across the four lists. In this way, we created four experimental subjects' groups conditions, corresponding to the four lists, and 12 subjects were run in each group condition. Study 2 consisted of the 65 object-relative construction sentences as described in the section on the creation and pretest of object-relative construction sentence materials above. There were two positions, Probe Position 2 and Probe Position 3, included in this study. Two almost identical versions of the same script were created with the same sentential materials combined with four visual probe list

176

Maxim I. Stamenov and Elena Andonova

conditions. The two versions were identical with the only exception that the occurrence of the two positions (Position 2 and Position 3) was counterbalanced across the two versions. In this way, Study 2 contained eight between-subjects groups, created by the 2 Versions X 4 list conditions. As pointed out by Love and Swinney (1996), this design allows each subject to experience all of the experimental conditions (probe types and probe positions), distributed across different items. There were eleven subjects in each of the eight experimental groups in this second study, accounting for a total of 88 subjects. 4.1.7. PROCEDURES FOR STUDIES 1 AND 2 In both studies subjects experienced the same procedure. They were comfortably seated in front of a computer monitor and were told that they had two tasks to perform at the same time. The first task was to listen to sentences that they would hear over the headphones and try to understand them. They were also told that the experiment would be interrupted several times and they would be tested on the content of the sentence they had heard. Their second task was to observe the computer screen, and they were asked to name as rapidly as possible the words that appeared on the screen with each sentence. The auditory sentence continued on without any interruption after the visual probe was presented. A black mask was used to focus the attention of subjects on the middle of the screen where words appeared. The subject's voiced response was recorded via microphone into the computer. Reaction time was measured for the onset of response after a visual probe had been presented. There were four interruptions of the experiment for each subject after a filler sentence and prior to a filler sentence each time. During these pauses subjects were asked to answer a multiple-choice question testing comprehension of the content of the sentence they had just heard. 4.2. Results In order to guarantee the validity and comparability of results and to eliminate errors in performance, we carried out a preliminary treatment of data in the following way. Cases of failure to respond and failure of the computer to register a response were discarded from the data. Individual subjects' data were screened for measures lying three standard deviations above or below his or her mean reaction time. When subjects' data contained more than 10% errors, the data were not included in subsequent analysis. In addition, we eliminated the data of subjects who gave an inadequate response to more than one of the questions in the comprehension test. The analyses will be based on the available data that remained after deletion of errors and individual deviations. The data were included in

Bulgarian: Lexical Access and Processing

177

TABLE 1 STUDY 1: MEAN REACTION TIMES (IN MS) FOR RELATED AND CONTROL PROBE NAMING FOR PRIMARY AND SECONDARY MEANINGS OF THE AMBIGUITY AT POSITION 1a Meaning of ambiguity Primary Secondary

Probe type

Probe position 1

Related Control Related Control

672* 729 687* 723

"following offset of ambiguity. * significant difference for related-control probe contrast.

planned comparisons and overall analyses of variance (ANOVA), as described below. In total, we analyzed the data from 1654 cases for study 1 and 2676 cases for Study 2. Study 1 investigated the activation of both meanings of the ambiguous word immediately following its occurrence in the sentence. Such activation has been demonstrated for English but had not been attempted for Bulgarian until the present study. Table 1 provides the means for each of the experimental conditions as calculated across all subjects. The results demonstrate priming for naming reaction times to the related probes as compared with the control probes for both the primary and secondary meanings of the ambiguity. There was a main effect for probe type (related vs. control) [F = 7.839, p < 0.01]. A planned comparison was run for related versus control probes [t = 4.46, p < 0.01]. There was no significant interaction with the ambiguity meaning factor (i.e., whether it was primary or secondary meaning that was activated). In this respect, the study provided replication and confirmation of results of previous similar studies demonstrating again that both meanings of the ambiguity are immediately activated following the occurrence of the ambiguous word in the sentence, even if the sentence was so constructed as to have a strong biasing context preceding the word. In this respect, the processing mechanisms in Bulgarian appear to be analogous to those established for English. Study 2 examined activation of the different meanings of the ambiguity at two subsequent points in the processing of the sentence. Probe Position 2 was a point 700 msec prior to the offset of the verb that licenses the gap, which in practice means before the verb is processed. Probe Position 3 was a point immediately following the verb (i.e. at the position of the gap). As with the data from Study 1, means for each of the experimental conditions (related vs. control for both the primary and secondary meanings of the ambiguity) at both positions 2 and 3 were calculated across all subjects. Table 2 presents data on Probe Position 2 and Probe Position 3.

178

Maxim I. Stamenov and Elena Andonova TABLE 2 STUDY 2: MEAN REACTION TIMES (IN MS) FOR RELATED AND CONTROL PROBE NAMING FOR PRIMARY AND SECONDARY MEANINGS OF THE AMBIGUITY AT POSITIONS 2 (700 MS BEFORE VERB OFFSET) AND 3 (FOLLOWING VERB OFFSET) Meaning of ambiguity Primary Secondary

Probe type

Probe position 2

Probe position 3a

Related Control Related Control

698* 750 721* 766

702 (n.s.) 686 730 (n.s.) 715

* significant difference between related and control probes; "n.s. = nonsignificant difference for related-control contrast.

The overall ANOVA was run on individual subjects' data. The two versions and four lists were used as between-subjects factors (with materials counterbalancing factors) and Probe Positions 2 and 3, ambiguity meaning (primary vs. secondary), and probe type (related vs. control) as within-subjects factors. Overall significant main effects of probe position (F = ll.52, p < 0.01), probe type (F = 5.ll7, p < 0.02) were established. Additionally, there was a significant effect of interaction for Ambiguity Probe Type and Position (F = 6.140, p < 0.01). These effects were then examined by employing a priori planned comparisons of the probe type condition (related vs. control probes) for each of the ambiguious meaning conditions (primary and secondary) at each of the two Probe Positions 2 and 3. At Probe Position 2, there was significant main effect for the factor Probe Type (F = 7.944, p < 0.01). There was significant priming effect for related, as compared with control, probes for the two meanings (primary and secondary) of the ambiguity. At Probe Position 3 we found a different result, which did not confirm the study by Love and Swinney (1996). At this position there was no significant priming effect for related, as compared with control, probes for either the primary or the secondary meaning of the ambiguity. This demonstrated lack of activation above the usual level for the antecedent at the position of the gap in Bulgarian.

5. DISCUSSION Our data provide partial replication of results obtained in studies with English language object-relative constructions. At Position 1, all meanings of the ambiguity were activated compared with controls. This confirms the cross-language

Bulgarian: Lexical Access and Processing

179

validity of the mechanism of the context-free nature of lexical access, at least for these two languages. At Position 2, both meanings were still activated, and the difference between related versus control conditions was not statistically significant. At Position 3 none of the meanings were activated significantly more than any other (unlike the results of Love and Swinney where only the context-relevant meaning of the ambiguity was reactivated). Having in mind the peculiarities of the Bulgarian language that are reflected in on-line sentence processing, we expected a close replication of results based on data from the English language (Love and Swinney, 1996) as far as there is syntactic parallelism in the sentence structure and no additional interference from other factors. Indeed, for Position 1, the results were replicated because of the close structural similarity. Positions 2 and 3 are within the subordinate clause, which in Bulgarian allows for the influence of several language-specific factors contributing to the on-line coreference disambiguation resolution (e.g., the relatively free word order, the presence of the gender- and number-marked relative pronoun, and the possibility for pro-drop due to the reliability of subject-verb agreement). Due to these specific features in the subordinate clause, which includes Positions 2 and 3, it is difficult to achieve a replication of the available effects from experiments with English language material. Thus, the relatively free word order in object-relative clauses in Bulgarian, the influence of gendernumber agreement between the relative pronoun and the object NP in Bulgarian provide earlier (than English) cues for filler-gap binding. Another (and most generally formulated) consideration concerns the language-specific interface(s) between structural (syntactic) and semantic (thematic role) constraints. To illustrate the latter point with an example, a study of Bulgarian language online sentence processing reveals a different hierarchy of cues for agency assignment in transitive sentences. Agreement (a syntactic cue) and Animacy (a semantically motivated cue) play a more dominant role than word order (cf. Andonova et al., 1996). The opposite is the case for English where word order is the most dominant cue (i.e., the decision appears to be more syntactically based). The difference in the results for English and Bulgarian language data may orient us toward a reconsideration of the theoretical model as well as to a more detailed analysis of the possible sources for cross-language variability in the components contributing to the gap-filler coreferencing. More studies with CMLP imposing further restrictions and variation in the test materials in order to control these language-specific components will provide more detailed evidence for the time course of the object-relative clause constructions processing. It is also appropriate to point out, from a theoretical point of view, that in Love and Swinney (1996) there is no explicit commitment to an interpretation of the "deep structure" one is supposed to test with CMLP. This deeper representation may be conceptual (even nonlinguistic) in nature (Love and Swinney, 1996).

180

Maxim I. Stamenov and Elena Andonova

It is not completely implausible to maintain that there are two different formats of deep structure, and CMLP tests the outcome of the joint and coordinated action of D-structure and LF (cf. Stamenov, 1997, for a discussion of different alternatives, in this respect). This possibility would be once again in compliance with generative grammar following the latest Minimalist Program (cf. e.g., Chomsky, 1995, p. 132) and maintaining that antecedent reactivation may require not just access to D-structure, but also to LF or to a coordinated interfacing of S-, D-structure, and LF in on-line processing in looking for a "common solution." This point has implications for the present study. If we have two different deep structure formats, for example D-structure and LF, their relative contribution could be of different value during different stages of sentence processing for different languages (e.g., at the offset of the relative pronoun vs. the offset of the verb vs. the offset of the sentence as a whole). The difference in the data between Bulgarian and English suggests that different features of sentence processing may contribute (probably at different stages of processing) to the local and global ambiguity resolution and coreferencing. Factors of this type may very well determine the earlier (or later) realization of ambiguity resolution and gap filling in Bulgarian versus English. Of course, this hypothesis is contingent on endorsing the possibility of interaction between the modules of lexical access, syntactic parsing, and semantic theta roles grid and establishing levels in modularity versus interaction (e.g., for local vs. global ambiguity resolution), but the details of modeling and experimentation await further research.

ACKNOWLEDGMENTS Equipment for this study was furnished by a grant from the MacDonnell Foundation for Cognitive Neuroscience in Eastern Europe and a grant from NIH-NIDCD to Elizabeth Bates, "Cross-linguistic Studies of Aphasia." We would like to express our special thanks to Ivailo Vlaev for his competent assistance in conducting the experiment. We are also grateful to Rumyana Kirova, Boryana Kadiyska, and Andrea Kulakov from the Cognitive Science Department for their help with the realization of different tasks related to the experiment.

REFERENCES Andonova, E., Gerganov, E., Misheva, A., and Petrov, A. (1996). Sentence interpretation in Bulgarian. Unpublished manuscript.

Bulgarian: Lexical Access and Processing

181

Boland, J., and Cutler. A. (1996). Interaction with autonomy: Output models and the adequacy of the great divide. Cognition 58, 309-320. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. Chomsky, N. (1981). Lectures on Government and Binding. Dordrecht: Foris. Chomsky, N. (1995). The Minimalist Program. Cambridge, MA: MIT Press. Friederici, A. D., and Mecklinger, A. (1996). Syntactic parsing as revealed by brain responses: First-pass and second-pass parsing processes. Journal of Psycholinguistic Research25, 157-176. Hillert, D. (1997). Language in time: Lexical and structural ambiguity resolution. In M. Stamenov (Ed.), Language structure, discourse and the access to consciousness (pp. 77-112). Amsterdam & Philadelphia: John Benjamins. Love, T., and Swinney, D. (1996). Coreference processing and levels of analysis in objectrelative constructions: Demonstration of antecedent reactivation with the CrossModal Priming paradigm. Journal of Psycholinguistic Research 25, 5-24. MacDonald, M., Pearlmutter, N. J., and Seidenberg M. S. (1994). Lexical nature of syntactic ambiguity resolution. Psychological Review 101, 676-703. Stamenov, M. (1997). Grammar, meaning and consciousness: What sentence structure can tell us about the structure of consciousness. In M. Stamenov (Ed.), Language structure, discourse and the access to consciousness (pp. 277-343). Amsterdam & Philadelphia: John Benjamins. Swinney, D. (1979). Lexical access during sentence comprehension: (Re) consideration of context effects. Journal of Verbal Learning and Verbal Behavior 18, 645-660. Swinney, D., Ford, M., Frauenfelder, U., and Bresnan, J. (1987). The time course of coindexation during sentence comprehension. Paper presented at the Psychonomic Society Meeting, Seattle. Tanenhaus, M. K., and Trueswell, J. C. (1995). Sentence comprehension. In J. L. Miller and P. D. Eimas (Eds.), Speech, language, and communication (pp. 217-262). San Diego, CA: Academic Press. Urbach, T. (1997). From brain to mind: Event-related potential evidence for sentence comprehension processes. American Behavioral Scientist 40(6), 754-781.

NOTE 1

It is important to acknowledge that this is an operation necessary in order to accomplish sentence processing. It is open to discussion to what degree discourse-specific intersentence coreferencing is enacted by the same mechanism.

This page intentionally left blank

THE CONTRIBUTION OF WORD FORM AND MEANING TO LANGUAGE PROCESSING IN SPANISH: SOME EVIDENCE FROM MONOLINGUAL AND BILINGUAL STUDIES JOSE E. GARCIA-ALBEA * ROSA M. SANCHEZ-CASAS* JOSE M.IGOA *Department of Psychology Universitat Rovira i Virgili Tarragona, Spain Department of Basic Psychology Universidad Autonoma de Madrid Madrid, Spain

1. INTRODUCTION One of the keynotes characterizing current language research is the special attention paid to the lexical component of the language faculty. Both in linguistic theory and in models of language processing, lexical information is playing a crucial role. Apart from specifying the particular features that identify each language, the lexicon constitutes a convenient meeting point for the other general components—phonological/orthographic, syntactic, and semantic—to interact, since all these types of information are included in our knowledge of a word. As a matter of fact, the increasing weight attributed to the lexicon in the explanation of syntactic phenomena (Bresnan, 1982; Chomsky, 1981, 1995; Grimshaw, 1990; Pustejovsky, 1995) has coincided with the progressive interest of psycholinguistics in drawing the corresponding links between word- and sentence-processing Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

183

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

184

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

mechanisms (see, for example, Clifton, Frazier, and Rayner, 1994; MarslenWilson, 1989; Simpson, 1991). To a varying extent, the way most languages project lexical information into syntactic processing is through the intervention of different morphological devices, thus coded in the structural form of words. In this sense, morphological features are primarily a matter of lexical description, and only subsequently they may affect the other linguistic levels, since in all of them—semantic, syntactic, and phonological/orthographic—morphological information may have noticeable consequences. That is also why the influence of morphological variables on language processing has been mainly addressed in the context of lexical representation and processes, the same context in which we will remain in this chapter. The psychological reality of morphological distinctions has been widely evidenced in all areas of psycholinguistic research, from language comprehension to language production, and from language acquisition to language breakdown (see Del Viso and Garcia-Albea, 1985, for a review of the classical studies in each area). However, and perhaps paradoxically, the lexical status of morphology, as a genuine level of representation and processing, has not been always acknowledged and still is the object of vivid debate both in linguistics (e.g., Aronoff, 1994; Halle and Marantz, 1993; Lieber, 1992) and psycholinguistics (e.g., Chialant and Caramazza, 1995; Marslen-Wilson, Tyler, Waksler, and Older, 1994; Sandra, 1994; Seidenberg, 1987). Perhaps because of its allegedly diverse functionality, morphology has tended to be subsumed under phonology or syntax, sometimes being also considered as no more than a by-product of form and meaning word relations. This chapter focuses specifically on this issue, which will be addressed from the perspective of studies on word recognition. In accordance with the crosslinguistic import of the volume, we will discuss research done with Spanish monolinguals, as well as with Spanish/English and Catalan/Spanish bilinguals. We use languages with rich morphological resources, such as Spanish and Catalan—both from the same Romance family—and contrast them with a less morphologically rich language, such as English. At the end, the possibility of generalizing over subjects and languages will provide a strong test of our conclusion's validity. The study of morphological processing within the domain of word recognition has been the focus of interest in psycholinguistic research, since the mid-1970s, when Taft and Forster published their seminal work on how polymorphemic words could be stored and retrieved during the process of visual word recognition (Taft and Forster, 1975, 1976). This work provided the starting point of what became a very productive line of research, which continues today (see, for example, Feldman, 1995; Frost and Katz, 1992; Henderson, 1985, 1989; Sandra and Taft, 1994; Taft, 1985). This steady interest for morphological issues has given rise to a large number of studies that, using a variety of tasks and experimental paradigms, have investigated why and how morphological structure could be involved at the level at which words are represented and accessed during word recognition.

Spanish Word Form and Meaning

185

In our attempt to examine to what extent the morphological relationships could serve as a criterion to model the organization of the internal lexicon, we have taken a different approach from that generally adopted in the majority of the initial studies carried out on this question. These studies have provided valuable evidence that suggests that morphological relatives are closely associated in the lexicon, and researchers have interpreted this evidence as indicating the existence of a level at which morphological information needs to be represented. However, this evidence does not seem to be by itself sufficient to exclude other alternative explanations that minimize the role of the morpheme. In particular, some current connectionist models question the psychological status of the morphological constituents of the words and claim that the apparent reported effects of morphological structure could be accounted for in terms of orthographic and phonological patterns of letter units (e.g., McClelland and Rumelhart, 1981; Rumelhart and McClelland, 1986; Seidenberg, 1987). For instance, Seidenberg (1987) has suggested that syllabic and morphological effects could be explained if one takes into account that the transitional probabilities that straddle syllabic or morphological boundaries are lower when compared to the probabilities of sequences within syllabic and morphological units. It is clear that words that are morphologically related share a common root or stem,1 and because of that, they share orthographic and phonological characteristics, as well as some of their semantic ones. Therefore, it might be possible that morphological relationships could be reduced to a convergence of semantic, phonological, and orthographic relationships, without being explicitly represented in the lexicon. Given these considerations, in order to determine whether or not morphology per se could play a role in lexical organization, it seems important to ask if morphologically related words are something special and different from phonologically, orthographically, and/or semantically related words. During the last decade, an increasing number of studies have addressed this basic question, resulting in a productive approach to this area (e.g., Beauvillain and Segui, 1992; Bentin and Feldman, 1990; Drews and Zwitserlood, 1995; Feldman, 1994; Grainger, Cole, and Segui, 1991; Marslen-Wilson, Tyler, Waksler, and Older, 1994; Napps, 1989; Napps and Fowler, 1987; Stolz and Feldman, 1995). In the second section of this chapter, we discuss some of the evidence, including the data we have obtained in Spanish, which supports the view that morphological relationships are explicitly represented at the level of lexical representation, and appear to be distinct from mere form and meaning relationships. Although the role of morphology in the organization of the mental lexicon has been investigated using different languages, studies on this issue have been restricted to monolingual experiments within each single language. However, it should be noted that models of bilingual lexical organization have traditionally distinguished between form and meaning levels of representation at which words from the two languages of the bilingual could be connected (e.g, De Groot, 1992;

186

J. E. Garcia-AIbea, R. M. Sanchez-Casas, and J. M. Igoa

Potter, So, von Eckhardt, and Feldman, 1984; Kroll, 1993). Recent experiments we have carried out with different bilingual populations (Spanish/English and Catalan/Spanish) have investigated the form and meaning relationships that exist between words across languages. For this purpose, words that have a common root (i.e., cognate translations) were contrasted with words that only have similar meaning (i.e., noncognate translations), and words that are just similar in terms of their form (the so-called false friends). The data from these experiments are discussed in the third section of this chapter, in order to demonstrate that the relations between cognate words cannot be reduced to a mere form and/or meaning similarity, being in this sense comparable to the relations between morphological relatives within the same language. The convergence of the results from the monolingual and bilingual experiments concerning the contribution of form and meaning in both morphological and cognate relationships led us to consider the possibility that the case of cognate words could be assimilated to a morphological relation of a special kind (see also Kirsner, Lalor, and Hird, 1993). This possibility has been investigated in a recent experiment we have carried with Spanish/Catalan bilinguals, where morphological effects across languages and within the same language for both cognate and noncognate words were compared. In the fourth section of this chapter, the results of this experiment are reported to provide evidence consistent with the claim that morphology could serve as a general principle of lexical organization for both monolingual and bilingual subjects. Finally, the implications that this claim may have for current proposals on lexical representation and access are discussed in relation to both monolingual and bilingual word recognition models.

2. ARE MORPHOLOGICAL RELATIONS REDUCIBLE TO FORM AND MEANING RELATIONSHIPS? The priming paradigm has been one of the main techniques used in word recognition studies to examine to what extent morphological relatedness can be dissociated from form and meaning similarity, in order to determine if morphological information as such is coded in the lexicon. This paradigm allows to test whether or not the presentation of a word (the prime) facilitates the recognition of another word (the target), which is subsequently presented. The type of relations between prime and target can be manipulated (identical, morphological, orthographic, or semantic) to see how this affects the pattern of priming effects, generally in a lexical decision task; in this task, subjects are asked to decide whether a sequence of letter constitutes a word or not. First, we refer to the priming studies that focus on the differences between morphology and orthography, reporting our results in

Spanish Word Form and Meaning

187

Spanish; and second, we examine some available evidence concerning the dissociation between morphological and semantic similarity. 2.1. Differences between Orthographic and Morphological Effects Evidence against an orthographic interpretation of the facilitation effect obtained with morphologically related words comes from a variety of studies, using different strategies and priming procedures. One strategy has been to manipulate the degree of orthographic similarity between morphologically related words, and test how this could influence the magnitude of the facilitation effects when primes and targets are displayed at different lags. For instance, a study manipulated orthographic similarity by altering the the two alphabets used in written Serbo-Croatian (Cirilic and Roman), obtaining equivalent facilitation for morphologically related prime-target pairs, regardless of whether the alphabet in which they were printed was the same or different (Feldman, 1992; Feldman and Moskovljevic, 1987). Alternatively, other studies have examined the influence of the orthographic similarity of morphological relatives by comparing regular (e.g., bringing—bring) and irregular inflected word forms (e.g., brought—bring), and have found that both facilitated the recognition of their corresponding base morpheme (e.g., Fowler, Napps, and Feldman, 1985; Napps, 1989; Stanners, Neiser, Hernon, and Hall, 1979; but see Kempley and Morton, 1982). A different strategy for showing that orthographic similarity does not underlie morphological facilitation has been to compare morphologically related primes to unrelated primes that only share orthographic characteristics. With such strategy, facilitation effects have been reported when the target was preceded by a morphologically related prime. However, if the relation between prime and target was only orthographic, the tendency was to find inhibition and not facilitation. This has been found to be the case for conditions in which prime and target were displayed at different lags of intervening items (around 10 items by average) (e.g., Feldman and Andjelkovic, 1992; Stolz and Feldman, 1995; Drews and Zwitserlood, 1995), and in contiguous priming presentation at a short delay (300 ms) between prime and target presentation (Drews and Zwitserlood, 1995). It is undeniable that the above-mentioned studies provide valuable evidence concerning the differences between morphological and orthographic relations. However, the unmasked priming procedure used in these studies does not exclude the possibility that strategic or episodic factors could influence the reported effects. It has been shown that episodic factors can contribute to long-term repetition priming (Feldman and Moskovljevic, 1987; Feustel, Shiffrin, and Salasoo, 1983; Forster and Davis, 1984; Oliphant, 1983). Moreover, even if short SOAs are used, there is no guarantee that the observed effects reflect automatic processing rather than strategic effects (De Groot, 1983; Neely, 1991). With short

188

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

SOAs, subjects are still fully aware of the prime, and therefore they may use predictive strategies. A procedure that has proven able to attenuate the application of such strategies is that of masked priming (Forster and Davis, 1984). In this procedure, the stimulus sequence in each trial is presented in quick sucession, each stimulus appearing superimposed on the preceding one. First in the sequence is a forward mask displayed for 500 ms; second, a prime is presented in lower case for around 60 ms; and third, a target that also acts as a backward mask is presented in upper case for 500 ms. The short presentation of the prime and the combined action of the forward and backward masking make the prime generally unavailable for conscious report, reducing to a minimum the possibility of strategic contamination. Evidence of the lexical nature of the priming effects obtained under these conditions has been reported in different studies (Davis, 1990; Forster and Davis, 1984; Forster, 1987; Forster, Davis, Schoknecht, and Carter, 1987; SanchezCasas, 1988, 1996; Segui and Grainger, 1990). Given these considerations, it seems more appropiate to test further the differences between orthography and morphology, using a masked priming procedure. Forster et al. (1987) were the first to demonstrate clear facilitation effects when the masked prime was morphologically related to the target (e.g., make-MADE, keep-KEPT). In addition, these effects were of the same size as those obtained in an identity condition. However, comparable effects were not found if the prime and target only share orthographic characteristics. In this case, inhibitory effects were observed in conditions where the prime has a higher frequency than the target (e.g., blue—BLUR) in relation to an unrelated condition (cart—BLUR) (Grainger and Segui, 1990). A direct comparison of the effects of morphological and orthographic priming have been carried out in two visual word recognition studies, using a masked priming procedure. Grainger, Cole, and Segui (1991) contrasted those effects in French derivationally related words (e.g., prenom— surnom, mural—muret) in a lexical decision task. Overall, they found facilitation effects for the morphologically related word pairs, regardless of whether the base or the derived word was presented as the prime, whereas orthographically similar pairs resulted in inhibition. Further evidence in support of the qualitative differences between these effects conies from experiments reported recently by Drews and Zwitserlood (1995) in Dutch and German. Using a lexical decision task, these authors only found consistent facilitation effects in the case of morphologically related parrs, specifically regular inflected forms (e.g., kellen-KELLER), being not significantly different from those observed with identical pairs. Primes and targets that were only orthographically related (e.g., keller-KELLE) resulted in inhibition, as previous findings had already shown (Grainger and Segui, 1990; Grainger et al., 1991). Morphological inflected priming was also obtained when the same experiments were carried out using a naming task, although in this case, orthographic similarity also produced facilitation effects.

Spanish Word Form and Meaning

189

Regardless of the possible explanation of the different pattern of results obtained with orthographically similar words across the two tasks, it is clear that these words do not consistently produce facilitation effects. Moreover, as Drews and Zwitserlood (1995) have pointed out, orthographic priming can be modulated by many factors (e.g., word frequency and/or neighborhood density, material constraints, etc.) that have not been found to affect morphological priming. Therefore, it can be concluded that masked priming data provide evidence in support of the different status of morphology and orthography, ensuring that the reported facilitation effects produced by morphological relatives do reflect processing at the lexical level. The Spanish priming experiment we describe in the remainder of this section was also designed to contrast morphological and orthographic priming effects under masked presentation conditions. The purpose of this contrast was not only to attempt to generalize the results across a different language from the ones that have been used, but also to provide further evidence that mere form similarity cannot account for the effect of morphological relatedness. In particular, we compared priming effects in inflected, derived, and orthographically related pairs, matching syntactic category, word length, and the degree of orthographic similarity of the members of each pair, within and across the three types of word relations. Given the strict orthographic match between the related pairs, any differences between the priming effects observed could be attributed to the role of morphology. Gender was used as the morphological relation that holds between the inflected words, which is one of the distinctive features of the inflectional morphology in Spanish. Elias-Cintron (1995) has recently proposed that Spanish words could be classified into four types based on their morphological, syntactic and semantic gender properties. One of these types is formed by words that accept both feminine and masculine values for morphological gender (as well as for syntactic and semantic ones) and whose meaning across the word pair varies only in the biological sex of the referent. In these words, the marker -a is added to their stem to construct the feminine gender (e.g., nina), and the marker -o to build the masculine gender (e.g., nino). Pairs of words of these characteristics were selected to conform the inflected morphological priming condition. All the words were bisyllabic nouns with a monosyllabic stem. Members of word pairs had the same length (4 or 5 letters) and orthographic overlap, differing just in the final letter which could be either "a" or "o" (e.g., nino-NINA). The second type of morphological relation examined in our experiment was derivation. We selected derived word pairs that closely matched the characteristics of the gender-inflected ones. That is, they were four- or five-letter-long nouns that shared all the letters except the last one, which was also the suffix "a" or "o," associated in this case with an invariant feminine and masculine gender respectively (e.g., rama-RAMO).2 Finally, a third type of related word was included whose relation was only orthographic. The orthographically similar word pairs

190

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

were also matched in syntactic category and letter length to the inflected word forms. Similarly, the words within each pair differed only in the final letter that was either "a" or "o" (e.g., foca-FOCO). As with derivationally related words, orthographically related forms have an invariant morphological gender (see EliasCintron, 1995, for a description of the words of this gender type), but they are semantically unrelated (i.e., foca means 'seal,' andfoco, 'floodlight'). Priming effects for the morphological and orthographically related word pairs described above were measured against two controls. In one of the controls, the word presented as the prime was unrelated to the target, without sharing any orthographic or semantic characteristics (e.g., celo-NINA, cita-RAMO, sumaFOCO). In the other, prime and target were similar in form but not in meaning, and differed from each other in the initial letter (e.g., pina-NINA, gamo-RAMO, loco-FOCO). This form control manipulation, together with the one provided by the orthographically related condition (e.g., foca-FOCO), allowed us to examine to what extent altering letters from the initial (i.e., the stem) or the final part of the word (i.e., the gender suffix) could affect the pattern of orthographic priming effects. If the initial information of the word plays an important role in word recognition (see Sanchez-Casas, 1988, 1996, for a review), and form similarity by itself results in inhibition, it might be possible that prime-target pairs that differ in the first letter produce smaller inhibitory effects than those pairs where the final letter is altered. In sum, there were three priming conditions for each type of word relation. For the inflected and derived related pairs, the word presented as the target could be preceded by three different primes: a morphological relative, a form-related word, and an entirely unrelated word. For the words that were only similar in form, the three types of primes were the following: a form-related prime that differed in the final letter from the target (comparable to the morphological relative prime); a form-related control word where the initial letter was different, and a totally unrelated word. A set of nonwords was constructed of the same size as the word set and distributed across comparable priming conditions. The experimental task we used was the lexical decision task with reaction time and percentage of errors as the two dependent variables. Words and nonwords were visually presented under masked priming conditions. The prime appeared in lower case for 64 ms and the target in upper case during 500 ms; the forward pattern mask was a sequence of hatches (i.e., #####) and was displayed for the same duration as the target (Forster et al., 1987). The word presented as the target ended with the letter "o" in aproximately half of the cases and in the letter "a" in the other half. It is important to notice that primes and targets in each pair were of exactly the same length, which is not common in the materials of the morphological priming experiments carried out in other languages. A group of students from the Universidad Autonoma of Madrid participated in the experiment. Overall, our data confirmed previous findings with regard to the pattern of mor-

Spanish Word Form and Meaning

191

TABLE 1 MEAN LEXICAL DECISION TIMES (MS) AND PERCENT ERROR RATES FOR EACH WORD RELATION, AS A FUNCTION OF PRIME TYPE

Prime type Inflection Morphological Form control Unrelated control Derivation Morphological Form control Unrelated control Orthographic Form related Form control Unrelated control

Word relation

Reaction time (ms)

Error rates (%)

nino-NINA pina-NINA celo-NINA

624 673 667

3.7 5.7 7.0

rama-RAMO gamo-RAMO cuna-RAMO

650 694 683

7.4 13.2 12.0

foco-FOCA loca-FOCA reto-FOCA

687 693 708

14.6 16.0 9.8

phological priming effects, though some discrepancies were found concerning orthographic priming. The results of our experiment are summarized in Table 1. A significant facilitation effect for both morphologically related words was found relative to the two controls. Morphological inflected forms showed a 43-ms difference with respect to the unrelated control, and a 49-ms difference in relation to the form-related control (where prime and target differed in the initial letter). A similar pattern of results was obtained for the morphologically derived words. Here, the size of the facilitation effect was 33 and 44 ms when compared to the unrelated and the form-related control, respectively. In contrast, the orthographically related words did not produce any reliable facilitatory or inhibitory effects. The 21-ms difference observed with respect to the unrelated control was only significant in the subject analysis, and the difference obtained relative to the form control was practically negligible. When comparing the form-related control with the unrelated control, neither significant facilitation nor inhibition was found. In the inflected and derived related words, the difference between the two control conditions was 6 and 11 ms, respectively, in favor of the unrelated control. In the orthographically related words, there was an advantage of 15 ms for the formrelated control, though this control produced significantly more errors than the unrelated one. It is important to mention that a similar pattern of priming effects was obtained in a further Spanish experiment, where basically the same materials and priming conditions were used. Once again morphological relatedness produced facilitation, while orthographic similarity by itself did not show any consistent effects (Sanchez-Casas, 1995).

192

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

Taken together, our results confirm previous masked priming studies that show facilitation effects with inflected and derived words (Drews and Zwitserlood, 1995; Grainger et al., 1991) as effects of a similar size. Critically, significant facilitation effects are not generally observed with orthographically similar primetarget pairs, reinforcing the claim that morphological priming cannot be reduced to form facilitation. When examining the pattern of effects concerning orthographically related pairs, our results do not support their reported inhibitory character (Grainger and Segui, 1990; Grainger et al., 1991; Drews and Zwitserlood, 1995). The manipulation of the orthographic similarity did not show any consistent pattern. When prime and target differ in the initial letter, there are no significant differences with respect to the unrelated control condition, and when varying the final letter, there is only a marginal trend towards facilitation. From this, we can conclude that the lack of consistent orthographic effects cannot be attributed to the position of the letters within the word (either in or outside the monosyllabic stem). We do not have a clear explanation to account for the differences with regard to orthographic priming between our results and previous ones. Although, as mentioned earlier, different factors seem to modulate orthographic priming, our form-related pairs share critical characteristics with those used in the studies where orthographic inhibition has been found (e.g., lexical status, short length, relative frequency of the prime). So, it does not seem likely that our different pattern of results could be due to the materials we used. Whatever may be the explanation for this conflictive result, the important finding to emphasize now is that this type of priming does not produce the same results as the morphological one. In addition to being orthographically related, words that are morphologically related also have a semantic relation. Therefore, it is possible that the semantic similarity between the prime and the target could contribute to the facilitatory effects obtained with morphological relations. We will briefly refer to some of the studies that have tested such a possibility, before discussing the bilingual experiments related to the issues under examination. 2.2. Differences between Semantic and Morphological Relations The extent to which the degree of semantic overlap influences morphological priming effects has been investigated in lexical decision experiments, using unmasked priming procedures and derivational morphology (e.g., Bentin and Feldman, 1990; Feldman, 1992; Feldman and Andjelkovic, 1992 ;Marslen-Wilson et al., 1994; Stolz and Feldman, 1995). For instance, Feldman and Stotko (cited in Feldman, 1992), obtained the same facilitation effects, at long lags, for English morphologically related pairs that were semantically transparent (e.g., creationcreate) and opaque (e.g., creature-create). Findings along the same line were reported by Bentin and Feldman (1990) in an experiment conducted in Hebrew.

Spanish Word Form and Meaning

193

These authors contrasted facilitation effects from primes that were related only morphologically to the target (opaque forms, e.g., number-library, which in Hebrew share the same root), both morphologically and semantically related (transparent forms, e.g., librarian-library), or only semantically (e.g., readinglibrary). The results showed that morphological priming did not change across the two prime-target lags used, whereas semantic priming did. Semantic transparency did not affect the magnitude of facilitation effect at long lags (15 intervening items), though opaque morphological relations produce smaller facilitation effects than the transparent ones when a 0 lag was used (no itervening items). MarslenWilson et al. (1994) have also found that semantic transparency affects morphological effects with contiguous prime-target presentation. Using a cross-modal word priming and a SOA of 200 ms, they only obtained facilitation with semantically transparent derivationally related words. These previous findings do not allow to rule out entirely the possible contribution of semantic transparency to morphological facilitation. However, evidence from masked priming studies, where the influence of episodic factors is minimized, suggests that semantic and morphological priming effects do not seem to be the same. On the one hand, and in contrast to the case of morphological priming effects, the results on semantic priming under masked presentation conditions have not shown consistent facilitation effects. While some lexical decision studies have found facilitatory semantic priming on targets preceded by masked primes (e.g., Balota, 1983; Fowler, Wolford, Slade, and Tassinary, 1981; Marcel, 1983), others have reported this priming to be lower than the one observed with unmasked prime presentation, and in some conditions, to be practically inexistent (De Groot, 1990; Durante and Hirshman, 1994). On the other hand, there is more direct evidence that morphological priming cannot be reduced to mere semantic facilitation. In a recent unpublished study carried out in Spanish, Dominguez, Cuetos and Segui (1997) compared priming effects in morphologically related prime-target pairs (e.g., loca-loco) and semantically related pairs (e.g., sana-odio) relative to an unrelated condition (e.g., cera-loco, lisa-odio). They used a masking procedure with SOAs of 32 and 64 ms. In both cases, morphological relations produced significant facilitation effects (29 ms and 64 ms, respectively), while semantic facilitation was almost absent. Some additional, though indirect evidence, that the degree of semantic similarity does not appear to affect morphological priming effects comes from our Spanish results described earlier. Although both the inflected and derived related word pairs have a semantic relation, the predictability of their meaning from the shared stem and the degree of semantic overlap is not the same. For instance, the meaning of nino (boy) is more closely connected to the meaning of nina (girl), whereas the meaning connection is more remote in the case of ramo (bunch of flowers) and rama (branch). Despite these differences, which are characteristic of

194

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

these two types of morphological relations, both inflections and derivations produced similar facilitatory effects. Nevertheless, further experiments need to test different meaning relations within derivations to draw definitive conclusions. The different studies reviewed in this first section have generally provided evidence that morphological priming effects can be dissociated from both orthographic and semantic effects. Such a dissociation can be interpreted to support the claim that morphological relations are something different from form or meaning relationships, and that they can be coded in the lexicon as a proper level of representation. In the next section, we will explore further the contribution of form and meaning to priming effects in bilingual visual word recognition, reporting some experiments we have carried out with Spanish/English and Spanish/Catalan bilinguals.

3. HOW DO FORM AND MEANING CONTRIBUTE TO THE COGNATE EFFECTS IN BILINGUALS? Words from two languages share form and/or meaning characteristics to different degrees. There are words, the so-called cognate words, that have a common root, and they are semantically and orthographically similar (e.g., rico-rich, torre-tower). There are other words that are only similar in meaning (i.e., noncognate words, e.g., libra-book, hoja-sheet) and others that are only orthographically related, commonly known as false friends (e.g., gamo-game, torno-torcti). The contrast between these types of word relations offers the possibility of examining, now across languages, the role that form and meaning similarity can have in lexical representation. Different priming studies, across a variety of languages, have shown that priming effects depend on the cognate status of the translation. Scarborough, Gerard, and Cortese (1984) first presented a set of Spanish words for lexical decision, followed by a set in English, where half were noncognate translations of the Spanish words in the first set. They did not find any evidence of facilitation effects. The same results have been also obtained in other studies with French-English bilinguals (Kirsner, 1986; Kirsner, Brown, Abrol, Chadha, and Sharma, 1980) and English-Hindi bilinguals (Kirsner, Smith, Lockhart, King, and Jain, 1984). In contrast, cognate translations have been found to produce facilitation (Cristoffanini, Kirsner, and Milech, 1986; Kerkman, 1984, cited by De Groot, 1993). In a study comparing English-French cognate and noncognate translations, as well as false friends with exactly the same spelling, Gerard and Scarborough (1989) reported facilitation effects with cognate translations, but no effect at all with noncognates. Interestingly, they also found facilitation with false friends, which was as large as that observed with cognates (see however, Kerkman, 1984, in

Spanish Word Form and Meaning

195

De Groot, 1993). This latter finding suggests that orthographic similarity contributes to the cognate priming effects. Cristoffanini et al. (1986) found data that support further this suggestion, by demonstrating that these effects decreased as a function of orthographic similarity (e.g., festival-festival, observation-ofoervacion, publicity-publicidad). All of these studies used unmasked priming procedures which, as mentioned earlier, can be subject to episodic contamination. A detailed comparison of cognate-noncognate effects has been carried out by De Groot and Nas (1991) in a series of experiments with Dutch-English bilinguals, using masked and unmasked priming procedures. Consistent with the view that there may be additional sources of (nonlexical) facilitation for unmasked primes, overall priming effects were larger with unmasked displays. Both cognate and noncognate translations produced facilitatory effects, though the absence or presence of masking interacted with translation type. With unmasked prime displays noncognate translations produced a larger facilitation effect than the cognate translation, but when prime presentation was masked, the situation was reversed. One caution that should be noted regarding De Groot and Nas's findings is that their procedure for masking the prime may have not been completely secure. In their procedure, primes were presented in upper case and the targets in lower case. Such presentation could have led to their primes being "more available" to a subject's general decision system than in the procedure of Forster and Davis (1984), and this may in turn have affected the pattern of effects. Support for this proposition is suggested by two findings. The first is that priming effects obtained using the upper-case prime and lower-case target procedure used by De Groot and Nas attenuates the standard frequency effect (see Sandra, 1990). This result is commonly found using an unmasked procedure when subjects are fully aware of the prime (e.g., Scarborough, Cortese, and Scarborough, 1977), but not with the masked priming technique (Forster and Davis, 1984; Segui and Grainger, 1990). Forster and Davis (1984) have suggested that the attenuation of the word frequency effect by a priming effect is evidence of the operation of nonlexical processes; proceses which may be similar to those which augment a subject's performance on low frequency words in a recognition task. More straightforwardly, it is clear from a series of contrasts made in the De Groot and Nas (1991) experiments that features of prime presentation do affect the size of priming effect. Indeed, the size of the noncognate translation priming in their experiments appeared to decrease across conditions of prime display, plausibly as a function of prime availability. For instance, the amount of noncognate priming was reduced from 113 ms, when the prime was clearly displayed, to 40 ms for the De Groot and Nas display and then to 22 ms for the display similar to Forster and et al's (with this residual effect not differing significantly from the control condition). This trend in the reduction of the priming effect begs the question of what would happen with even tighter controls on masking the prime. In this respect, it is important to mention that when De Groot and Nas used

196

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

lower-case primes with upper-case targets, they found a large (although only marginally significant) difference between the amount of priming produced by cognate and noncognate translation types. Consistent evidence that facilitation effects are only observed with cognate words is provided in a series of experiments we have carried out with different bilingual populations. In all the experiments, we used the masked priming procedure as described before (see section 2.1.), and a lexical decision task. Our first experiments with Spanish-English bilinguals (Davis, Sanchez-Casas, GarcfaAlbea, 1991) compared priming effects in cognate (e.g., rico-RICH) and noncognate translations (e.g., luna/MOON), in both language directions (i.e., Spanish prime-English target and English prime-Spanish target). Primes and targets within each pair were of the same length (4 or 5 letters). Priming effects with translation pairs were measured with respect to two conditions: identity and control. In the identity condition, the prime was the same word as the target to determine whether or not priming within and across languages was the same. In the control condition, the target was preceded by a nonword prime that had the same orthographic overlap with the target as the translation pairs (e.g., rict-RICH, suna-MOON). In order to test to what extent cognate priming effects could be due to their degree of orthographic similarity, we selected cognate pairs that varied in the number of letters they shared (e.g., rico-RICH, torre-TOWER, claro-CLEAK). We were also interested in investigating the influence of language dominance that had not been considered in previous studies but has been shown to affect performance in other bilingual tasks (see Kroll, 1993; Kroll and Sholl, 1992, for a review). With this aim, we tested balanced, Spanish-dominant and Englishdominant bilinguals, as well as a group of subjects who, at the time of testing, had not acquired a good competence in their second language (so-called semibilinguals). Different groups of students from St. Louis University (Madrid Campus) participated in the experiments. The overall pattern of results could be summarized as follows: (a) cognate translations showed facilitation effects, being of the same magnitude as that obtained with identical pairs, whereas noncognate translations did not show significant facilitation; (b) the cognate effects were of a similar size over target language; (c) the degree of orthographic similarity between cognate words did not affect the magnitude of the facilitation obtained, contrasting with previous findings with unmasked priming procedures (see Cristoffanini et al., 1986); (d) finally, the language dominance of the subject did not affect the cognate priming effects, with the three groups of bilinguals (balanced, Spanish-dominant and Englishdominant) showing the same facilitation effects. However, no evidence of these effects was observed by the semibilingual subjects, which suggests that a sufficient degree of competence in both languages is required for cognate effects to emerge (see also Mildred, 1986). These findings clearly showed that cognate words are different from noncog-

Spanish Word Form and Meaning

197

nate words with regard to priming effects. More importantly, this difference suggests that both form and meaning similarity seem to be required for crosslanguage facilitation to emerge. In addition, the same facilitation effects were obtained with identity and cognate primes. Assuming that identity priming is the consequence of the activation of the same lexical representation, this result points towards the possibility that cognate translations may share a common entry in the lexicon (see Sanchez-Casas, Davis, and Garcia-Albea, 1992). In order to provide further evidence for these conclusions, we carried out recently an experiment with Spanish/Catalan bilinguals (Garcia-Albea, SanchezCasas, and Valero, 1996). Being two Romance languages, Spanish and Catalan have a great orthographic and meaning overlap. Thus, it is possible to examine in more detail the contribution of form and meaning to bilingual lexical representation. The masked priming procedure and the task we used were the same as in the Spanish-English bilingual experiments. Translation priming effects were also examined in both language directions. Language dominance was not manipulated in this experiment, because our Spanish-Catalan bilinguals were similarly fluent in both languages. We selected a set of cognate (e.g., tarde-TARDA) and noncognate translations (gabia-JAULA), as well as a group of false friends pairs (fleco-FLECA). False friends are words only similar in form, and in this respect they are comparable to the orthographically related words in our morphological experiment (e.g.,focoFOCA, see section 2.1). Within each set of pairs, the word presented as the target could be preceded by an identical prime, a translation prime, and a nonword control. This nonword prime had the same orthographic overlap with the target as the prime-target translation pairs (e.g., tardi-TARDA, prama-JAULA, fleceFLECA). Primes and targets within each pair were of the same length (4 or 5 letters). A group of students from the Universitat Rovira i Virgili from Tarragona (Spain) participated in the experiment. The results are summarized in Table 2 (Spanish-Catalan direction, SC) and Table 3 (Catalan-Spanish direction, CS). They replicated the pattern of priming effects of the Spanish-English bilingual data. Only facilitatory effects were found with the cognate prime-target pairs when compared to the form control, for both language-priming directions (23 ms and 36 ms for SC and CS respectively). However, null-significant effects were found in the case of noncognate pairs in the two directions. Similarly, identity priming was not significantly different from the cognate priming. In contrast to data from unmasked priming studies (Gerard and Scarborough, 1989), false friends did not produce any facilitation effects. It is clear from the above-reported data with both Spanish-English and SpanishCatalan bilinguals that mere form similarity is not enough by itself to account for the cognate facilitation effects. Although our experiments did not manipulate the semantic similarity of the translations pairs, the consistent absence of facilitation effects with noncognate words (only semantically similar), suggests that meaning

198

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa TABLE 2 MEAN LEXICAL DECISION TIMES (MS) AND PERCENT ERROR RATE FOR EACH WORD RELATION AS A FUNCTION OF PRIME TYPE SPANISHCATALAN LANGUAGE DIRECTION

Prime type Cognates Identity Translation Control Noncognates Identity Translation Control False Friends Identity Form related Control

Word relation

Reaction time

Error rate (%)

cotxe-COTXE coche-COTXE corde-COTXE

658 678 701

12.0 8.0 14.0

sorra-SORRA arena-SORRA cordi-SORRA

654 702 709

11.0 12.0 15.0

curta-CURTA curva-CURTA cuma-CURTA

668 721 722

12.0 21.0 17.0

TABLE 3 MEAN LEXICAL DECISION TIMES (MS) AND PERCENT ERROR RATE FOR EACH WORD RELATION AS A FUNCTION OF PRIME TYPE: CATALANSPANISH LANGUAGE DIRECTION

Prime type Cognates Identity Translation Control Noncognates Identity Translation Control False friends Identity Form related Control

Word relation

Reaction time (ms)

Error rate (%)

tarda-TARDE tarda-TARDE tardi-TARDE

621 625 661

5.0 8.0 7.0

jaula-JAULA gabia-JAULA prama-JAULA

618 674 675

7.0 8.0 10.0

fleco-FLECO fleca-FLECO flece-FLECO

615 660 653

6.0 9.0 8.0

alone cannot account for the cognate effects either. The degree of semantic overlap between translations in cognate and noncognate pairs has been investigated in a series of experiments by Sanchez-Casas, Igoa, and Suarez-Buratti (1992) with Spanish-English bilinguals. They used a translation recognition task, where subjects were visually presented with pairs of words and had to decide whether or not

Spanish Word Form and Meaning

199

the words were translations from each other. Within both cognate and noncognate translations, semantic overlap was manipulated by selecting pairs that had a complete overlap in meaning across the two languages (e.g., leon-lion, vida-life), and pairs that had only a partial overlap (e.g., papel-paper, hoja-sheet). The results indicated that the degree of semantic overlap affected subject's response times in the case of noncognates, but not in the cognates. That is, noncognate translations were recognized faster when they had an identical meaning than when had only part of their meaning in common, while no difference was found between the corresponding cognate conditions. Going back to the masked priming results, the aspect of our findings that we would like to emphasize here is that the pattern of priming effects with cognate words across languages resembles closely those obtained, under comparable conditions, with words that are morphologically related within the same language. That is, in any of the types of word-related pairs, facilitation does not seem to be due to mere orthographic or semantic similarity. This led us to consider the possibility that the cognate relation could be a special kind of morphological relation. Cognate translations are indeed morphologically related words, sharing a common root or stem across languages, and it is possible that they have the same lexical representational status as morphologically related words. This possibility would imply that morphology and not language would be the critical feature of lexical organization in bilinguals, as it has been proposed to be the case in monolinguals. In the next section, we will present some Spanish and Catalan data that support such possibility.

4. IS THE COGNATE EFFECT A SPECIAL KIND OF MORPHOLOGICAL EFFECT? As far as we know, there are no reported studies that have compared cognate priming effects with morphological priming effects within and across languages. We recently carried out a masked priming experiment with Spanish-Catalan bilinguals specifically designed with this purpose (Sanchez-Casas, 1997). In particular, we were interested in testing morphological priming effects across languages with cognate and noncognate words, and to compare these effects with the ones produced by morphologically related primes within the same language, and with the corresponding translation effects. The type of morphological relations we used were gender and number inflections. As we described in section 2.1., gender in Spanish is formed by adding the suffix -a (feminine) or -o (masculine) to the stem. With regard to number formation, either an -s or an -es is added to the stem, depending on whether the word ends in a vowel (-s) or a consonant (-es) (e.g., nino-ninos, leon-leones). All the

200

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

words we selected were bisyllabic nouns or adjectives with monosyllabic stems and had either a cognate or a noncognate translation; their length ranged between four and seven letters. Half of the item set were gender-inflected words and the other half were number-inflected ones. The latter had an invariant morphological gender and formed the plural by adding an -5. The prime was of the same length as the target in the case of the gender-inflection pairs and had one letter less in the number-inflected pairs (i.e., the prime was always the singular form of the word and the target the plural form). Since previous findings had shown that language direction did not affect the pattern of priming effects, only one language direction was used, with target words always in Spanish. Within each of the cognate and noncognate word sets, the Spanish word presented as the target could be preceded by a morphological relative in Spanish, its translation in Catalan, a morphological relative in Catalan, and a nonword unrelated control. Examples of the four priming conditions for cognate and noncognate words are shown in Table 4. A different group of Spanish-Catalan bilingual students from the Universltat Rovira i Virgili (Tarragona) participated in the experiment. Based on previous results, we would expect to find facilitation for morphologically related words within the same language for both cognate and noncognate words. Similarly, facilitation should be obtained in the case of cognate translation pairs but not with noncognate ones. Additionally, if cognate translations can be equated to morphological relations, we would now expect that cognate priming effects will be the same as the morphological priming effects, both within and across languages. Given that noncognate words are morphologically unrelated, no facilitation would be expected with these words either in the translation condition or in the cross-language morphological condition. The results from the experiment are summarized in Table 4. The pattern of results clearly confirms the above predictions. When compared with the control, cognate words showed the same size facilitation effect in the translation condition (52 ms), the within-language morphological condition (56 ms), and the cross-language morphological condition (48 ms). In the case of noncognate words, there were no significant differences between translation, cross-language morphological condition, and unrelated control. Morphologically related primes within the same language produced a facilitation of 51 ms with respect to the unrelated prime, and significant differences of 30 and 31 ms when compared with translation and cross-language morphological conditions, respectively. The above findings clearly show that cognate priming effects are not different from morphological effects, supporting the view that cognate translations can be considered to be a special kind of morphological relations. More interesting, they are consistent with the claim that morphology could be the critical principle of organization not only of the monolingual lexicon, as it has been suggested in some

Spanish Word Form and Meaning

201

TABLE 4 MEAN LEXICAL DECISION TIMES (MS) AND PERCENT ERROR RATE FOR COGNATE AND NONCOGNATE TRANSLATIONS AS A FUNCTION OF PRIME TYPE Prime type Cognates Within-language Morphologically related Translation Cross-language Morphologically related Unrelated control Noncognates Within-language Morphologically related Translation Cross-language Morphologically related Unrelated control

Translation maja-MAJO puerta-PUERTAS maco-MAJO portas-PUERTAS maca-MAJO porta-PUERTAS tica-MAJO cucela-PUERTAS loca-LOCO pato-PATOS boig-LOCO . . . anecs-PATOS boja-LOCO . . . anec-PATOS vima-LOCO selia-PATOS

Reaction Error time rate (%) 579

6.7

583

6.5

587

6.7

635

9.0

570

4.4

600

6.0

601

9.1

621

6.9

proposals, but also in the bilingual lexicon. In our opinion, this claim has important implications for models of the bilingual lexicon, where morphological considerations have not usually been taken into account (see Kirsner et al., 1993, as an exception in this respect). These implications will be discussed in the final section of the chapter, where we will also attempt to examine what the evidence from both monolingual and bilingual data can say about general issues of lexical representation and access.

5. CONCLUDING REMARKS Throughout this chapter, we have intended to show the high degree of convergence reached by two lines of research that have not always been as closely related as it might have been expected. The experimental work with monolinguals and bilinguals presented here starts from a common interest in lexical representation and processes within the field of visual word recognition. The general question we have addressed throughout the chapter has dealt with the psychological reality of morphological word relations, contrasting them with pure orthographic (phonological) and semantic ones. There has also been a strict coincidence in the

202

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

methodological aspects of the study, using equivalent designs and the same experimental paradigm (a lexical decision task under masked priming conditions) in both the monolingual and the bilingual experiments. We first separately examined within- and between-language priming effects for different types of word relations, obtaining comparable results across the two populations: much in the same way as morphological priming within a language could not be reduced to form or meaning facilitation, cognate effects across languages stand well above mere form ("false friends") or meaning (noncognate) relations (sections II and III). This finding led us to plan the last experiment we have reported (section IV), in order to test and compare directly within- and between-language morphological priming effects, using cognate and noncognate word pairs. First, our results replicated the (within-language) morphological effects and the cognate effects found in the former experiments; second, and most important, they revealed the existence of a cross-language morphological effect only for cognate words, which is comparable in size to the cognate-translation effect and the within-language morphological effect. Although not tested in this experiment, all these effects appear to be also comparable to the identity priming effect usually found with the masked priming technique. Taken together, these results give support to the hypothesis that morphologically related words would share a common representation in the internal lexicon. Lexical entries would then better correspond to morpheme-like units than to whole words, which implies that morphologically complex words would be represented in a decomposed format, subject to the application of word-formation rules. As has been clear throughout the chapter, this conclusion is of primary concern for models of lexical representation and organization, without prejudging how lexical access will proceed, in terms of the relevant access units and (prelexical) segmentation strategies. In principle, our results are neutral about the role of morphology in these prelexical stages; only optimality considerations, together with the accumulated evidence since the seminal work by Taft and Forster (1975, 1976), make more plausible the characterization of those units and strategies in morphological terms. Limiting ourselves to representational issues at the lexical level, the main implications of the work discussed here are straightforward. All the evidence points towards the existence of a morphological level of representation for the individuation and organization of the entries in the mental lexicon. This may be the proper and genuine level of lexical representation, with a convenient intermediate degree of abstraction to serve as the interface between the information relative to the phonological/orthographic form and the more abstract meaning information. This morph-like level of representation reflects also the form of the words, but at that more abstract level that enables the connection with meaning that is susceptible to entering the computational (syntactic) machinery for sentence processing. In sum,

Spanish Word Form and Meaning

203

this morphological level of lexical representation is the best suited to capture the confluence of the different types of linguistic information that is included (and that plausibily interact) in each lexical entry. Furthermore, the appeal of the proposal, as developed in this chapter, is reinforced by its utility to connect models of the bilingual lexicon with more general models of lexical representation (mainly conceived under a monolingual perspective). Research on the bilingual lexicon has been persistently guided by the distinction between conceptual and lexical levels of representation, tending to characterize the latter in pure orthographic-phonological terms (from Potter et al.'s, [1984], models up to the more recent proposals, such as Beauvillain, 1992, or Kroll, 1993). The main implication of our results to this respect is that a morphlike characterization of the lexical level will provide a better understanding of the reported cross-language effects. We want to conclude by briefly mentioning the prospects for further research along the lines sketched in this chapter. There are several distinctions within the morphological level that deserve more detailed analysis and experimentation, mostly the ones related to the transparency, both of form and meaning, in the morphological relations between words. In close association with the transparency parameters, other distinctions will be also in need of clarification, such as regular versus irregular morphology, inflections versus derivations, prefixation versus suffixation, and so on. At the end, and once the morphological level is better established, an important question will remain: Where, in the language module, are the levels of "pure" form and "pure" meaning to be located?

ACKNOWLEDGMENTS Research presented in this chapter has been supported, in part, by the project grant number PB93-0363 from the DGICYT (Ministry of Education and Culture of Spain). The authors want to thank Toni Valero, Pep Demestre, Yolanda Almagro, and Silvia Padro for their assistance in the experiments.

REFERENCES Aronoff, M. (1994). Morphology by itself. Cambridge, MA: MIT Press. Balota, D. A. (1983). Automatic semantic activation and episodic memory encoding. Journal of Verbal Learning and Verbal Behavior, 22, 88-104. Beauvillain, C. (1992). Orthographic and lexical constraints in bilingual word recognition. In R. J. Harris (Ed.), Cognitive processing in bilinguals. Amsterdam: Elsevier.

204

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

Beauvillain, C., and Segui, J. (1992). Representation and processing of morphological information. In R. Frost and L. Katz (Eds.), Orthography, phonology, morphology and meaning (pp. 377-388). Amsterdam: Elsevier. Bentin, S., and Feldman, L. B. (1990). The contribution of morphological and semantic relatedness to repetition priming at short and long lags: Evidence from Hebrew. Quarterly Journal of Experimental Psychology, 42A, 693-711. Bresnan, J. (Ed.) (1982). The mental representation of grammatical relations. Cambridge, MA: MIT Press. Chialant, D., and Caramazza, A. (1995). Where is morphology and how is it processed? The case of written word recognition. In L. B. Feldman (Ed.), Morphological aspects of language processing. Hillsdale, NJ: Erlbaum. Chomsky, N. (1981). Lectures on government and binding. Foris: Dordrecht. Chomsky, N. (1995). The minimalist program. Cambridge, MA: MIT Press. Clifton, C., Frazier, L., and Rayner, K. (1994). Perspectives on sentence processing. Hillsdale, NJ: Erlbaum. Cristoffanini, P. M., Kirsner, K., and Milech, D. (1986). Bilingual lexical representation: The status of the Spanish-English cognates. Quarterly Journal of Experimental Psychology, 38, 367-393. Davis, C. W. (1990). Masked repetition priming. Unpublished doctoral dissertation, Monash University. Davis, C. W., Sanchez-Casas, R. M., and Garcia-Albea, J. E. (1991). Bilingual lexical representation as revealed using a masked priming procedure. Unpublished manuscript. De Groot, A. M. B. (1983). The range of semantic spreading activation in word priming. Journal of Verbal Learning and Verbal Behavior, 22, 417-436. De Groot, A. M. B. (1990). The locus of the associative-priming effect in the mental lexicon. In D. A. Balota, G. B. Flores d'Arcais, and K. Rayner (Eds.), Comprehension processes in reading (pp. 101-122). Hillsdale, NJ: Erlbaum. De Groot, A. M. B. (1992). Bilingual lexical representation: A closer look at conceptual representations. In R. Frost and L. Katz (Eds.), Orthograpy, phonology, morphology and meaning (pp. 389-412). Amsterdam: Elsevier. De Groot, A. M. B. (1993). Word type effects in bilingual processing tasks: Support for a mixed representational system. En R. Schreuder and B. Weltens. The bilingual lexicon (pp. 191-214). Amsterdam: John Benjamins. De Groot, A. M. B. and Nas, G. L. J. (1991). Lexical representations of cognate and noncognates in compound bilinguals. Journal of Memory and Language, 30, 90-123. Del Viso, S., and Garcia-Albea, J. E. (1985). Variables morfblogicas en el procesamiento del lenguaje. Estudios de Psicologia, 19-20, 195-214. Dominguez, A., Cuetos, F., and Segui, J. (1997). Distintividad lexica: el caso de la morfologfa flexiva de genero. Paper presented at the III Symposium de Psicolinguistica, Abril. Oviedo (Spain). Drews, E., and Zwitserlood, P. (1995). Morphological and orthographic similarity in visual word recognition. Journal of Experimental Psychology: Human Perception and Performance, 21 (5), 1098-1116. Durante, R., and Hirshman, H. (1994). Retrospective priming and masked semantic priming: the interfering effects of prime activation. Journal of Memory and Language, 33, 112-127.

Spanish Word Form and Meaning

205

Elias-Cintron, R. (1995). Towards a general theory of agreement: A psycholinguistic study of Spanish gender. Unpublished doctoral dissertation, Cornell University. Feldman, L. B. (1992). Morphological relationships revealed through the repetition priming task. In M. Noonan, P. Downing, and S. Lima (Eds.), Linguistics and literacy (pp. 239-254). Amsterdam: John Benjamins Pub Co. Feldman, L. B. (Ed.) (1995). Morphological aspects of language processing. Hillsdale, NJ: Erlbaum. Feldman, L. B., and Andjelkovic, D. (1992). Morphological analysis in word recognition. In R. Frost and L. Katz (Eds.), Orthography, phonology, morphology and meaning (pp. 343-360). Amsterdam: Elsevier. Feldman, L. B., and Moskovljevic, J. (1987). Repetition priming is not purely episodic in origin. Journal of Experimental Psychology: Learning, Memory and Cognition, 13, 573-581. Feustel, T. C, Shiffrin, R., and Salasoo, A. (1983). Episodic and lexical contributions to the repetition effect in word identification. Journal of Experimental Psychology: General, 112,309-346. Forster, K. I. (1987). Form-priming with masked primes: The best-match hypothesis. In M. Coltheart (Ed.), Attention and Performance XII. London: Erlbaum. Forster, K. I., and Davis, C. (1984). Repetition priming and frequency attenuation in lexical access. Journal of Experimental Psychology: Learning, Memory and Cognition, 10, 680-698. Forster, K. I., Davis, C., Schoknecht, C., and Carter, R. (1987). Masked priming with graphemically related forms: Repetition or partial activation? Quarterly Journal of Experimental Psychology, 39A, 211-251. Fowler, C. A., Napps, S. E., and Feldman, L. B. (1985). Relations among regular and irregular morphologically related words in the lexicon as revealed by repetition priming. Memory and Cognition, 13, 241-255. Fowler, C. A., Wolford, G., Slade, R., and Tassinary, L. (1981). Lexical access with and without awareness. Journal of Experimental Psychology: General, 110, 341-362. Frost, R., and Katz, L. (Eds.) (1992). Orthography, phonology, morphology and meaning. Amsterdam: Elsevier. Garcia-Albea, J. E., Sanchez-Casas, R. M., and Valero, T. (1996). Form and meaning contribution to word recognition in Catalan-Spanish bilinguals. Paper presented at the Ninth Conference of the European Society for Cognitive Psychology, September, University of Wiirzburg. Gerard, L. D., and Scarborough, D. L. (1989). Language-specific lexical access of homographs by bilinguals. Journal of Experimental Psychology: Learning, Memory and Cognition, 15,305-315. Grainger, J. (1990). Word frequency and neighborhood frequency effects in lexical decision and naming. Journal of Memory and Language, 29, 228-244. Grainger, J., Cole, P., and Segui, J. (1991). Masked morphological priming in visual word recognition. Journal of Memory and Language, 30, 370-384. Grainger, J., O'Regan, J. K., Jacobs, A. H., and Segui, J. (1989). On the role of competing word units in visual word recognition. Perception and Psychophysics, 45, 189195. Grainger, J., O'Regan, J. K., Jacobs, A. H., and Segui, J. (1992). Neighborhood frequency

206

J. E. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

effects and letter visibility in visual word recognition. Perception and Psychophysics, 51,49-56. Grainger, G., and Segui, J. (1990). Neighbourhood frequency effects in visual word recognition: A comparison of lexical decision and masked identification latencies. Perception and Psychophysics, 47, 191-198. Grimshaw, J. (1990). Argument structure. Cambridge, MA: MIT Press. Halle, M., and Marantz, A. (1993). Distributed morphology and the pieces of inflection. In K. Hale and S. J. Keyser (Eds.), The view from building 20. Cambridge, MA: MIT Press. Henderson, L. (1985). Towards a psychology of morphemes. In A. Ellis (Ed.), Progress in the psychology of language (Vol. 1, pp. 15-72). London: Erlbaum. Henderson, L. (1989). On mental representation of morphology and its diagnosis by visual access speed. In Marslen-Wilson (Ed.), Lexical representation and process (pp. 357391). Cambridge, MA: MIT Press. Kemply, S. T., and Morton, J. (1982). The effects of priming with regularly and irregularly related words in auditory word recognition. British Journal of Psychology, 73, 441-445. Kirsner, K. (1986). Lexical representation: Is a bilingual account necessary? In J. Vaid (Ed.), Language processing in bilinguals: Psycholinguistic and neuropsychological perspectives (pp. 21-45). Hillsdale, NJ: Erlbaum. Kirsner, K., Brown, H. L., Abrol, S., Chadha, A., and Sharma, N. K. (1980). Bilingualism and lexical representation. Quarterly Journal of Experimental Psychology, 32, 565-574. Kirsner, K., Lalor, E., and Hird, K. (1993). Exercise, meaning and morphology. In R. Schreuder and B. Weltens (Eds.), The bilingual lexicon (pp. 215-248). Amsterdam: John Benjamins. Kirsner, K., Smith, M. C., Lockhart, R. S., King, M-L., and Jain, M. (1984). The bilingual lexicon: Language-specific effects in an integrated network. Journal of Verbal learning and Verbal Behavior, 23, 519-539. Kroll, J. (1993). Accessing conceptual representations for words in a second language. In R. Schreuder and B. Weltens (Eds.), The bilingual lexicon (pp. 53-81). Amsterdam: John Benjamins. Kroll, J., and Sholl, A. (1992). Lexical and conceptual memory in fluent and nonfluent bilinguals. In R. J. Harris (Ed.), Cognitive Processing in bilinguals (pp. 191-204). North-Holland: Elsevier. Lieber, R. (1992). Deconstructing morphology. Chicago: University of Chicago Press. Marcel, A. J. (1983). Conscious and unconscious perception: Experiments on visual masking and word recognition. Cognitive Psychology, 15,197-237. Marslen-Wilson, W. (Ed.) (1989). Lexical representation and process. Cambridge, MA: MIT Press. McClelland, J. L., and Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. Marslen-Wilson, W, Tyler, L., Waksler, R., and Older, L. (1994). Morphology and meaning in the English mental lexicon. Psychological Review, 101, 3-33.

Spanish Word Form and Meaning

207

Mildred, H. V. (1986). Masked priming effects between and within languages. Unpublished honours thesis, Monash University. Napps, S. E. (1989). Morphemic relationships in the lexicon: Are they distinct from semantic and formal relationships? Memory and Cognition, 17, 729-739. Napps, S.E., and Fowler, C.A. (1987). Formal relationships among words and the organization of the mental lexicon. Journal of Psycholinguistic Research, 16, 257-272. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner and G. Humphreys (Eds.), Basic process in reading: Visual word recognition (pp. 264-336). Hillsdale, NJ: Erlbaum. Oliphant, G. (1983). Repetition and recency effect in word recognition. Australian Journal of Psychology, 35, 393-403. Potter, M. C., So, K.-R, von Eckardt, B., and Feldman, L. B. (1984). Lexical and conceptual representations in beginning and proficient bilinguals. Journal of Verbal Learning and Verbal Behavior, 23, 23-38. Pustejovsky, J. (1995). The generative lexicon. Cambridge, MA: MIT Press. Rumelhart, D. E., and McClelland, J. L. (1986). Parallel distributed processing: explorations in the microstructure of cognition, Vol. 1: Foundations. Cambridge, MA: MIT Press. Sanchez-Casas, R. M. (1988). Access representation in visual word recognition. Unpublished doctoral dissertation, Monash University. Sanchez-Casas, R. M. (1995). Priming de froma en el reconocimiento de palabras relacionadas morfologicamente. Paper presented at the II Simposium de Psicolinguistica, Abril. Tarragona (Spain). Sanchez-Casas, R. M. (1996). Lexical access in visual word recognition: The contribution of word form. In M. Carreiras, J. E. Garcfa-Albea, and N. Sebastian-Galles (Eds.), Language processing in Spanish. Mahwah, NJ: Erlbaum. Sanchez-Casas, R. M. (1997). La organizacion y la representacion lexica en el hablante bilingue: datos del castellano, ingles y Catalan. Paper presented to III Simposium de Psicolingiifstica, Abril. Oviedo (Spain) Sanchez-Casas, R. M., Davis, C. W., and Garcia-Albea, J. E. (1992). Bilingual lexical processing: exploring the cognate/noncognate distinct. European Journal of Cognitive Psychology, 4 (4), 293-310. Sanchez-Casas, R. M., Igoa, J. M., and Suarez-Buratti, B. (1992). Are bilingual lexical representations interconnected? Paper presented at the Fifth Conference of the European Society for Cognitive Psychology. September, Paris. Sandra, D. (1990). Representational and processing aspects of compound words in visual word recognition: An experimental approach and methodological appraisal. Unpublished doctoral dissertation, University of Amsterdam. Sandra, D. (1994). The morphology of the mental lexicon: internal word structure viewed from a psycholinguistic perspective. In D. Sandra and M. Taft (Eds.), Morphological structure, lexical representation and lexical access (pp. 227-269). Hove UK: Erlbaum. Sandra, D., and Taft, M. (Eds.) (1994). Morphological structure, lexical representation and lexical access. Hove UK: Erlbaum.

208

J. £. Garcia-Albea, R. M. Sanchez-Casas, and J. M. Igoa

Scarborough, D. L., Cortese, C., and Scarborough, H. (1977). Frequency and repetition effects in lexical memory. Journal of Experimental Psychology: Human Perception and Performance, 7, 3-12. Scarborough, D. L., Gerard, L., and Cortese, C. (1984). Independence of lexical access in bilingual word recognition. Journal of Verbal Learning and Verbal Behavior, 23, 84-99. Segui, J., and Grainger, G. (1990). Priming word recognition with orthographic neighbours: effects of relative prime-target frequency. Journal of Experimental Psychology: Human, Perception and Performance, 16, 65-76. Seidenberg, M. S. (1987). Sublexical structures in visual word recognition: Access units or orthographic redundacy? In M. Coltheart (Ed.), Attention and Performance XII. London: Erlbaum. Simpson, G. B. (Ed.) (1991). Understanding word and sentence. New York: Elsevier. Stanners, R. R, Neiser, J. J., Hernon, W. P., and Hall, R. (1979). Memory representation for morphologically related words. Journal of Verbal Learning and Verbal Behavior, 18, 399-412. Stolz, J. A., and Feldman, L. B. (1995). The role of orthographic and semantic transparency of the base morpheme in morphological processing. In L. B. Feldman, (Ed.), Morphological aspects of language processing (pp. 109-129). Hillsdale, NJ: Erlbaum. Taft, M. (1985). The decoding of words in lexical access: A review of the morphographic approach. In D. Besner, T. G. Waller, and G. E. MacKinnon (Eds.), Reading research: Advances in theory and practice. New York: Academic Press. Taft, M. (1991). Reading and the mental lexicon. London: Erlbaum. Taft, M., and Forster, K. I. (1975). Lexical storage and retrieval of prefixed words. Journal of Verbal Learning and Verbal Behaviour, 14, 638-647. Taft, M., and Forster, K. I. (1976). Lexical storage and retrieval of polymorphemic and polysyllabic words. Journal of Verbal Learning and Verbal Behavior, 15, 607-620. Varela Ortega, S. (1992). Fundamentos de morfologia. Madrid: Smtesis.

NOTES 1 From here on, we will use root, stem, or even base form indistinctly, when referring to the nuclear part shared by morphologically related words. Otherwise, the context will clearly indicate when the specific meaning of each term would need to be preserved. In this respect, we follow the standard use by which root means the atomic lexical morpheme that stands when all affixes (derivational and inflectional) have been removed; by stem we refer to that, simple or complex, part of the word that remains when inflections are removed; base form is used as a cover term, applied to any of the above, or even to whole words, in a relational sense associated with word formation processes in general. 2 We use the term derivation to mark the contrast with inflections. In Spanish, as in most morphologically complex languages, these are the two main processes of affixation, and jointly with compounding, the main processes of word formation. The distinction between inflection and derivation has been usually based on external criteria such as regularity,

Spanish Word Form and Meaning

209

meaning predictibility, productivity, changes in grammatical category, or the relevance for syntactic operations. On internal grounds (i.e., attending to the structure of the word), it is not always easy to find a clear demarcation. However, we took this as an advantage to select very well-matched materials for our experiments. In fact, inflected and derived pairs have the same apparent internal structure (root + gender suffix); the important difference between them is that, in the case of derivations, that gender mark really does not operate as an inflectional suffix (no gender variation at any level), but as a derivational one (words of each pair are derived from the same root). As Varela Ortega (1992) indicates, these are cases that illustrate how an inflectional paradigm can be used to serve a derivational process.

This page intentionally left blank

ACCOUNTING FOR CROSSLINGUISTIC VARIATION: A CONSTRAINT-BASED PERSPECTIVE ROBERT THORNTON MARIELA GIL MARYELLEN C. MACDONALD Department of Psychology University of Southern California Los Angeles, California

1. INTRODUCTION Comrie (1989) makes a distinction between two different approaches to crosslinguistic research. The first approach follows from the work of Greenberg, who attempted to establish descriptive grammatical universals by examining a wide array of languages. These universals are stated in terms of concrete properties of language, for example "all languages have pronominal categories involving at least three persons and two numbers" (Greenberg, 1963, p. 96). According to this approach, the job of the researcher is to compile a descriptive, theory-independent database of a representative sample of the world's languages and then to formulate concrete universals on the basis of general tendencies among the languages. Alternatively, Chomsky's early work stressed the need to look at a few languages in great depth. While acknowledging "a great diversity in the surface structure of languages" (Chomsky, 1965, p. 118), this approach focuses on universals that are stated in terms of fairly complex abstract knowledge about grammatical relations and generalizations. Accordingly, Chomsky (1965, p. 118) notes: There is no reason to expect uniformity of surface structures, and the findings of modern linguistics are thus not inconsistent with the hypotheses of universal Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

211

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

212

R. Thornton, M. Gil, and M. C. MacDonald grammarians. Insofar as attention is restricted to surface structures, the most that can be expected is the discovery of statistical tendencies, such as those presented by Greenberg (1963).

Within this framework, universal grammar is thought to be innate and consequently independent of any particular language. Therefore, the focus of this program of research has been deep analysis of a relatively small number of languages in the hope of discovering the relationship between surface structure and its underlying grammatical basis, rather than broad typology, which would, according to this view, only indicate how languages differ from each other. Although both of these perspectives have stressed the need for linguistic analysis resulting in the formulation of general principles, they differ in two fundamental ways: (a) the nature of the data needed for analysis (i.e., the number of languages studied) and (b) the abstractness of the proposed universals. In the following section, we discuss parallels to this distinction in recent psycholinguistic research and outline a number of current theories of sentence processing. We then present relevant crosslinguistic data on modifier attachment in English and Spanish. These data suggest several conclusions concerning the role of crosslinguistic data in forming theories of sentence processing.

2. PSYCHOLINGUISTIC APPROACHES TO CROSSLINGUISTIC VARIATION 2.1. Garden-Path Theory Principle-based accounts of parsing propose that the assignment of syntactic structure is initially guided solely by abstract structural principles. Although a number of researchers have proposed such accounts of parsing (Berwick and Weinberg, 1984; Crocker, 1996; Gibson, Pearlmutter, Canseco-Gonzalez, and Hickok, 1996; Gorrell, 1989; Hawkins, 1994; Pritchett, 1988, to name a few), we will focus on the best known of these: garden-path theory (De Vincenzi and Job, 1995; Ferreira and Clifton, 1986; Frazier, 1987; Frazier and Rayner, 1982; Rayner, Carlson, and Frazier, 1983). This theory proposes a parser that constructs a single syntactic representation according to the phrase structure rules of a given language. Given a choice between alternative structures, parsing is initially governed by two general principles: (a) late closure, which directs the parser to attach new material into the phrase currently being processed rather than to a more distant attachment site (Frazier and Rayner, 1982); and (b) minimal attachment, which directs the parser to construct the simplest syntactic structure consistent with the input (Rayner et al., 1983). These principles are motivated by the need to quickly structure information in memory (i.e., a simple syntactic structure should be easier to construct than more complex ones). Frazier (1987) has suggested that because

Constraint-Based Crosslinguistic Variation

213

these abstract principles are nonarbitrary extensions of general processing constraints, they are universal. In this respect, this program of research echoes the Chomskyan program, in that rather than attempting to catalog concrete differences between languages, several general abstract principles are said to govern processing universally. Although late closure and minimal attachment predict no crosslinguistic differences in initial processing, a number of recent studies provide data suggesting that the local attachment preference taken as evidence for late closure does not hold for all languages (see Mitchell and Brysbaert, this volume, for review). Frazier and Clifton (1996, 1997) have recently proposed that late closure and minimal attachment are indeed universal, but apply only to the processing of primary phrases, which are, for the most part, the main subject-predicate and all of its obligatory arguments. A nonprimary phrase is one that cannot, even temporarily, be considered primary (e.g., relative clauses, adjunct predication, conjoined phrases). Nonprimary phrases are handled by construal, a mechanism that associates incoming material into the current thematic processing domain, using a variety of structural and nonstructural information. Thus, crosslinguistic variation in initial interpretation is predicted only in the nonprimary cases, whereas primary phrases should still universally conform to late closure and minimal attachment. In a recent series of questionnaire studies on relative clause (RC) attachment to complex noun phrases (NPs) in English and Spanish, Gilboy, Sopena, Clifton, and Frazier (1995) presented data in support of construal. They found that attachment preferences for ambiguous RC modifiers were mediated by nonstructural information, specifically the thematic and referential properties of the NPs. They reported a striking similarity between the languages while finding a large amount of variation within each language. Following Frazier (1990), they proposed a Gricean account of crosslinguistic variation to explain the differences found between the languages, which we will discuss in greater detail in a later section. 2.2. Statistical Tuning In contrast to garden-path theory, which posits universal, abstract parsing principles, the tuning hypothesis (Cuetos, Mitchell, and Corley, 1996; Mitchell, Cuetos, Corley, and Brysbaert, 1995) invokes an experience-based parsing mechanism that is sensitive to information about the frequency distributions of different structural possibilities. On this view, tuning occurs at a purely structural level, on the basis of syntactic category information; all lexical influence is initially ignored. This limitation in sensitivity to statistics concerning only structural information is thought to owe to memory limitations (Mitchell et al., 1995). Accordingly, when comprehenders are faced with a structural ambiguity in their language, the preferred resolution of the ambiguity will have the structure that they have most frequently been exposed to; structural processing preferences are determined by the

214

R. Thornton, M. Gil, and M. C. MacDonald

frequency with which they occur in a particular language. Consequently, tuning predicts that a comprehender's parsing preferences should highly correlate with phrasal-level statistical patterns found in various corpora. Crosslinguistic variation is explained by differing frequency distributions in different languages. This approach most closely echoes the Greenbergian program, in that theorizing occurs at a relatively concrete level of description; no abstract constraints are proposed and initial parsing decisions are based exclusively on a person's exposure to statistical distributions of structural information. As Cuetos et al. (1996) note, although tuning does not propose any specific universal parsing principles, it does place universal constraints on parsing; the suggestion is that what is universal is a frequency-sensitive processing mechanism that is capable of attending to the phrasal-level structural regularities of a particular language, rather than being directed by abstract parsing algorithms. In summary, Tuning makes the relatively straightforward prediction that initial parsing decisions are predictable solely from overall structural patterns in the language. 2.3. Constraint Satisfaction Both the principle-based and statistical accounts outlined so far posit that the processing mechanism initially makes use of a very limited domain of information. For garden-path theory, this information concerns the syntactic category of each word in the input. For tuning, it is information regarding the frequency of occurrence of different structural possibilities. Both models propose that the influence of all other sources of constraint is limited to a secondary reanalysis stage. By contrast, constraint-based models propose that a number of sources of information are initially relevant to sentence comprehension; furthermore, there are no architectural limitations to the use of certain kinds of information, as in the theories discussed above. On this view, some constraints may affect sentence interpretation earlier than others, but these effects may be traced to (a) distributional information, such that more local and frequent information has a larger influence than other constraints, and (b) architectural properties of the cognitive system, such as the nature of activation and competition mechanisms (McRae, SpiveyKnowlton, and Tanenhaus, 1997). Thus the focus in these models has not been on the initial influence of a single source of information; rather, these models have critically examined the integration and time course of a number of constraints that are available early in processing (Boland, Tanenhaus, and Garnsey, 1990; Garnsey, Pearlmutter, Myers, and Lotocky, 1997; MacDonald, Pearlmutter, and Seidenberg, 1994; Spivey-Knowlton and Sedivy, 1995; Trueswell, Tanenhaus, and Garnsey, 1994). Although much of the recent constraint-based work has focused on the influence of lexical information on syntactic ambiguity resolution (MacDonald, 1994; Trueswell, 1996), these models in fact propose three broad levels of constraint

Constraint-Based Crosslinguistic Variation

215

(MacDonald et al., 1994; Spivey-Knowlton and Sedivy, 1995): (a) Lexical: in understanding sentences, comprehenders must recognize individual words, making fairly detailed semantic and frequency information available from the individual lexical items (Garnsey et al., 1997; MacDonald et al., 1994; Trueswell et al., 1994). (b) Structural: word order and processing constraints result from the serial order of language (see Hawkins, 1994; Weckerly and Elman, 1992, for examples). Additionally, the processing mechanism is sensitive to a number of types of contingent frequency information (Juliano and Tanenhaus, 1994). (c) Pragmatic/discourse: as each word is being integrated with the rest of the input, a higher levels discourse representation of the utterance begins to form, which in turn acts as a top-down source of constraint on processing (Altmann and Steedman, 1988; Spivey-Knowlton and Sedivy, 1995). In other words, like tuning, these models predict that the processing mechanism is sensitive to the contingent frequencies of different structural alternatives in the input (Juliano and Tanenhaus, 1994), and as a result ambiguity resolution may differ crosslinguistically. Unlike tuning, however, these models predict that the frequencies of structural patterns are only one of a number of sources of soft constraint on initial processing; processing is mediated by lexical and pragmatic/ discourse information as well. In the next section, we discuss cases in which crosslinguistic variation is constrained by a broad pragmatic constraint: the felicity of NP modification.

3. PRAGMATIC CONSTRAINT ON CROSSLINGUISTIC VARIATION Much of the recent crosslinguistic research on sentence processing has focused on word order, specifically on modifier attachment to complex NPs. Part of the reason that this has been such an active area of research is that although a number of languages share these types of ambiguities, different languages seem to resolve them in different ways (see Mitchell and Brysbaert, this volume, for review). For example, Cuetos and Mitchell (1988), among others, have found that for sentences like (1), Spanish speakers preferred to attach the RC to the more distant of two NPs, such that subjects interpreted the daughter, rather than the colonel, as having the accident. Given comparable English materials, however, English speakers preferred to attach the RC to the local NP, with the interpretation that it was the colonel who had the accident. (1) La hija del coronel que tuvo el accidente The daughter of the colonel who had the accident' Garden-path theory predicts that the initial interpretation of complex NPs is governed by abstract, innate, universal parsing principles. Tuning predicts that

216

R. Thornton, M. Gil, and M. C. MacDonald

initial interpretation is governed simply by exposure to the structural patterns of the language. We, however, propose a constraint-based model of crosslinguistic variation in which exposure to the statistical distributions (at structural and other levels) of one's language interacts with more general pragmatic constraints to guide initial comprehension. Accordingly, there may be differences between languages in initial attachment preference owing to different distributional properties, but any such differences should be mediated by other sources of constraint that operate in the same manner for all languages. We hypothesize that certain pragmatic constraints hold crosslinguistically and are a major limitation on the degree of crosslinguistic variation that is observed in attachment to complex NPs. We have recently conducted several experiments that illustrate the importance of pragmatic constraint in resolving structural attachment ambiguities (Thornton, MacDonald, and Gil, 1997). In two self-paced reading experiments, we examined modifier attachment to complex NPs in English and Spanish in structures of the form NP1 prep NP2 (prep NP3), where the material in parentheses could modify either NP1 or NP2. We modified a broad pragmatic constraint that we hypothesized to hold crosslinguistically, the felicity of NP modification. Modifiability, in this sense, refers to the amount of information that a comprehender needs to uniquely identify entities in the input. For example, in (2a), the head of NP2 has received strong modification in the sense that a particular hall is identified, whereas in (2b), it has received weak modification in the sense that the focus operator only sets up a contrast set of a number of halls and more modification is needed to uniquely identify a particular hall. (2) a. The computer down my front hall with ... b. The computer down the only hall with... Similarly, the head of NP2 is more strongly modified in (3a) than in (3b) because the use of the possessive pronoun our selects a single, specific referent for red car, as people generally only have a single red car. Although the definite determiner the in (3b) presupposes a unique referent (see Hawkins, 1991; Kadmon, 1990, for discussion of uniqueness), it is not infelicitous to add additional modification in order to select a specific red car from a number of other ones. (3) a. The mechanic by our red car with ... b. The mechanic by the red car with . .. Following work that stresses the importance of this kind of pragmatic constraint on processing (Altmann and Steedman, 1988; Spivey-Knowlton and Sedivy, 1995), we predicted that as the modification of a noun becomes stronger, that is, identifying it more uniquely in the discourse, it is less likely to receive further modification. Neither tuning nor garden-path theory predict a difference between the initial processing of (2a) and (2b) or (3a) and (3b); tuning predicts that pro-

Constraint-Based Crosslinguistic Variation

217

cessing initially occurs without reference to this sort of pragmatic constraint, and garden-path theory predicts that sensitivity to pragmatic factors applies only to the processing of nonprimary phrases, such as relative clauses, but not to primary phrases, such as the prepositional phrase (PP) modifiers used in our experiments. We constructed 32 sentences of the type in (2) and (3) with endings that pragmatically disambiguated the PP as either modifying NP1 or NP2, as in (4). (4) a. NP1 modifying disambiguation: The computer down the only hall with expanded memory, b. NP2 modifying disambiguation: The computer down the only hall with a drinking fountain. We presented these items in counterbalanced fashion to 48 native English speakers in a self-paced reading task. Twenty-eight comparable Spanish items were also presented to 40 native Spanish speakers in the same task. We examined reading times at the disambiguation and subsequent words and found overall crosslinguistic differences in initial attachment preference between the English and Spanish speakers: English speakers exhibited a greater overall preference for local (NP2) attachment than Spanish speakers did, replicating previous findings (see Mitchell and Brysbaert, this volume, for review). In both languages, however, subjects' preferences were mediated by the modifiability of NP2. Specifically, when modification to the head of NP2 was weak, English subjects' reading times were longer at a disambiguation phrase that promoted NP1 modification than for one that promoted NP2 modification, whereas Spanish subjects showed no significant differences across disambiguation type. In other words, in the weak modification condition, which promotes NP2 attachment, English speakers exhibited an NP2 attachment preference, while Spanish speakers exhibited no clear attachment preference. When the head of NP2 had received strong modification, promoting NP1 attachment, however, English subjects showed no significant differences across disambiguation type, whereas Spanish subjects' reading times were significantly longer for the NP2 attachment disambiguation than for the NP1 attachment disambiguation. This pattern of reading times demonstrates the very early effect of this pragmatic constraint for both languages: As modification to the head of NP2 becomes stronger, additional NP2 modification becomes infelicitous, and attachment to NP1 becomes more likely. Thus, we can see that this type of pragmatic information, NP modifiability, constrains the range of crosslinguistic variation insofar as it has a similar effect on attachment preferences in both English and Spanish. Modifiability, however, does not explain the overall stronger tendency for local (NP2) modification in English than in Spanish. We hypothesize that the overall difference in modification preferences emerges from variations in the use of alternative syntactic structures in English and Spanish, which we discuss in the next section.

218

R. Thornton, M. Gil, and M. C. MacDonald

4. THE CASE OF WORD ORDER AND GENITIVES In some respects, our explanation of the overall differences between the English and Spanish data is similar to the Gricean argument put forth by Frazier (1990) and elaborated by Frazier and Clifton (1996). There are a number of constructions that have acceptable alternatives in English but not in Spanish. For example, English has two types of genitives: prepositional (e.g., the dress of the singer) and prenominal (e.g., the singer's dress), whereas Spanish (and romance languages in general) has only prepositional genitives (e.g., el vestido de la cantante). Part of our explanation for the overall preference for local (NP2) modification in English relative to Spanish stems from the fact that a number of different structural options for noun modification are frequently used in English, whereas only one is available in Spanish for the structures in question. We argue that the availability or nonavailability of alternative structures constrains interpretation of ambiguous modifiers. Consider first the situation in English, with its two alternative modification forms. The choice of one versus the other form can have a substantial effect on what kinds of postnominal modification are grammatical. For example, if a speaker chooses a prenominal genitive such as the singer's dress, further postnominal modification of dress is allowable, as in the singer's dress with gold sequins, but further modification of singer is not permitted (e.g., *the singer's dress with a lovely voice). In order to further modify the singer, the prepositional genitive must be used (e.g., the dress of the singer with the lovely voice). Thus use of the prepositional genitive is obligatory when NP2 modification is desired, and we suggest that these grammatical constraints create frequency asymmetries in English, such that there is frequent use of the prepositional genitive (i.e., NP1 prep NP2) with NP2 modification in English. We suggest that the resulting frequency distributions of form-meaning pairs, which can be traced to these grammatical constraints, are a major factor in the general NP2 modification preference that is observed in English for NP1 prep NP2 structures. By contrast, Spanish has no alternative prenominal genitive, so the NP1 prep NP2 structure must be used in all cases, no matter what additional modification might be desired. Thus there is no strong association between a particular genitive form and particular modification in Spanish, and therefore no overall NP2 modification preference for NP1 prep NP2 structures in Spanish. Interestingly, there are a set of relations for which the prenominal genitive is not allowed in English, and the prepositional NP1 prep NP2 form is required. In PPs in which the preposition conveys a thematic role, a prenominal form is not a grammatical option. This pattern is illustrated in (5); in attempting to express the prepositional modifier in (5a) as a prenominal modifier in either (5b) or (5c), all semantic and thematic content of the preposition near is lost.

Constraint-Based Crosslinguistic Variation

219

(5) a. the farmer near the barn b. *the barn's farmer c. *the barn farmer In other words, locative relationships as in (5) and other relationships in which the preposition conveys a particular thematic role must always be expressed in the prepositional structure in English, regardless of what kind of further modification might be needed. Thus there is likely to be no strong association between the use of the prepositional structure and a particular modification site. The locative case in English therefore mirrors the situation with all prepositional phrases in Spanish. By contrast, possessive relationships in English, in which the preposition of conveys no thematic role, may be expressed in two structures in English, but the use of these structures is limited by the type of modification that the speaker intends, resulting in large frequency asymmetries in the pairings of particular structures with modifications. We have predicted a relationship between the availability of alternative structures and interpretation preferences of ambiguous modifiers crosslinguistically, and the same availability predictions should hold within a language as well. Specifically, for English, modification of NP2 in a NP1 prep NP2 sequence should be strongly preferred when the sequence conveys a possessive relation, as the NP1 of NP2 (NP2 modifier) sequence is very frequent in English, owing to constraints on the use of the prenominal genitive alternative structure. For locative relationships, however, for which the same structure must be used independent of modification, we predict that there will not be a strong preference for NP1 or NP2 modification. We have recently collected off-line data that test this prediction. English speakers wrote completions for complex NPs with the structure NP1 prep NP2 PP or NP1 prep NP2 RC. We manipulated the type of complex NP so that it expressed either a possessive relationship or a locative relationship. We also manipulated whether the last word of the fragments introduced a relative clause (either that or who) or a PP (with). We predicted that subjects will prefer low-attachment completions for the possessive NPs, which can be expressed with an alternative structure, than for the locative ones, which cannot. We further predicted that the availability of this type of information will affect the attachment of RC and PP modifiers in a similar manner. By contrast, the construal hypothesis (Frazier and Clifton, 1996) makes very different predictions for interpretation of these items. Because PPs are possible primary relations, the attachment of the PP items should be governed by late closure, leading to an NP2 interpretation preference for all items, independent of their possessive or locative status. Relative clauses are nonprimary phrases, however, so interpretation of these items should be governed by construal. Thus the RCs are predicted to be associated into the current thematic processing domain, which

220

R. Thornton, M. Gil, and M. C. MacDonald

Frazier and Clifton (1996) define as the maximal projection of the last theta assigner. Because the prepositions in the locative items are theta assigners, only one site (NP2) is open for modification, whereas both NP1 and NP2 are open for the possessive items, as of only assigns case, and not a thematic role. As a result, construal predicts that for RC attachment, NP2 modification should be strongly preferred for the locative items and less so for the possessive items. 4.1. Method 4.1.1. SUBJECTS Sixteen University of Southern California undergraduates were paid for participation. All were native English speakers. 4.1.2. MATERIALS Twenty-four ambiguous sentence fragments of the structure NP1 prep NP2 with/who/that were used. There were two experimental manipulations. First, the relationship between NP1 and NP2 was either locative or possessive. For the locative condition, all NPs were inanimate. Three items were constructed for each of the four locative prepositions used: behind, beneath, beside, and near, for a total of twelve locative items. For the possessive condition, all NPs were animate. Six items each were taken from the kinship and functional items (types B1 and B2, respectively) used by Gilboy et al. (1995), for a total of twelve possessive items. Second, the last word of the fragment either introduced a PP (with, in all cases) or an RC (who for the 12 animate items and that for the 12 inanimate items). See (6a) for example of a possessive-relation NP and (6b) for an example of a locativerelation one. (6) a. The assistant of the inspector with/who b. The table near the window with/that 4.1.3. PROCEDURE Stimuli were presented on paper in random order. Two presentation lists were created containing an equal number of items from each condition, with each item appearing only once in each list. Two practice items of the same structure were composed. Subjects were asked to complete the sentence fragments with the first thought that came to mind. Following completion of the questionnaire, the experimenter, who was blind to the hypotheses, reviewed the questionnaire with the subject and asked, for each item, which person or object was modified by the subject's written completion. The entire procedure took about 10 minutes.

Constraint-Based Crosslinguistic Variation .

221

4.2. Results and Discussion The sentence completion data are presented in Figure 1. Overall, 68.75% of the completions indicated attachments to NP2, consistent with a number of other studies that found a local attachment preference in English (Cuetos and Mitchell, 1988; Gibson et al., 1996). There was a main effect of the type of relationship between NP1 and NP2, Fl(l, 15) = 20.38; p < .001;F2(1,22) = 22.63; p < .001, such that subjects were more likely to modify NP2 for the possessive items than for the locative items. There was no main effect of the type of modifier, and the interaction of these variables was not significant, all Fs < 1. These data support both of our predictions and run counter to the construal hypothesis in several important ways. First, the possessive items, which have an acceptable alternative form, yielded more NP2 completions than did the locative items, which have no acceptable alternative forms. This result demonstrates that information about the availability of alternative structures has a robust effect on modifier attachment, illustrating the important role this information plays in sentence processing. Although the survey was conducted only in English, the results suggest that differences in overall attachment preference between the languages may owe to the availability of alternative prenominal modifier structures in English that are absent in Spanish. Gilboy et al. (1995) also provide data in support of this claim. They found a stronger local attachment preference in English than in Spanish for items that have alternative structure in English, but not in Spanish (cf. their animate possessive items). For thematic items, for which there are no alternatives in either language, the attachment preferences were much closer. Although they attributed the results to construal, the present experiment suggests

Figure 1. Percentage NP2 attachment to complex NPs containing locative or possessive modification.

222

R. Thornton, M. Gil, and M. C. MacDonald

that their results, at least in part, may be due to the availability of alternative structures, which we have shown has an effect on this type of survey data. Second, this pattern of results held equally for PP and RC items. This finding is inconsistent with construal, which predicts a difference in attachment preference for these items, as it makes a distinction between the attachment of primary and nonprimary phrases. Further, construal makes the wrong predictions for either type of modifier. As construal does not apply to the processing of primary phrases, garden-path theory asserts that our PP items should be attached according to late closure, predicting no differences between our possessive and locative items, which was not the case. Construal also asserts that our RC items should be associated into the current thematic processing domain, predicting the opposite pattern of results: the locative items should have shown more NP2 attachment, as the current thematic processing domain is only NP2, than the possessive items, for which both sites are open. In summary, the findings of the present experiment, along with the self-paced reading data presented above, suggest that pragmatic information, whether it be felicity of NP modification or availability of alternative structures, exerts a strong influence in constraining crosslinguistic variation and on processing in general. It is also important to note that our interpretation of the data differs from a strict Gricean interpretation in two important ways. First, it is a probabilistic constraint. We do not predict that the mere existence of options is what has an effect; it is their frequency of usage that is important. Second, it is only one of a number of soft constraints. As our self-paced reading data demonstrates, crosslinguistic differences are mediated by the pragmatic context in which they occur.

5. CONCLUSIONS Green (1990), in addressing controversy related to the universality of Gricean principles, makes an important point about crosslinguistic variation and pragmatics in general. She argues that not only is pragmatic constraint on meaning not specific to particular languages, it is not specific to language at all. As such, crosslinguistic variation should not be random; rather, it should follow predictably from nonlinguistic principles. Along these lines, the purpose of this chapter has been to explore the predictable role of pragmatic information in constraining and explaining crosslinguistic differences in sentence processing. The data that we have presented make an important point: within each language we see the effect of pragmatic constraint, which is useful in explaining differences between languages. Simply noting that, apart from this manipulation, Spanish readers are more likely than English ones to modify more distant NPs is interesting, but it would miss the point of why this occurs.

Constraint-Based Crosslinguistic Variation

223

Constraint-based accounts of the sort that we have suggested have been criticized for simply observing ad hoc correlations between frequency and parsing preference (Frazier, 1995). To return to Connie's distinction, Frazier's critique casts constraint satisfaction as a completely Greenbergian enterprise, merely cataloging statistical patterns without a true theory. On the contrary, as we hope to have illustrated in this chapter, the goal of our enterprise is not simply to characterize the statistical properties of a number of languages; we propose that frequency sensitivity emerges from general processing constraints, and is constrained by various kinds of pragmatic information. In our view, both crosslinguistic similarity and variation emerge from the interaction of the speaker's environment and specific pragmatic properties of the cognitive system.

ACKNOWLEDGMENTS This research was supported by National Science Foundation grant SBR-9511270 to the third author. We thank Jared Layport for his assistance in running subjects.

REFERENCES Altmann, G. T. M., and Steedman, M. (1988). Interaction with context during human sentence processing. Cognition, 30, 191-238. Berwick, R. C., and Weinberg, A. (1984). The grammatical basis of linguistic performance. Cambridge, MA: MIT Press. Boland, J. E., Tanenhaus, M. K., and Garnsey, S. M. (1990). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language, 29, 413-432. Chomsky, N. (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. Comrie, B. (1989). Language universals and linguistic typology (2nd ed.). Chicago: University of Chicago Press. Crocker, M. (1996). Computational psycholinguistics: An interdisciplinary approach to the study of language. Dordrecht, The Netherlands: Kluwer Academic Publishers. Cuetos, F, and Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the use of the Late Closure strategy in Spanish. Cognition, 30,73-105. Cuetos, F, Mitchell, D. C., and Corley, M. M. B. (1996). Parsing in different languages. In M. Carreiras, J. Garcia-Albea, and N. Sabastian-Galles (Eds.), Language processing in Spanish (pp. 145-187). Mahwah, NJ: Lawrence Erlbaum Associates. De Vincenzi, M., and Job, R. (1995). An investigation of late closure: The role of syntax, thematic structure, and pragmatics in initial and final interpretation. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 1303-1321.

224

R. Thornton, M. Gil, and M. C. MacDonald

Ferreira, F, and Clifton, C., Jr. (1986). The independence of syntactic processing. Journal of Memory and Language, 25, 348-368. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII: The psychology of reading (pp. 559-586). Hillsdale, NJ: Lawrence Erlbaum Associates. Frazier, L. (1990). Parsing modifiers: Special-purpose routines in the human sentence processing mechanism? In D. A. Balota, G. B. Flores d'Arcais, and K. Rayner (Eds.), Comprehension processes in reading (pp. 303-330). Hillsdale, NJ: Lawrence Erlbaum Associates. Frazier, L. (1995). Constraint satisfaction as a theory of sentence processing. Journal of Psycholinguistic Research, 24, 437-468. Frazier, L., and Clifton, C., Jr. (1996). Construal. Cambridge, MA: MIT Press. Frazier, L., and Clifton, C., Jr. (1997). Construal: Overview, motivation and some new evidence. Journal of Psycholinguistic Research, 26, 277-295. Frazier, L., and Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Garnsey, S. M., Pearlmutter, N. I, Myers, E., and Lotocky, M. A. (1997). The relative contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of Memory and Language, 37, 58-93. Gibson, E., Pearlmutter, N. J., Canseco-Gonzalez, E., and Hickok, G. (1996). Recency preference in the human sentence processing mechanism. Cognition, 59, 23-59. Gilboy, E., Sopena, J. M., Clifton, C., Jr., and Frazier, L. (1995). Argument structure and association preferences in Spanish and English complex NPs. Cognition, 54, 131-167. Gorrell, P. (1989). Establishing the loci of serial and parallel effects in syntactic processing. Journal of Psycholinguistic Research, 18, 61-73. Green, G. (1990). The universality of Gricean interpretation. Proceedings of the Berkeley Linguistic Society, 16, 411-428. Greenberg, J. H. (1963). Some universals of grammar with particular reference to the order of meaningful elements. In J. H. Greenberg (Ed.), Universals of language (pp. 5890). Cambridge, MA: MIT Press. Hawkins, J. A. (1991). On (in)definite articles: Implicatures and (un)grammaticality prediction. Journal of Linguistics, 27, 405-442. Hawkins, J. A. (1994). A performance theory of order and constituency. Cambridge, UK: Cambridge University Press. Juliano, C., and Tanenhaus, M. K. (1994). A constraint-based lexicalist account of the subject/object attachment preference. Journal of Psycholinguistic Research, 23,459472. Kadmon, N. (1990). Uniqueness. Linguistics and Philosophy, 13, 273-324. MacDonald, M. C. (1994). Probabilistic constraints and syntactic ambiguity resolution. Language and Cognitive Processes, 9, 157-201. MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676-703. McRae, K., Spivey-Knowlton, M. J., and Tanenhaus, M. K. (1997). Modeling the influence

Constraint-Based Crosslinguistic Variation

225

of thematic fit (and other constraints) in on-line sentence comprehension. Unpublished manuscript. Mitchell, D. C., Cuetos, F, Corley, M. M. B., and Brysbaert, M. (1995). Exposure-based models of human parsing: Evidence for the use of coarse-grained (non-lexical) statistical records. Journal of Psycholinguistic Research, 24,469-488. Pritchett, B. (1988). Garden path phenomena and the grammatical basis of language processing. Language, 64, 539-576. Rayner, K., Carlson, M., and Frazier, L. (1983). The interaction of syntax and semantics during sentence processing. Journal of Verbal Learning and Verbal Behavior, 22, 358-374. Spivey-Knowlton, M. J, and Sedivy, J. C. (1995). Resolving attachment ambiguities with multiple constraints. Cognition, 55, 227-267. Thornton, R., MacDonald, M. C., and Gil, M. (1997). Pragmatics constrain the initial interpretation complex noun phrases in English and Spanish. Unpublished manuscript, University of Southern California. Trueswell, J. C. (1996). The role of lexical frequency in syntactic ambiguity resolution. Journal of Memory and Language, 35, 566-585. Trueswell, J. C., Tanenhaus, M. K., and Garnsey, S. (1994). Semantic influences on parsing: Use of thematic role information in syntactic disambiguation. Journal of Memory and Language, 33, 285-318. Weckerly, J., and Elman, J. L. (1992). A PDP approach to processing center-embedded sentences. In Proceedings of the Fourteenth Annual Conference of the Cognitive Science Society (pp. 414-419). Hillsdale, NJ: Lawrence Erlbaum Associates.

This page intentionally left blank

THE TIME COURSE OF ATTACHMENT DECISIONS: EVIDENCE FROM FRENCH JOEL PYNTE Centre National de la Recherche Scientifique, and University of Provence Aix-en-Provence, France

For more than 15 years, research in the field of human sentence processing has focussed on this sole question: how are syntactic ambiguities dealt with? In this chapter, I am going to summarize a series of recent experiments conducted in French, whose aim was to (modestly) contribute to the debate. Let us first briefly consider the solutions that have been proposed so far: 1. Solution (a): The different possible structures are processed in parallel, together with other sources of information. 2. Solution (b): The choice is posponed until more information is available. 3. Solution (c): The parser forces the decision (e.g., in favor of the simplest structure). Solution (a) has recently received considerable attention in the framework of the constraint satisfaction approach of sentence comprehenion (MacDonald, Pearlmutter, & Seidenberg, 1994). As for solution (b), a formal framework is provided by the Decription Theory (D-theory) developed by Marcus, Hindle, and Fleck (1983) (see also Perfetti's, 1990, pieces parser). Although different on many critical aspects, solutions (a) and (b) share the assumption that syntactic parsing is basically a lexically driven process, in which chunks of syntactic structures, or subtrees, are generated on-line while a sentence is being parsed. Moreover, in both types of models, these chunks are assumed to combine with thematic information (e.g., concerning the possible argument structures and thematic girds associated with each lexical item) in order to form a complete syntactic structure. Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

227

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

228

Joel Pynte

As emphasized by Frazier (1987), solutions (a) and (b) are likely to raise a serious problem in terms of resource requirements (since memory capacity for unorganized materials is known to be limited). By contrast, the memory requirement is minimal in the case of solution (c) since each incoming item is assumed to be immediately integrated in a unique syntactic structure. In Frazier's gardenpath model (Frazier, 1987; Frazier & Rayner, 1982), this is achieved by means of two parsing heuristics: minimal attachment (do not postulate any potentially unnecessary nodes) and late closure (if grammatically permissible, attach new items into the clause or phrase currently being processed). For example, in sentence (1) below, the late-closure principle states that the relative clause "who . . ." will be preferentially attached to the second noun in the "Nl-of-N2" construction (the actress, not her servant, was on the balcony). (1) Someone shot the servant of the actress who was on the balcony with her husband. Although this may be true as far as English is concerned (Cuetos and Mitchell, 1988), the reverse effect has, in fact, been found in a variety of languages ranging from Spanish (Cuetos and Mitchell, 1988) to Dutch (Brysbaert and Mitchell, 1996a). Frazier and Clifton's (1996) answer to these studies was to propose a distinction between two types of attachment procedures, corresponding to two types of syntactic relationships. Although primary relationhips (basically the relations between a main verb an its various arguments within a single clause) would still be submitted to minimal attachment and late closure, the processing of nonprimary relationships, such as relative clauses (RC), would rely on different principles. Instead of being attached in a fully determinate syntactic representation, nonprimary relationships are assumed to be associated with an entire theta domain, until a decision can be made on the basis of semantic and pragmatic factors. In fact, Frazier and Clifton's (1996) proposal amounts to applying solution (b) to a certain class of sentences while maintaining solution (c) for another class. Such a half-way position is somewhat unsatifactory, and one can be tempted to go one step further, by applying solution (b) to primary relationships as well (see the last three sections of this chapter). Another (obvious) solution, if one wishes to maintain a unified view of sentence processing, consists in going along with solution (c) for both primary and nonprimary relationships. This approach, which assumes that all types of attachments are decided at the syntactic level, is explored in section 1.

1. RELATIVE CLAUSE ATTACHMENT This section further examines the nature of the bias (if any) favoring high attachment for sentences like (1) above. According to Frazier and Clifton's (1996)

Time Course of Attachment Decisions

229

Construal theory, the decision to attach high in Mitchell et al.'s experiments is made at the interpretation level and can thus be influenced by semantic and/or pragmatic factors. An interesting proposal, which will not be discussed here, was that conversational maxims (Grice, 1975) could be responsible for some of the effects reported in the crosslinguistic literature. For example, the fact that "the servant of the actress" could have been replaced by "the actress's servant" in sentence (1) (but not in the Spanish counterpart) could explain why English readers, unlike Spanish ones, tend to attach the RC to N2 (but see Brysbaert and Mitchell, 1996a, for a counterargument). Another way to examine whether the Nl preference found by Mitchell et al. is structurally or semantically and/or pragmatically determined consists in manipulating the context in which the experimental sentences are presented. The rationale underlying the eye-tracking experiment reported below is that it should be possible, in this way, to turn the generally preferred Nl attachment into a N2 preference when this is contextually relevant.

1.1. Experiment 1: The Effect of Context A more detailed description of the experiment summarized here can be found in Zagar, Pynte, and Rativeau (1997). The experimental sentences were of the type shown in (2). The sentences were disambiguated by the adjective whose gender agreed either with Nl (forcing high-attachment) or N2 (low attachment), (see also Table 1 where the noun agreeing in gender with the adjective is underlined). Disambiguation could thus, in principle, be performed without the intervention of any high-level interpretation processes. (2)

Un journaliste aborda l'avocat (male N1) de la chanteuse (female N2) qui paraissait plus confiant(e) (masculine or feminine gender) que les autres.

Each sentence was embeded in a short paragraph whose role was to provide a consistent "discourse model" (Garnham, 1983) for the test sentence by introducing a set of potential referents for each noun phrase. Two versions of each paragraph were prepared (Table 1). In the Nl-biased context, the discourse model contained several potential referents for Nl and only one for N2. In the N2-biased version, the context contained several potential referents for both N2 and Nl. The reason for not having a symmetrical context with one Nl and several N2s is that the construction "Nl of N2" is already restrictive. If there were only one referent for Nl, the Nl-of-N2 would sound slightly odd. When there are several Nls and N2s, N2 is still not uniquely specified and can be restricted by the RC that follows. Under the contextual influence hypothesis the RC restricts N2, and the context is consistent with a late-closure reading of the target sentence. Table 1 also presents the two versions (attachment forced in favor of Nl vs. in favor of N2) of the corresponding experimental sentence. Twenty-four such texts were presented (in French) to 24 French native speakers. In one condition of the experiment (the only

230

Joel Pynte TABLE 1 THE FOUR EXPERIMENTAL CONDITIONSa

Nl-biased context (several Nls and one N2): The hearing was about to begin and everyone was waiting for the judge. The audience was chatting noisily and talking about the case. The singer [female N2] and her barristers [male Nl] were standing in a corner of the courtroom.... Nl-disambiguation: A journalist approached the barrister of the singer who seemed more confident [masc.] than the others. (330 ms) N2-disambiguation: A journalist approached the barrister of the singer who seemed more confident [fem.] than the others. (380 ms) N2-biased context (several N1's and several N2 's): The hearing was about to begin and everyone was waiting for the judge. The audience was chatting noisily and talking about the case. The singers [female N2] and their barristers [male Nl] were standing in a corner of the courtroom. Nl-disambiguation: A journalist approached the barrister of the singer who seemed more confident [masc.] than the others. (310 ms) N2-disambiguation: A journalist approached the barrister of the singer who seemed more confident [fem.] than the others. (353 ms) a

Target sentences shown in italic.

one reported here), the subjects had to answer a very precise question after having read each text. In order to track the very moment when disambiguation occurred, a critical region was defined, comprising the nine characters around the last letter in the disambiguating adjective (i.e., four letters on the left and four letters on the right). The total amount of attention received by this region during first-pass reading was calculated and is shown in Table 1 (between brackets). Rayner and Frazier's (1987) definition of first-pass reading time was used. The results indicated that French readers, like Spanish and Dutch ones, prefer to attach the RC to Nl. The difference was of 47 msec in favor of high attachment, Fl(l, 23) = 3.82, p = .06; and F2(l, 23) = 5.34, p = .03 for the main effect of attachment type (type of disambiguation). It is important to note that this effect was obtained on first-pass reading, which suggests that attachment decisions may be very rapid. Moreover, this effect was not qualified by context. Neither the main effect of context, nor the interaction between context and type of attachment was significant (F's close to or below 1). The general pattern of results obtained here seems difficult to explain in the Construal framework, which assumes that RC attachment results from discourse-level influences. In contrast with the Construal theory, Mitchell and Cuetos (1991) maintained that RC attachment is settled on the basis of purely structural information. In order to explain crosslingustic differences, they suggested that the preference for one interpretation over another is the simple result of the respective frequencies of each of the interpretations in the language (the Linguistic-Tuning hypothesis): "The basic proposal is that ambiguities are initially resolved in line with the sta-

Time Course of Attachment Decisions

231

tistical prevalence of the alternative readings in the language as a whole" (Cuetos, Mitchell, & Corley, 1996). Note, however, that the Tuning hypothesis remains relatively neutral concerning the type of mechanisms responsible for such frequency biases (see MacDonald et al., 1994; and the last section of this chapter, for possible accounts in terms of constraint satisfaction and minimal commitment, respectively). In accordance with the Tuning hypothesis, Mitchell, Cuetos, and Corley (1992) reported corpus analysis data showing that in Spanish, 60% of the relative clauses were attached to Nl, whereas in English, only 38% high attachment was observed. To our knowledge, the only available corpus analysis for French (based on 1384 occurrences of "Nl-de-N2-qui") is consistent with the Nl preference found in the Zagar et al. study, since 62% high attachment was observed (Baltazart & Kister, 1995). However, a discrepancy between attachment preference and corpus analysis data has been reported for at least one language, namely Dutch, where a highattachment preference was obtained in eye-movement data and a low-attachment preference was observed in corpus analyses (Brysbaert and Mitchell, 1996b). Before abandoning the Tuning hypothesis, we should consider whether corpus analysis is in fact suitable for assessing the availability of such and such a syntactic construction at a given moment in time. Frequency counts may vary for different sub-corpora (e.g., newspapers articles vs. novels), for differents subgroups of individuals, and even for different types of semantic contents. 1.2. Experiment 2: Syntax Setting Effects Another way to deal with this question consists in artificially manipulating the relative availability of high and low attachments by varying the composition of the experimental list. For example, the bias in favor of Nl-attachment found by Zagar et al. (1997) can be expected to be weakened if, prior to being presented with a sentence like (3), the subject is presented with a great deal of sentences for which attachment to N2 is massively preferred (example 4) (syntax-setting hypothesis). Such an experiment was reported by Pynte and Frenck-Mestre (1996). Eye movements were recorded while French subjects (N = 16) read sentences (N = 16) of the following type: (3)

Yves surprend la boulangere de Marc qui parait plus hargneux(se) que la nuit derniere. 'Yves caught Mark's baker (fem) who seemed crabbier (fem/masc) than last night.'

(4)

Yves surprend la boulangere avec Marc qui parait plus hargneux(se) que la nuit derniere. 'Yves caught the baker (fem) with Mark who seemed crabbier (fem/masc) than last night.'

232

Joel Pynte

Two versions of each sentence were prepared. In one version the gender of the adjective was the same as the gender of Nl. In this case, the adjective forced high attachment of the relative clause (Nl attachment). In the other version the adjective agreed with N2 and was consistent with low attachment (late closure). Sentences were presented in two blocks dependent upon the preposition ('de' vs. 'avec') between Nl and N2. Half of the subjects received the 'de' condition first, and half received the "avec" condition first. A low-attachment preference was obtained for Nl-avec-N2 sentences (shorter reading times on the adjective for N2 attachment whatever the block order). This aspect of the results will not be discussed here (see Frenck-Mestre and Pynte, in press). The reverse effect was obtained for N1-de-N2 sentences. First-pass reading times recorded on the disambiguating adjective were shorter for high attachment in this case (442 ms vs. 483 ms, Fl(l, 8) = 10.07, p < .01; F2(l, 14) = 4.51, p — .052). This effect tended to be qualified by block order however (Fl(l, 8) = 4.09, p < .08; F2(l, 14) =1 .71, p = .021). When subjects were confronted with the 'avec" condition first (for which Nl attachment was often unacceptable), virtually no difference was obtained between Nl and N2 attachments for the 'de' sentences (18 ms, Fl and F2 < 1, in the 'avec" first vs. 83 ms, Fl(l, 4) = 16.53, p = .015; F2(l, 7) = 6.89, p = .034, in the 'de' first condition). These results confirm Zagar et al.'s findings that high attachment is preferred in French for Nl-of-N2-RC constructions. However, when subjects are confronted with Nlwith-N2 sentences immediately prior, the high attachment preference is disturbed for Nl-de-N2 constructions. This gives some support to the notion that the availability of such and such a syntactic construction can indeed influence syntactic decisions. However, the fact that the degree of availability of a given interpretation can apparently be modulated through experimental manipulation suggests that the underlying mechanisms are somewhat open to strategic control. This notion will be further discussed in section 4.

2. ADJECTIVAL PHRASE ATTACHMENT This section further examines Frazier and Clifton's (1996) distinction between primary and nonprimary relationships. Of particular interest, from this point of view, are examples such as (5a) and (5b), where the adjectival phrase (AP) can be analyzed either as a primary phrase (the adjective dubitative is an argument of the verb in example 5a, see Schwarze, 1988) or as a nonprimary phrase (''dubitatif(ve)' is an adjunct predicate in 5b). The meaning of (5a) can be paraphrased by 'when he left his wife she was dubious.' The adjective must agree in gender with the object NP ('sa femme') in this case. By contrast, in (5b), the adjective is a predicate whose subject can be either the subject or the object of

Time Course of Attachment Decisions

233

the preceding clause (Frazier and Clifton, 1996), and gender agreement can occur with either the first or the second NP. Note that the meaning of (5b) (whatever the subject of the adjunct predicate) also includes a temporal aspect, but much less clearly than in (5a). (5) a. Il a laisse sa femme dubitative. 'He left his wife dubious (feminine).' b. Il a laisse sa femme, dubitatif(ve). 'He left his wife, dubious (masc/fem).' How does the reader decide between the two interpretations? Frazier and Clifton's answer was that any potential primary phrase is first analyzed as primary. The mere fact that examples such as (5a) can be found means that any AP in this position is a potential primary phrase and, according to Frazier and Clifton, should thus be attached to the VP, following the minimal attachment principle. Only at the thematic processing level should the (possible) inadequacy of such an attachment be detected, and the adjunct predicate interpretation be considered. In other words, in examples such as (5a) and (5b), the argument interpretation should be tried first, and should thus lead to shorter reading times, when semantically permitted (see Frazier and Clifton, 1996, for a dicussion of this question). Two different hypotheses will be considered here, namely, (a) that the argument interpretation critically depends on the verb's argument structure, and (b) that "extralinguistic" cues such as punctuation (or prosody in the case of an auditory presentation) also exert some influence. Concerning the first point, it should be noted that the argument interpretation is only possible for certain types of verb, a constraint that Schwarze (1988) expressed as follows: "X-complements are possible only if the verb has stored them in its lexical form" (p. 273). According to Schwarze, the final adjectival phrase is an argument in sentence (6) below, while it is not in sentence (7): (6) Nous I'avons repeche vivant. 'We fished him out alive.' (7) ? Nous l'avons regarde vivant. ? 'We looked at him alive.' The difference between these two examples may seem quite tenuous. This merely means that the distinction between primary and nonprimary relationships is not always as clear-cut as suggested by Frazier and Clifton (1996), and that some lexical processing is probably necessary before a given syntactic structure can be chosen. Concerning the second point, it should be noted that in French, the adjunct predicate intrepretation is critically dependent on the presence of a comma (or prosodic break for auditory presentation) in front of the adjective (example 5b).

234

Joel Pynte

In the absence of a comma or prosodic break, the adjective will be interpreted either as a normal noun-modifying adjective (postnominal adjectives are frequent in French), or as an argument of the verb if permitted by the verb's argument structure. The correct interpretation thus seems to be triggered, in French, by "nonlinguitic" cues (such as punctuation or prosody), which suggests that the parsing procedures are somewhat open to strategic control, just as those involved in processing relative clauses (see previous section).

2.1. Experiment 3: The Role of Prosody This prediction was examined in an acceptablity judgement experiment. A trial consisted in the auditory presentation of a short NP1 + V + NP2 clause (synthesized speech, e.g., 'Paul laisse Marie') followed by the visual presentation of an adjective (e.g., 'dubitative'). The subjects (N = 24) were required to decide whether the adjective could be incorporated into the preceding clause in order to form an acceptable utterance by pressing one of two buttons marked "yes" and "no." Twelve different verbs (all permitting a clear temporal interpretation), 12 adjectives (all denoting a nonpermanent state), and 24 first names were used. The adjective agreed in gender with either the first or second NP. Two prosody conditions were compared. In the "strong break" condition, a fall in fundamental frequency (F0) was introduced in NP2's last syllable. In the "weak break" condition, an FO rising was introduced instead (suggesting that the sentence was not complete, see Hirst, 1987). The four experimental conditions are shown in Table 2. Table 2 also contains the mean proportion of yes answers and the corresponding response times (two first columns). The sentence was judged significantly more acceptable when the adjective agreed in gender with NP2, whatever the prosody

TABLE 2 THE FOUR EXPERIMENTAL CONDITIONSa Experiment 3 Responseb

Experiment 4

Latency Latency (ms) Responseb (ms)

Christophe laisse Delphine

. . . dubitative (fem.)

.72

1635

.77

2122

Christophe laisse Delphine

. . . dubitatif(masc.)

.35

1790

.37

2545

Christophe laisse Delphine

. . . dubitative (fem.)

.63

1667

.66

2315

Christophe laisse Delphine

. . . dubitatif(masc.)

.42

1797

.35

2505

a NP1 is masculine and NP2 is feminine here (they were varied in the experiment). The one agreeing with the adjective is underlined. * Yes answer frequency.

Time Course of Attachment Decisions

235

(Fl(l, 16) = 20.89, p < .01 for the main effect of type of agreement). However, the tendency was more marked for the raising intonation than for the falling intonation (Fl(l, 16) = 10.87, p < .01 for the interaction between prosody and agreement). No significant difference was observed on latencies. The overall preference for NP2 agreement observed here is in keeping with Frazier and Clifton's (1996) suggestion that the argument interpretation is preferred whenever semantically acceptable (remember that all the verbs used in the present experiment permitted an argument interpretation). According to Frazier and Clifton the effect merely results from the application of the minimal attachment strategy. It is also possible to argue that the AP was, in fact, immediately appropriated by the thematic processor to fill a role in the verb's theta grid. The fact that the bias in favor of NP2 was higher in the raising intonation condition, as compared to the falling intonation condition, suggests that prosody can interact with the verb's argument structure in determining the correct interpretation. It seems tempting to conclude that the falling intonation was actually perceived as a full intonation boundary by the subjects, which may have blocked the theta-role mechanism so that the following AP could not (immediately) be equated with a role in the verb's theta-grid (see Marcus and Hindle, 1990, for a similar proposal). Note, however, that the proportion of acceptablity judgments did not increase in the falling intonation condition, suggesting that the adjunct predicate interpretation was not immediately triggered by the presence of an intonation boundary. Adjunct predicates are assumed to accept NP1 as well as NP2 as their subject (Frazier and Clifton, 1996), and a 100% acceptability score could thus have been expected in the falling intonation condition. 2.2. Experiment 4: Contextual Influence The aim of this experiment was to examine the nature of the bias favoring the argument interpretation. The stimulus sentence was preceded by a context sentence (visual presentation). Apart from this modification, all other aspects of the experiment were similar to the previous one. The context sentence was aimed at favoring a temporal intrepretation of the test sentence but did not share any semantic relationship with it (Example: "the storm had just stopped" for the previous example). If the preference for NP2 agreement is structurally determined (e.g., due to the operation of the minimal attachment strategy), then no effect of context should be found. By contrast, such an effect can be expected if lexical information is involved (e.g., presence of a specific role in the verb's theta-grids). This prediction derives from the observation that those verbs that accept an argument AP happen to include specific temporal aspects in their meaning. Many of them are also polymorphemic in French (e.g., retrouver, raccompagner, revoir, etc). Interestingly,

236

Joel Pynte

one also finds the temporal morpheme 're' in (6) above (repecher), which is presented as a typical example of argument interpretation by Schwarze (1988). Note that 're' is a pseudomorpheme in ''regarded (example 7), whose theta-grid does not comprise any role for an argument AP. The results are presented in the last two columns of Table 2. A general preference for NP2 agreement was observed, even in the case of the falling intonation. Only the main effect of type of agreement (NP1 vs. NP2) was significant, both for the frequency of yes answers and for latencies (Fl(l, 16) = 30.05, p < .01, and Fl(l, 16) = 9.05, p < .05, respectively. The main effect of prosody approached significance in the frequency analysis (Fl(l, 16) = 3.45). The bias in favor NP2 agreement was thus strengthened by the context. The influence of context was particularly clear in the falling intonation condition, where an argument interpretation was apparently promoted, in spite of an adverse prosodic cue. The effect found here may seem surprising, given the lack of any contextual influence on RC attachment reported in the previous section. Why should the context exert an influence on the processing of APs, and not on RCs? Rather than looking for an answer in the primary versus nonprimary opposition (RCs are clearly "nonprimary"), a distinction between two types of contexts seems to be necessary here. In this last experiment, the context probably exerted its influence in an indirect way, by activating one of the verb's possible argument structures, which, in turn, permitted the AP to be treated as an argument by the thematic processor. The key point is that context is assumed to operate at the lexical level here. The sentential level would only be affected in a secondary step, as a consequence of lexiccal selection (see MacDonald et al., 1994, for a discussion of this notion in the framework of the constraint satisfaction model).

3. PREPOSITIONAL PHRASE ATTACHMENT The possible interaction of prosody and argument structure in attachment decisions is further examined in this section. Let us consider examples (8) and (9) below. From the point of view of a phrase-structure parser, these two sentences are equivalent, with the final prepositional phrase (PP) being attached to the VP in both cases. Note, however, that the PP is an argument of the verb in (8) ('informer' is a di-transitive verb), while it is an adjunct in (9) ('choisit' is monotransitive). In other words, the parser deals with a primary relationship in (8), and with a nonprimary relationship in (9). (8) Les espions informent les gardes du complot. 'The spies inform the guards of the conspiracy.'

Time Course of Attachment Decisions

237

(9) L'etudiant choisit un appartement avec soin. 'The student chooses a flat with care.' This should not have any effect on performance if Frazier and Clifton (1996) are correct (since the PP is a potential primary phrase in both cases). On the other hand, a difference between the two types of verbs can be expected if attachment decisions are influenced by lexical (e.g., thematic) information. The experimental literature seems to favor Frazier and Clifton's position. For example, Clifton, Speer, and Abney (1991) found a general preference for VP-attachment over NPattachment, whatever the status (adjunct vs. argument) of the PP. In this experiment, processing times were found to be shorter for sentences like (8) and (9), as compared to sentences like (10) and (11) (see, however, Taraban and McClelland, 1988; Pynte and Prieur, 1996; and Frenck-Mestre and Pynte, 1997, for somewhat different findings). (10) Les espions informent les gardes du palais. 'The spies inform the guards of the palace.' (11) L 'etudiant choisit un appartement avec balcon. 'The student chooses a flat with a balcony.' These results will not be discussed in any detail here. Instead, I wish to concentrate on the possible role of intonational structuring on attachment decisions in examples like (8) to (11). 3.1. Experiment 5: The Interaction of Argument Structure and Prosody In this experiment (Pynte and Prieur, 1996, experiment 3), a set of 64 sentences, based on 16 di-transitive and 16 monotransitive verbs were presented in two prosody conditions. The eight experimental conditions are presented in Table 3, where the # sign stands for an intonational phrase boundary, and the target word is underlined. The task consisted of detecting the presence of a target word in an utterance. The target word was presented visually on the computer screen, in lowercase letters. It remained displayed for 2 sec and was immediately followed by the auditory presentation of the utterance. The subjects (N = 24) had to press the space bar on the keyboard as soon as they heard the target word. Underlying this technique was the assumption that the time needed to detect a target word in an utterance depends on the complexity of the operations required to incorporate the word into the utterance. For the experimental items, the target word was the last word in the utterance (underlined in Table 3). For the filler and practice sentences, the location of the target word was variable and unpredictable. In 25% of the cases, the visual

238

Joel Pynte TABLE 3 EXAMPLES OF SENTENCES USED IN THE EXPERIMENTa Reaction time (ms) Di-transitive verb VP-att: Les espions # informent les gardes du complot Les espions # informent les gardes # du complot NP-att: Les espions # informent les gardes du palais Les espions # informent les gardes # du palais

(505) (496) (482) (553)

Monotransitive verb VP-att: L' etudiant # choisit un appartement avec soin L 'etudiant # choisit un appartement # avec soin NP-att: L 'etudiant # choisit un appartement avec balcon

(559) (538) (520)

L 'etudiant # choisit un appartement # avec balcon (520) a VP-att, Attachment to the Verb Phrase; NP-att, Attachment to the second Noun Phrase. #, intonational phrase boundary. Target word is underlined.

target had no counterpart in the utterance heard. For these utterances, the subject was not supposed to respond, and the next trial started two seconds later (go/no-go task). Natural speech was used. Four stimuli were contructed from each verb. These were obtained though a splicing procedure, so that, for a given intonation condition, the signal was physically the same up to the preposition for both types of continuation word. Similarly, the last word, which determined the type of attachment, was physically the same for both intonation conditions. The acoustic analysis (syllable duration and fundamental frequency (F0) recorded on the last syllable of NP2) revealed clear differences between the two prosody conditions. Importantly, the phrasing was similar for di-transitive and monotransitve verbs, and was consistent across the set of test sentences. In summary, three within-subject factors were manipulated, namely, Phrasing (the PP was separated or not from NP2 by a prosodic break), attachment (the last word of the sentence forced either VPattachment or NP-attachment), and verb type (di-transitive vs. monotransitive). Mean reponse times are given in parentheses in Table 3. There was a significant interaction between prosody and attachment: Fl(l, 20) = 7.89, p < .05; F2(l, 24) = 12.64, p < .01. The presence of a prosodic break before the PP apparently favored its attachment to the VP, and hindered its attachment to NP2 (this effect is only apparent for di-transitive verbs, however). This overall two-way interaction is consistent with the view that a prosodic break before the PP temporarily "blocks" its attachment to NP2 (see a discussion of this notion in Pynte and

Time Course of Attachment Decisions

239

Prieur, 1996). In accordance with what was said in the preceding section, it will be assumed here that prosodic breaks operate at the thematic level (e.g., by avoiding the PP to be processed as an argument of NP2). The three-way interaction between prosodic phrasing, attachment, and verb type was significant (Fl(l, 20) = 7.79, p < .05; F2(l, 24) = 3.81, p < .063). As can be seen in Table 4, the two longest response times were recorded in those conditions where the final interpretation of the sentence conflicted with both the verb's argument structure and prosody, namely, the monotransitive/no-break/VPattachment condition and the di-transitive/break/NP-attachment condition (i.e., when prosodic structure reinforced an incorrect lexically determined attachment, see examples 12 and 13). (12) La police previent les gardes # du palais.

(553 ms)

(13) L'etudiant choisit un studio avec soin.

(559ms)

How can this interaction be explained in accordance with the view that prosodic breaks operate by temporarily blocking the operation of the thematic processor? A possible solution would be to admit that listeners develop expectations concerning the surface properties of the complements likely to appear after each type of verb and, critically, concerning the approximate processing time that will be required before theta assignment can take place. For example, French listeners know that when two complements are present, the first one is usually quite short. This means that in a sentence like (8) (where two complements are expected), a NP2modifying expression is unlikely. For this reason, the thematic processor will tend to immediately appropriate NP2 to fill a role in the verb's theta grid. Moreover, if we assume that a prosodic break in front of the PP (example 12) hinders its assignment as an argument of NP2, the outcome will be that 'les gardes' will be equated with the theme role of 'previent' before 'du palais' has got a chance to be processed. By contrast, only one (possibly long) complement is expected after the verb 'choisit,' and any materials following NP2 are thus likely NP2 modifiers. The thematic processor will thus tend to "wait" until the complement is complete before filling the verb's theta grid. This strategy is favored by the absence of any prosodic break in example (13), and the theta assigner will (wrongly) try to treat the PP 'avec soin' as a possible argument of NP2, thus explaining the long response times observed. 3.2. Experiment 6: Completion Data This type of interpretation critically relies on the assumption that listeners develop expectations concerning an argument PP in the case of di-tranitive verbs and an NP2-modifier PP in the case of monotransitive verbs. This hypothesis

240

Joel Pynte

was confirmed in a sentence-completion experiment reported in detail in FrenckMestre and Pynte (1997). A total of 56 French, undergraduate psychology students were asked to provide completions to a total of 70 sentence beginnings (in French) presented in a single list, which began with four complete sentences. Of the 70 sentence beginnings, 12 contained di-transitive verbs (e.g. "They accused the woman of..."), and 12 contained monotransitive verbs (e.g., 'Elle regarde la vendeuse de ...'). The 46 remaining items were fillers containing verbs of various types. Subjects were asked to complete the sentences with the first word(s) that came to mind. The results were quite clear-cut. For sentence beginnings that contained a ditransitive verb, the PP was more likely to be an argument of the verb than a NP2 modifier (76% vs. 20% of completions). That is, when given the beginning, "They accused the gangsters of...", they were more likely to complete the sentence with "robbery" than with "Marseille." In contrast, for sentence beginnings that contained monotransitive verbs, 82% of subjects' completions implied NP attachment of the PP, and only 15% implied the VP attachment of an adjunct. For example, given the beginning, "They looked at the saleswoman o f . . . , " subjects answered with nouns such as "store" or "clothes" more often than with nouns like "anger" (in English this would be equivalent of "in anger"). Analysis of variance of the total number of NP and VP attachments for each of the two verb types confirmed the significance of the interaction described above (Fl(l, 55) = 744.24, p < .001; F2(l, 22) = 54.43, p < .001). As a whole, the pattern of results obtained thus gave some support to the idea that a verb's argument structure is likely to trigger expectations concerning the status and attachment site of a subsequent PP.

4. THE DYNAMICS OF ATTACHMENTS This section proposes a tentative generalization of the notions dicussed in section 3. As a whole, the interpretations that have been proposed so far are consistent with the minimal commitment principle, which implies that certain structural decisions are postponed until helpful semantic information becomes available. Moreover, the results discussed in the two preceding sections suggest that the same parsing principles probably apply to nonprimary and primary relationships. This is at odds with Frazier and Clifton's (1996) model, which maintains that primary relations must be fully and immediately settled at the syntactic level. Following Perfetti (1990), I am going to assume here that only "local" attachments (e.g., article + noun) are obligatory, and that all "remote" ones (including predicate-argument relations) can be delayed. The main criticism that can be opposed to this proposal is linked to the assumed memory limitations of the parser (see the introduction section). A first possible answer to this argument can be

Time Course of Attachment Decisions

241

looked for in section 3 of this chapter, where it was suggested that the listener's knowlege concerning certain regularities of the language might help him to predict the moment when a given relationship is likely to be settled. Going further along this line of reasonning, one can imagine a less memory-consuming process in which "waiting times" will be optimally reduced. Syntactic and lexical informations become available in a continuous way during parsing, and the main problem for the parser is one of synchronizing these different sources of information as they emerge from the input stream. According to Perfetti, two distinct levels must be distinguished, namely the constituent builder stage, where low-level phrasal units such as NPs and PPs are first constructed, and the argument filler stage, where these low-level units are subsequently integrated according to the specifications associated to the lexical entry of the sentence's main verb (with all the pending attachments being presumably settled at the same time). The proposal made here can be seen as an attempt to further describe the possible dynamics of this second stage. For the sake of simplicity, local relationships are assumed to be first built up in the way described by the pieces model. The novelty is that each low-level contituent is assumed to be associated with a delay grossly indicating its forseeable processing rank. This delay is then passed to the thematic processor, with the consequence of blocking the interpretation of the phrase in question for the corresponding time lapse. It could be argued that the sole delays (potentially) determine the final structure of the sentence, and that Perfetti's second step thus becomes unnecessary. Indeed, any (binary) tree is fully specified by the ranks at which two successive (recursively defined) nodes are allowed to combine. This is illustrated in example (14) below, where the letters correspond to terminal nodes and the numerical values to processing ranks. It must be assumed that all the values are decremented at each processing step and that two nodes combine when separated by a zero value. (14) stepl: step 2: step 3: step 4:

a 3b l c 2d a 2 (b c) 1 d a l ((b c) d) (a ((b c) d))

Note, however, that we are not dealing here with a total order. The delays do not directly indicate the exact rank at which the corresponding phrase will be processed. They are assumed to correspond, instead, to approximate time lapses imposed to the thematic processor, and the actual processing rank of a given element is likely to depend on the subsequent elements (and associated delays) present in the input stream. If a sufficently long delay is associated with a given element n, then it may happen that a subsequent element, say n + 2, will present itself to the parser before n has been processed by the thematic processor. If the new element is asociated with a sufficently short delay (i.e., granted a high priority), then it may

242

Joel Pynte

happen that n + 2 will be interpreted before n has got a chance to be. This permits monotonic on-line "revision" in the same way as the D-theory does (see Sturt & Crocker, 1996). For example, sentence (15) will be correctly interpreted as meaning "He knows the truth," whereas sentence (16) will be interpreted as "He knows that the truth is hard": (15) He 5 knows 3 (the truth) (16) He 5 knows 3 (the truth) 1 (is hard) In fact, introducing grossly defined delays amounts to partly specifying the structure. However, instead of expressing the relationships in terms of dominance and precedence between nodes, like in the D-theory, the listener's knowlege concerning attachment possibilities are assumed to be translated in temporal terms. Such an assumption seems quite natural concerning the human parser, whose final purpose is to provide on-line interpretations. Moreover, it should be noted that the notion of delay provides an easy way to express statistical regularities (whether in the language as a whole or resulting from an experimental manipulation) as well as grammatical knowledge. For example, the observed preference for highattaching the relative clause in N1-de-N2-qui... constructions (see section 1) can be expressed by systematically introducing a delay in front of any relative pronoun. Delays also provide an obvious way to express prosody and punctuation. Let us consider again the NP1-V-NP2-AP construction discussed in section 2. An F0 fall at the end of NP2 was shown to hinder the final AP from being interpreted as an argument of the verb. This can be explained if one admits that, in this condition, the role of agent was equated to NP1 before AP interpretation was allowed to begin (with the consequence that, at that time, the thematic processor had to deal with the full clause and not with the sole verb's theta-grid). A similar reasoning can account for the interaction between argument structure and prosody found in section 3 for PP-attachment. The hypothesized processing steps for each of the eight conditions in Table 3 are presented in Table 4. The delay introduced after the verb has been set to 1 for di-transitives (e.g., "informer") and to 2 for monotransitives. This is because the first complement of a di-transitive verb is assumed to be quite short (see section 3). The delay corresponding to a prosodic break has been set to 1. The left part of Table 4 corresponds to syntactic parsing (building up low-level constituents and determining the delays), whereas the right part corresponds to thematic processing (equating low-level constituents with thematic roles). Each delay value is decremented at each step. A phrase is presented to the thematic processor when preceded by a zero value. If two zero delays are present, the leftmost prevails. Attachment can only be made if theta assignment is succesful. Failure at this level does not entail any penalty (apart from hindering attachment). The system merely proceeds to the following step. However, if the same phrase is then

243

Time Course of Attachment Decisions TABLE 4 THE HYPOTHESIZED TIME COURSE OF PP-ATTACHMENT DECISIONSa Parser informent 1 (les gardes) 0 (du complot) informent 0 (les gardes) 0 (du complot) (informent (les gardes)) 0 (du complot) ((informent (les gardes)) (du complot)) informent 1 (les gardes) 1 (du complot) informent 0 (les gardes) 0 (du complot) (informent (les gardes)) 0 (du complot) ((informent (les gardes)) (du complot)) informent 1 (les gardes) 0 (du palais) informent 0 ((les gardes) (du palais)) (informent ((les gardes) (du palais))) informent 1 (les gardes) 1 (du palais) informent 0 (les gardes) 0 (du palais) (informent (les gardes)) 0 (du palais) (informent (les gardes)) 0 (du palais) * ((informent (les gardes)) (du palais)) choisit 2 (un appartement) 0 (avec soin) choisit 1 (un appartement) 0 (avec soin) choisit 0 ((un appartement) (avec soin)) * (choisit ((un appartement) (avec soin))) choisit 2 (un appartement) 1 (avec soin) choisit 1 (un appartement) 0 (avec soin) choisit 0 (un appartement) 0 (avec soin) (choisit (un appartement)) 0 (avec soin) ((choisit (un appartement)) (avec soin))

Theta assigner decision les gardes du complot? informent les gardes ? informent du complot?

N Y Y

informent les gardes? informent du complot?

Y Y

les gardes du palais ? informent les gardes?

Y Y

informent les gardes? informent du palais ? informent du palais!!!

Y N

un appart. avec soin? un appart. avec soin!!! choisit un appart. ?

N Y

un appartement avec soin? choisit un appartement? choisit avec soin ?

N Y Y

choisit 2 (un appartement) 0 (avec balcon) un appart. avec balcon? choisit 1 ((un appartement) (avec balcon)) choisit 0 ((un appartement) (avec balcon)) choisit un appartement? (choisit ((un appartement) (avec balcon))) choisit 2 (un appartement) 1 (avec balcon) choisit 1 (un appartement) 0 (avec balcon) un appart. avec balcon? choisit 0 ((un appartement) (avec balcon)) choisit un appartement? (choisit ((un appartement) (avec balcon))) a

Y Y

Y Y

!!!, forced attachment; * garden-path condition.

presented again to the theta assigner, it is assumed here that the attachment is forced (thus leading to a garden path). This is indicated by the !!! sign. Out of the eight conditions, two conducted to a garden-path (* sign). Interestingly these two conditions are precisely the ones which gave rise to the longest response times (see Table 3 and examples 12 and 13 above).

244

Joel Pynte

REFERENCES Baltazart, D., and Kister, L. (1995). Correlation entre determination et selection d'un anaphorise dans une structure N de N. Paper presented at the Seminaire "Anaphore et Reference". Nancy (CRIN): 20-22 September. Brysbaert, M., and Mitchell, D. C. (1996a). Modifier attachment in sentence parsing: Evidence from Dutch. Quarterly Journal of Experimental Psychology, 49A, 664-695. Brysbaert, M., and Mitchell, D. C. (1996b). Modifier attachment in Dutch: Deciding between garden-path, construal and statistical tuning accounts of parsing. Paper presented at the workshop on Computational Models of Human Syntactic Processing, NIAS, Wassenaar, June 5-8. Carreiras, M., and Clifton, C. (1993). Relative clause interpretation preferences in Spanish and English. Language and Speech, 36, 353-372. Clifton, C., Speer, S., and Abney, (1991). Parsing arguments: Phrase structure and argument structure a determinants of initial parsing decisions. Journal of Memory and Language, 30, 251-271. Crain, S., and Steedman, M. (1985). On not being led up the garden-path: The use of context by the psychological syntax processor. In D. R. Dowty, L. Karttunen, and A. M. Zwicky (Eds.), Natural language parsing: Psychological, computational and theoretical perspectives. Cambridge, UK: Cambridge University Press. Cuetos, F., and Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the use of the Late Closure strategy in Spanish. Cognition, 30, 73-105. Cuetos, F, Mitchell, D. C., and Corley, M. (1996). Parsing in different languages (pp. 145187). In M. Carreiras, J. Garcia-Albea, & N. Sebastian-Galles (Eds.), Language Processing in Spanish. LEA Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII. Hillsdale, NJ: Lawrence Erlbaum Associates. Frazier, L., and Clifton, C. (1996). Construal. Cambridge, MA: MIT Press. Frazier, L., and Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Frenck-Mestre, C., and Pynte, J. (1997). Syntactic ambiguity resolution while reading in second and native language. Quarterly Journal of Experimental Psychology, 50A, 119-148. Frenck-Mestre, C., and Pynte, J. (in press). Resolving syntactic ambiguities: Crosslinguistic differences? In V. Lombardo and M. De Vincenzi (Eds.). In preparation. Garnham, A. (1983). Mental models as representations of discourse and text. Chichester: Ellis Horwood. Grice, H. P. (1975). Logic and conversation. In P. Cole & J. L. Morgan (Eds.), Syntax and Semantics, Vol. 3: Speech acts (pp. 41-58). New York: Academic Press. Hirst, D. (1987). La Decription Linguistique des Systemes Prosodiques: Une Approche Cognitive. These de doctoral d'Etat. Univerite de Provence, Aix-en-Provence, France. MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). Lexical nature of syntactic ambiguity resolution. Pychological Review, 101, 676-703. Marcus, M., and Hindle, D. (1990). Description theory and intonation boundaries. In

Time Course of Attachment Decisions

245

G. T. M. Altmann (Ed.), Cognitive models of speech processing: Psycholinguistic and computational perspectives. Cambridge, MA: MIT Press. Marcus, M., Hindle, D., and Fleck, M. (1983). D-theory: Talking about talking about trees. In Proceedings of the Twenty-First Annual Meeting of the Assiciation for Computational Linguistics. Cambridge, MA: MIT Press. Mitchell, D. C., and Cuetos, F. (1991). The origins of parsing strategies. Conference Proceedings: Current Issues in Natural Language Processing. University of Texas at Austin, Austin, TX. Mitchell, D. C., Corley, M. M. B., and Garnham, A. (1992). Effects of context in human sentence parsing: Evidence against a discourse-based proposal mechanism. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 69-88. Mitchell, D. C., Cuetos, F, and Corley, M. B. B. (1992). Statistical versus linguistic determinants of parsing bias: Cross-linguistic evidence. Paper presented at the Fifth annual CUNY conference on Human Sentence Processing, CUNY, NY. Mitchell, D. C., Cuetos, F, and Zagar, D. (1990). Reading in different languages: Is there a universal mechanism for parsing sentences? In D. A. Balota, G. B. Flores d'Arcais, and K. Rayner (Eds.), Comprehension processes in reading. Hillsdale, NJ: Lawrence Erlbaum Associates. Perfetti, C. A. (1990). The cooperative language processor: Semantic influences in an autonomous syntax. In D. A. Balota, G. B. Flores d'Arcais, and K. Rayner (Eds). Comprehension processes in reading. Hillsdale, NJ: Erlbaum. Pynte, J., and Frenck-Mestre, C. (1996). Early-closure attachment in French: A replication. Poster presented at AMLAP'96, Torino, Italy, September. Pynte, J., and Prieur, B. (1996). Prosodic breaks and attachment decisions in sentence parsing. Language and Cognitive Processes, 11, 165-191. Rayner, K., Carlson, M., and Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: eye movements in the analysis of semantically biased sentences. Journal of Verbal Learning and Verbal Behavior, 22, 358-374. Rayner, K., and Frazier, L. (1987). Parsing temporarily ambiguous complements. Quarterly Journal of Exprimental Psychology, 39A, 657-673 Schwarze, C. (1988). The treatment of French adjectif detache in Lexical Functional Grammar. In U. Reyle & C. Rohrer (eds.), Natural language parsing and linguistic theories (pp. 262-288). Boston, MA: Reidel. Sturt, P., and Crocker, M. W. (1996). Monotonic syntactic processsing: a cross-linguistic study of attachment and reanalyis. Language and Cognitive Processes, 11,449-494. Zagar, D., Pynte, J., and Rativeau, S. (1997). Evidence for early-closure attachment on firstpass reading times in French. Quarterly Journal of Experimental Psychology, 50A, 421-438.

This page intentionally left blank

VERB PROCESSING IN GERMAN AND ENGLISH: AMBIGUITY, DISCONTINUOUS FORMS, AND THEMATIC COMPLEXITY DIETER HILLERT Department of Psychology University of California, San Diego La Jolla, California

1. INTRODUCTION To comprehend language, a continuous phonetic string is first segmented to access lexical information (syntactic and semantic entries), which is then organized into various syntactic relationships; in turn such sentential units (propositions) must be combined to give rise to the more elaborate units of meanings embodied in conceptual processing (pragmatics). Our approach taps on-line into sentence processing to investigate the extent to which different levels of linguistic representations work independently (modular) or interactively during the temporal course of language comprehension. Accordingly, there are two general accounts of normal sentence comprehension in the literature: (a) the constraint satisfaction account, and (b) the contextindependent account. The constraint satisfaction account claims that language processing is a highly interactive operation. Access to words is facilitated by sentence context (Marslen-Wilson, Brown, and Tyler, 1988; Tabossi, Colombo, and Job, 1987), and the selection of a certain parsing strategy occurs on encountering lexical information (MacDonald, Pearlmutter, and Seidenberg, 1994). In contrast, the context-independent account claims that initial processes are modular (Fodor, 1983; Foster, 1979; Shapiro, Grimshaw, and Zurif, 1987; Swinney, 1979). Thus, parsing is an autonomous process taking place independent of lexical entries Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

247

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

248

Dieter Hillert

(Frazier and Rayner, 1982), and words are initially accessed without contextual penetration (Seidenberg, Tanenhaus, Leiman, and Bienkowski, 1982; Swinney, 1979; Tanenhaus, Leiman, and Seidenberg, 1979). Whether evidence can be found for a modular or an interactive model seems to be largely dependent on the experimental technique used to examine on-line sentence processing. A "context-free" on-line technique is, for example, the crossmodal lexical priming (CMLP) or cross-modal lexical decision (CMLD) tasks (and modifications such as CML naming or CM picture priming). Subjects perform two tasks: They listen to a sentence and simultaneously perform a lexical decision/naming task on a written (or pictorial) stimulus during the sentence presentation. These methods have been shown to be both sensitive and reliable for tapping on-line into sentence processing (e.g., Fodor, 1983; Forster, 1981; Nicol, Fodor, and Swinney, 1994; Nicol, Swinney, Love, and Hald, 1997; Shapiro et al., 1987; Swinney, Onifer, Prather, and Hirshkowitz, 1979). In contrast, "contextsensitive" on-line techniques such as cross-modal lexical continuation tasks (lexical decision or naming) seem to be more reliable in measuring interactive activation patterns. While listening to a sentence, subjects are asked to make a lexical decision on a word (or to name a word or picture) that is part of the sentence itself. Thus, the auditory sentence presentation is interrupted while subjects perform a secondary task. These techniques seem to be sensitive and reliable for examining sentential probabilities; that is, they measure the linguistic and pragmatic probability that word X will occur at a certain position Y in a sentence context Z (e.g., Grosjean, Dommergues, Cornu, Guillelmon, and Besson, 1994; Gurjanov, Lukatela, Savic, and Turvey, 1985; Hillert and Bates, 1996). The results of experiments using other kinds of so-called on-line techniques are more difficult to interpret. For example, monitoring and repetition tasks (Marslen-Wilson, 1973; Marslen-Wilson, Brown, and Tyler, 1988; Marslen-Wilson and Tyler, 1980) seem to reflect effects due to contextual integration (postlexical processing) rather than a bottom-up lexical access mechanism (Hillert, 1997). The present approach explores a "universal feature" of human sentence processing by considering the crosslinguistic comparison of on-line verb processing in German and English sentence comprehension. A "context-free" paradigm is applied to generate data not confounded with inter- and intrasubjective variability; at the same time abstract linguistic structures are considered because they provide specific predictions for how sentence processing unfolds over time.

2. VERB AMBIGUITIES In English there are only few syntactic structures that provide the opportunity to examine the extent to which the theta-role(s) of a verb's argument might constrain access to the verb's meaning. So-called cleft sentences provide such an op-

Verb Processing in English and German

249

portunity. According to Chomsky's (1981) Government Binding (GB) theory, the canonical SVO-structure in English is the basic syntactic structure from which all other noncanonical structures are generated. For example, sentences (2) and (3) are derived from sentence (1). The structure in (2) illustrates a cleft-subject sentence having an SVO structure in the relative clause. However, (3) is a cleft-object sentence; that is, the direct object Mary is moved out of its canonical position (Spec, specifier) to a preverbal position leaving behind a trace. (1) Mary kissed John. (2) It is Johni whoi kissed Mary. (3) It is Maryi whoi John kissed

i.

In one study cleft-sentences are presented to examine access to verb ambiguities (paper in progress). According to previous findings with noun-noun ambiguities, it is to expect that verb ambiguities are exhaustively accessed in canonical structures (e.g., Onifer and Swinney, 1981; Swinney, 1979; Tanenhaus, Leiman, and Seidenberg, 1979; but see Tabossi, Colombo, and Job, 1987).Therefore, cleftsentences were presented to provide the strongest semantic-syntactic and conceptual bias for a single verb meaning. To guarantee the strongest bias either an object (argument) or an adjunct was presented in the preverbal position. As shown in (4) and (5) the cleft-sentences examined strongly biased a single reading of the verb homophone. Both meanings of the verb homophone were equally frequent. In applying a CMLP task the semantically related probes ARREST and RESERVE were presented at the immediate vicinity of the verb (# marks the test point). (4) It was for robbing the bank that the police booked # the suspect in the red hat. (5) It was through the travel agent that the tourist booked # the train ticket. The response patterns of this study clearly show that both verb meanings were initially activated independent of the context information provided in the preverbal position (across 64 subjects; one-tailed t-tests: p < .05 for (M = meaning, C = context) Ml X Cl, Ml X C2, M2 X Cl, and p < .01 for M2 X C2) (see Figure 1). Thus, neither a specific PP in Spec nor the semantic-conceptual bias did constrain access to a single verb meaning. Instead, both meanings were exhaustively accessed at the vicinity of the verb. These response patterns replicate previous findings found with noun-noun verb ambiguities (see above). The issue of an autonomous parser is based mostly on evidence found in English sentence processing (e.g., Frazier, 1989), but the frequency of sentence structure must also be considered (e.g., MacDonald, Pearlmutter, and Seidenberg, 1994). Cleft-sentences certainly are not very common structures used in everyday spoken English. Thus, it is possible that the English parser cannot be sensitive enough to take into account lexical information preverbally presented. The parser is simply not tuned through everyday conversation to compute the PP distinction

250

Dieter Hillert

Figure 1. Access to verb homophones in cleftsentences (M = meaning, C = context). Significant priming was found in all conditions (see text).

related to different verb meanings in such idiosyncratic structures. Thus, it may be important to examine noncanonical sentence structures that are as frequent as canonical sentence structures. In particular, we find this constellation in languages allowing a relatively free word order, such as the so-called SOV language German.

3. DISCONTINUOUS VERB PARTICLE COMBINATIONS This section presents the data from two experiments that investigated how discontinuous German verbs are accessed during on-line sentence comprehension. The results challenge traditional claims about how the parser handles a morphosyntactically realized object placed in Spec position (Hillert, 1998a). Both experiments investigated the following questions: 1. Are meanings of discontinuous verbs exhaustively accessed at the offset of the basic verb?

Verb Processing in English and German

251

2. Do overt case markers placed in Spec constrain access to syntactically possible discontinuous verb meanings? Schreuder (1990) used an isolated word-priming paradigm to examine processing differences between separable and nonseparable prefixed verbs in Dutch. He found that a separable prefix primes its stem (and conversely), but a nonseparable prefix does not. For the present approach the inverse order (stem primes a possible separable prefix) is of particular interest because this order is comparable to the order of discontinuous verb information in Dutch and German sentences for example, but with numerous intervening words (cf. Frazier, Flores d'Arcais, and Collen, 1993). An ambiguity arises at the verb stem or basic verb in present tense and simple past. If we assume that German is an SVO-language, the verb remains in the second position (V2), and the verb's particle moves to the sentence-final position. However, if we assume that German is an SOV-language, movements of the verb stem and its particle are explained quite differently (Chomsky, 1986). For example, morphologically complex verbs such as mitsingen (vt. 'with + sing'; tr. 'to sing with someone') have their own entry in the mental lexicon and are represented as such at the D-structure (deep structure) of a sentence, as illustrated in Figure 2, with the sentence Fritz sang in dem Chor mit (vt. 'Fritz sang in the choir with'; tr. 'Fritz sang in the choir'): (a) The verb stem sang ('sang') first moves

Figure 2. Verb stem movement in an SOV language.

252

Dieter Hillert

right, leaving behind a trace, and (b) then moves to C (constituent). The particle mit ('with') is left behind and remains left-adjacent to the trace of the verb + sang ('sang'). The final movement (c) generates the subject out of the inflectional phrase (IP) to the Spec position. In other words, for an SOV-language, the verb stem moves and it is the particle that remains in the sentence-final position. This generative model has no direct implications for how we compute ambiguous verb forms on-line from left to right. An ambiguity arises in present and past tense because the basic verb can refer to several possible discontinuous verb meanings. For example, (6) shows that the basic verb geben ('to give') can be linked to 10 different particles, some of these discontinuous verb forms are also ambiguous (e.g., aufgeben: vt. 'on + give', tr. 1, 'to give up'; tr. 2 'to send off', or wiedergeben: vt. 're + give', tr. 1 'to reproduce', tr. 2 'to return'). The relatively free word order in German allows the speaker to fill the Spec position with an accusative or dative object. Intransitive meanings such as OVERSTATE (6a) and GIVE UP (6b) should not be directly accessed at the offset of the verb stem, because Spec is not filled with a subject, and therefore a transitive or ditransitive verb meaning is required. Again, no syntactic violation occurs in (6c-6f). However, in the marked cases (6g) and (6h), a dative object sounds inappropriate, but this syntactic template may still be possible. (A native speaker would prefer to use a prepositional object such as fur ihnlsie ['for him/her'].) Finally, a dative object in Spec contradicts the verb meanings pretend (6i) and admit (6j) because both verbs subcategorize only for an accusative object or a sentence clause (S'), but not for a dative object. (6) a. b. c. d. e. f. g. h. i. j.

angeben aufgeben abgeben weitergeben wiedergeben mitgeben eingeben weggeben vorgeben zugeben

'overstate' 'give up' 'hand over' 'pass on' 'reproduce' 'give along' 'enter something' 'giveaway' 'pretend' 'admit'

[-] [—] [__(NP)NP] [_ (NP) NP] [_ (NP) NP] [_ (NP) NP] [_ (*NP) NP] [_ (*NP) NP] [_NP][_S'] [_NP][_S']

The CMLP-paradigm was used by Hillert (1998a) to examine access to discontinuous verb meanings during on-line sentence comprehension. The first experiment tested whether the complete meaning of two different verb-particle combinations, which are comparable in frequency, is accessed at the offset of the verb stem. The discontinuous verbs were presented in two different contexts biasing a specific verb-particle meaning. The Spec position was filled either by a subject or an accusative object to provide a strong semantic bias. In this case, it was possible

Verb Processing in English and German

253

to examine ambiguous discontinuous verbs with an identical particle. For example, in the semantically-biasing sentences (7) and (8) it was probed for both meanings of the verb aufgeben (vt. 'off + give'; tr. 1, 'sending out'; tr. 2 'quit') by presenting the related target words SCHICKEN ('sending') and ENDEN ('finishing') at the offset of the verb stem (#). (7) Das aus Spanien kommende Paket gab # der Bote so schnell wie moglich auf. vt. 'The from Spain arriving parcel gave # the carrier as soon as possible off.' tr. 'As soon as possible the carrier sent off the parcel that arrived from Spain.' (8) Der am Boden liegende Boxweltmeister gab # den Kampf nach einigem Zogern auf. vt. The on the ground laying world champion in boxing gave # the fight after some hesitation off.' tr. The boxing champion who was on the ground quit the fight after some hesitation.' Examining on-line access to discontinuous verbs is therefore a crucial test for the hypothesis of form-driven access. The question is whether verb meanings are immediately activated, even though only a part of the complete word form is presented. The results of this study are shown in Figure 2 (left-hand side). Significant lexical priming was found for all relevant conditions, that is, both discontinuous verb meanings were immediately activated independent of biasing context (across 40 subjects, one-tailed Mests: p < .05 for Ml X Cl, Ml X C2, M2 X Cl, and p < .01 for M2 X C2). This result supports an exhaustive-access account: Multiple lexical access is measured even in case of partial input of the relevant word form. The second experiment examined discontinuous verb meanings that were comparable to the first study. Thus, pairs of verb-particle combinations having a common verb stem and being equal in frequency were used, but they were presented in sentences that had either an accusative or dative object in the Spec position. Four different types of discontinuous verb meanings were examined at the offset of the verb stem. The first two meanings were congruent to noncanonical sentence structures; that is, they allowed an accusative or dative object. These constellations are called ACC X ACC (accusative context X accusative meaning) and DAT X DAT (dative context X dative meaning). For example, in (9) and (10) the congruent meaning schicken ('to send off': [ (NP)ACC]) of the ambiguous verb aufgeben was examined. The listener is informed of the occurrence of a dative object at the inflection (the last letter -n to be precise) of kommenden, which is called participle I or gerundive (adjective verb (infinitive) + endjen}). The determiner der alone refers to the case/gender constellations "nominative-mas-

254

Dieter Hillert

culine-singular," "dative-feminine-singular," "genitive-feminine-singular," or "genitive-plural." The second two verb meanings were incongruent to noncanonical sentence structures; they violated the DAT- or ACC-context. By examining the intransitive (INTR) meaning of Aufgeben with the related probe arrogant (Germ = Engl. 'arrogant'), we examined in sentences (9) and (10) the incongruent condition *ACC X INTR and *DAT X INTR. Thus, if the parser takes into account morphosyntactically realized case information, no activation of the intransitive meaning 'arrogant' should be measured. Otherwise, if we measure a multiple access, the results found for the English parser reported so far would be confirmed; that is, syntactic context does not influence initial lexical access. (9)

Der aus der Nachbarschaft kommenden FrauIO gab # erSUBJ das PaketDO auf. vt. 'The from the neighborhood coming womanIO gave # heSUBJ the parcelDO on.' tr. 'He sent off the parcel for the woman who is from the neighborhood.'

(10) Den von vielen Firmen begehrten Rohstoffoo gab # DanielSUBJ auf dem Postamt auf. vt. 'The of many firms demanded raw materialDO gave # DanielSUBJ on the post office on.' tr. 'At the post office Daniel sent off the raw material demanded by many firms.' The second part of this study also examined the congruent condition DAT X DAT and ACC X ACC. Thus, in (11) and (12) the related words leihen ('to borrow') and Gewehr ('rifle') were presented. In addition, we probed for DAT meanings violating both sentence structures: In the first incongruent condition *DAT X ACC, the verb meaning does not allow an NP. For example, in (11) the morphologically marked NP (dem: "masculine/neuter-dative-singular") in the sentenceinitial position informs the listener that she or he will not encounter a verb that does not allow an indirect object (called here ACC meaning). In the second incongruent condition, *ACC X DAT, the direct object requires different semantic features. For example, it is possible to borrow money from someone, but taking into account our common sense it is not possible to borrow something from the bear, as examined in (12). The verb meaning vorlegen ('to borrow') requires the semantic feature [+HUMAN] if the object is [+ANIMATE]. (11) Dem immer zuvorkommenden und hoflichen KollegenDAT legte # sieNOM ein MonatsgehaltACC vor. vt. 'The always courteous and polite colleagueDAT laid # sheNOM a monthly paymentACC before.'

Verb Processing in English and German

255

tr. 'She borrowed a monthly payment from the courteous and polite colleague.' (12) Den aus den Wdldern entkommenen BarACC legte # die SpezialeinheitNOM um. vt. 'The from the woods fleeing bearACC laid # the special forceNOM to.' tr. 'The special force killed the bear coming from the woods.' Overall, for the congruent conditions ACC X ACC and DAT X DAT, highly significant lexical priming was found (across 57 subjects: p < .0001). Thus, this confirms the results reported so far; that is, verb-particle meanings are multiply activated immediately at the offset of the verb stem independent of whether the Spec position is filled by a subject, direct object, or indirect object. Moreover, this finding holds even though the semantic contexts provided in the second experiment were relatively neutral. However, for the incongruent condition a different response pattern was found. Surprisingly, a highly significant inhibitory effect was found for *ACC X INTR (p < .0001), and significant priming for *DAT X INTR (p < .001). Thus, the German parser seems to inhibit access to an intransitive verb meaning if the Spec position is filled by an accusative object, but does not restrict access to an intransitive meaning if Spec is filled by a dative object. Comparable response patterns were found for DAT violations. A significant inhibitory effect was found for *ACC X DAT (p < .05), and highly significant priming for *DAT X ACC (p < .0001). Thus, if Spec was filled by a morphosyntactically marked dative object, significant priming was still observed, but not when Spec was filled with a morphosyntactically marked accusative object (see Figure 3: right-hand side). The results were interpreted according to the distinction between lexical case and structural case. Structural case, such as nominative or accusative, reflects a configuration that depends solely on government. In contrast, lexical case, such as dative or genitive, depends on theta-role assignment in addition to government; that is, it depends on the inherent properties of the lexical element assigning the argument. Nominative and accusative alternate with respect to the same verb (theta-role), and case can be absorbed (e.g., Sie sieht ihnACC / 'She sees him'; ErNOM wird gesehen / 'He is seen'). However, no "case absorption" takes place with lexical case. The German speaker must learn for each individual lexical element (e.g., prepositional object, verb, adjunct) whether it can assign lexical case, such as dative or genitive. With other words: lexical case is an inherent property of the verb, whereas structural case is assigned on the syntactic surface structure. Thus, an exhaustive access has been found, even for incongruent conditions, if Spec was filled by a dative object. This response pattern may be explained by the verb-internal representation of lexical case: All lexical entries of a verb are immediately accessed independent of semantic-syntactic context. Conversely, the German parser constrains verb access by considering structural case in the

256

Dieter Hillert

Figure 3. Left-hand side: Access to discontinuous verb meanings at the offset of the ambiguous verb stem within a semantically biased context. Right-hand side: Access to syntactically possible and impossible discontinuous verb meanings at the offset of the ambiguous verb stem. An exhaustive access was found for both congruent conditions. (ACC X ACC; DAT X DAT). For incongruent conditions, an exhaustive access was also found if Spec was filled by a dative object (*DAT X ACC, *DAT X INTR), but not if Spec was filled by an accusative object (*ACC X DAT, *ACC X INTR) (see text for details).

Spec position. For this reason automatic inhibitory effects were found for the incongruent conditions that did not allow either an intransitive meaning or a dative meaning. However, in considering a constraint-satisfaction account, an alternative interpretation of these data might be possible. The question arises whether the dissociation found between lexical case and structural case can be explained by the frequency distribution of different German verb types. One might assume that the patterns observed in English may also appear in German: transitive verbs are more common than intransitive verbs, which are in turn more common than ditransitive verbs (cf. MacDonald et al., 1994). However, because the present study examined equally frequent discontinuous verb meanings in semantically neutral contexts, the interplay between semantic-syntactic context and verb frequency would be a plausible explanation of the outcome. For example, a dative object in Spec does

Verb Processing in English and German

257

not necessarily prime a subordinate (less frequent) ditransitive meaning, because in German there are a large number of transitive dative verbs. Thus, the argument that an exhaustive access was found within a dative context because the dative marker primes the subordinate ditransitive verb meaning, while the more frequent intransitive verb meaning does not require a contextual bias to be immediately accessed, seems to be an inappropriate explanation. Similarly, the accusative context might bias either a transitive or ditransitive meaning (e.g., verbs with two accusative objects or with dative and accusative objects). Thus, it appears to be premature to claim that the intransitive verb meaning was inhibited because the accusative marker in Spec biased the more frequent transitive meaning and suppressed the subordinate intransitive meaning. In addition, there are other reasons to believe that this line of argumentation does not explain the response pattern. According to different frequency norms (Meier, 1969; Ruoff, 1981), dative and accusative case are equally frequent in German. However, no information is available on whether noncanonical sentence structures with an accusative or dative marker in Spec position are less frequent than the canonical SVO-structure. If the fact is taken into account that accusative and dative markers are equally frequent in everyday spoken German, it is difficult to stay with the hypothesis that syntactic frequency explains the dissociation found. For these reasons the distinction between lexical and structural case is regarded to be fundamental for the prediction of immediate access to (discontinuous) verb meanings in a specific syntactic context.

4. THEMATIC COMPLEXITY Another approach to examine how verb meanings are accessed during on-line sentence comprehension is to measure the processing load in the immediate vicinity of the verb. The processing load (computational costs) may reflect the difficulty involved in accessing a verb's lexical.entries: The greater the processing load at the offset of a verb, the greater the difficulty in accessing the lexical entries of this verb. By careful distinction between the number of possible subcategorizations, the number of possible argument structures, and the number of possible theta-roles, the opportunity can be provided to examine optional and obligatory semantic and syntactic information inherent to a verb. Several studies in English found evidence that the more argument structures are inherent to a verb, the more computationally expensive it is to access the verb. Thus, the listener seems to activate all possible arguments of a verb, independent of whether a theta-role is assigned or not (e.g., Shapiro et al., 1987; Shapiro, Brookins, Gordon, and Nagel, 1991; but see Schmauder, 1991; Schmauder, Kennison, and Clifton, 1991). In a

258

Dieter Hillert

more recent study Ahrens and Swinney (1995) explained different verb complexity effects not by the number of arguments, but by the number of theta-roles associated with each verb. Hillert's (1998b) cross-modal lexical decision study examined whether a thetarole complexity effect can be found in a "semiconfigurational" language; that is, the type and number of theta-roles of a German verb can be constrained by presenting either a dative or accusative object in Spec. Five different sentence structures were examined. In (13), the verb versprechen (to promise) must assign three theta-roles because the listener encounters a dative object first. However, in (14) the dative object is optional, and either one or two theta-roles can be assigned immediately after the verb. In contrast, the verb bewundern ('admiring') must assign two theta-roles whether the accusative object is in Spec position (15) or not (16). Finally, in (17) only one theta-role can be assigned because schnarchen ('snoring') is an intransitive verb.

(13) 3 [ NP NP]: DAT X DAT Der reizenden NichteDAT versprach # der reiche OnkelNOM einen DiamantenACC aus Brasilien. vt. 'The nice nieceDAT promised # the rich uncleNOM a diamondACC from Brazil.' tr. 'The rich uncle promised the nice niece a diamond from Brazil.' (14) 3-2 [ (NP) NP] : NOM X DAT Der reiche OnkelNOM versprach # {der reizenden NichteDAT} einen DiamantenACC aus Brasilien. vt./tr. 'The rich uncleNOM promised # {the nice nieceDAT} a diamondACC from Brazil.' (15) 2 [ NP] : ACC X NOM Den hohen LeuchtturmACC bewunderte # der bekannte ArchitektNOM auf der Insel. vt. 'The high lighthouseACC admired # the well-known architectNOM on the island.' tr. 'The well-known architect admired the high lighthouse on the island.' (16) 2 [ NP]: NOM X ACC Der bekannte ArchitektNOM bewunderte # den hohen LeuchtturmACC auf der Insel. vt./tr. 'The well-known architectNOM admired # the high lighthouse on the island.' (17)

1[ ] : NOM X NOM Der kleine PudelNOM schnarchte # unheimlich laut. vt./tr. 'The little poodleNOM snored # terribly loud.'

Verb Processing in English and German

259

In considering the results found for the distinction between lexical case and structural case (Hillert, 1998a), a significant difference in processing load should be measured between sentence contexts constrained either by nominative/accusative or by dative. No significant difference in processing a dative verb should be measured according to the distinction between canonical and noncanonical sentence structures: If all possible theta-roles are activated in both contexts using a dative verb (13, 14), then the processing load is equal (maximal 3). Similarly, no difference should be measured among the accusative verb conditions because their number of theta-roles is equal (2). However, across the verb types the theta-role

Figure 4. Theta-role complexity: A significant difference between canonical and noncanonical sentence structures according to verb type was only found with dative verbs (DAT X DAT vs NOM X DAT). No significant difference was found among accusative verbs (ACC X ACC vs. NOM X ACC) and between these accusative verbs and intransitive verbs (NOM X NOM).

260

Dieter Hillert

complexity is different (dative-3; accusative-2; nominative-1), and therefore significant differences should be observed between lexical (dative) and structural case (accusative and nominative) (Figure 4). The data from this experiment revealed a significant theta-role complexity effect between canonical and noncanonical sentences only when a dative object was presented in Spec. If an accusative object filled Spec, no significant difference was found compared to the canonical sentence structure (see Figure 4). The results found for the single-case conditions did not confirm the complexity issues stated in (16-20). The reaction times were significantly faster for the NOM X DAT condition (canonical) than for the DAT X DAT condition (noncanonical) (across 49 subjects: p < .05). The morphosyntactic realization of the indirect object in the canonical subject position seems to affect the processing load in the immediate vicinity of the verb. One explanation might be that only the maximum number of theta-roles is exhaustively activated, and only if this number is obligatory. According to this view, the significantly faster reaction times for the NOM X DAT condition are a result of the optional number of theta-roles (2 or 3). Thus the listener does apply a minimal theta-role assignment strategy. The result that no significant difference was found between the conditions NOM X ACC (canonical) and ACC X ACC (noncanonical) may support the data reported previously: Structural case assignment does not affect verb access. The responses found for the different verb types seem to support this contention: There were no significant reaction time differences between both accusative conditions (NOM X ACC and ACC X NOM) and the nominative condition (NOM X NOM).

5. CONCLUSIONS On-line comprehension of English cleft-sentences supports the multiple lexical access account. However, the evidence that verb particle information does not constrain access to a single verb meaning might be premature, if the frequency of cleft-sentence structures in everyday conversation is considered: The English parser might be not sensitive enough for such idiosyncratic syntactic structures. However, support for the multiple-lexical access account was also found for discontinuous verb access in German. Although the listener encounters only the verb stem at the V2-position, both meanings of the possible discontinuous verbs are exhaustively accessed. However, the multiple-lexical access account was partly falsified in those sentences, in which subjects access discontinuous verb meanings in morphosyntactically incongruent conditions. The German parser did inhibit access to intransitive meanings and dative meanings if a structural case (accusative object) appeared in the Spec position. Otherwise, facilitation of inappropriate verb meanings (intransitive and accusative) is measured, if lexical case (dative) in Spec

Verb Processing in English and German

261

constrained access to the verb stem. The findings on accessing theta-role complexity of different verb types supports the response patterns found for discontinuous verbs: A dative marker in the preverbal position caused an exhaustive access, but the difference between an intransitive verb and an accusative verb did not. The crosslinguistic evidence on German and English sentence processing discussed here reveals that lexical facilitation might not be the only process occurring during initial on-line access to lexical information. Taking into account the reported response patterns, a parser might automatically inhibit meanings excluded by the structural case marker. This leads to the interesting hypothesis that the multiple lexical access mechanism occurs if the inherent properties of a lexical item must be accessed, whereas lexical access is morphosyntactically constrained if structural case determines the type of verb meaning being accessed downstream.

ACKNOWLEDGMENTS Supported by a Feodor-Lynen fellowship (V.3-FLF) of the Alexander von Humboldt Foundation. (Bonn, F. R. G.). The author is grateful to David Swinney for providing access to his research laboratories.

REFERENCES Ahrens, K., and Swinney, D. (1995). Participant roles and the processing of verbs during sentence comprehension. Journal of Psycholinguistic Research, 24 (6), 533-547. Boland, J. (1993). The role of verb argument structure in sentence processing: Distinguishing between syntactic and semantic effects. Journal of Psycholinguistic Research, 22 (2), 133-152. Chomsky, N. (1981). Lectures on government and binding. Dordrecht: Foris. Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MIT-Press. Ford, M., Bresnan, J., and Kaplan, R. (1982). A competence-based theory of syntactic closure. In J. Bresnan (Ed.), The mental representation of grammatical relations. Cambridge, MA: MIT Press. Forster, K. I. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper and E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett. Hillsdale, NJ: Lawrence Erlbaum. Forster, K. I. (1981). Priming and the effects of sentence and lexical contexts on naming time: Evidence for autonomous lexical processing. Quarterly Journal of Experimental Psychology, 33A, 465-495. Frazier, L. (1989). Against lexical generation of syntax. In W. Marslen-Wilson (Ed.), Lexical representation and process. Cambridge, MA: MIT Press. Frazier, L., Flores d'Arcais, G., and Collen, R. (1993). Processing discontinuous words: On the interface between lexical and syntactic processing. Cognition, 47, 219-249.

262

Dieter Hillert

Frazier, L., and Rayner, K. (1982). Making and correcting errors during sentence comprehension. Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Grosjean, F, Dommergues, J.-Y., Cornu, E., Guillelmon, D., and Besson, C. (1994). The gender-marking effect in spoken word recognition. Perception and Psychophysics, 56,590-598. Gurjanov, M., Lukatela, G., Lukatela, K., Savic, M., and Turvey, M. T. (1985). Grammatical priming of inflected nouns by the gender of possessive adjectives. Journal of Experimental Psychology: Learning, Memory, and Cognition, 11 (4), 692-701. Hillert, D., and Bates, E. (1996). Morphological constraints on lexical access: Gender priming in German. Technical Report 9601. Center for research in Language, University of California, San Diego. Hillert, D. (1997). Language in time: Lexical and structural ambiguity resolution. In M. I. Stamenov (Ed.), Language structure, discourse and the access to consciousness. Amsterdam, Philadelphia: Benjamins. Hillert, D. (1998a). Discontinuous words: Access to ambiguous stems of verb particle combinations (submitted). Hillert, D. (1998b). Theta-role complexity effects in semi-configurational language (submitted). MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101 (4), 676-703. Marslen-Wilson, W. D. (1973). Linguistic structure and speech shadowing at very short latencies. Nature, 244, 522-523. Marslen-Wilson, W. D., and Tyler, L. K. (1980). The temporal structure of spoken language understanding. Cognition, 8, 1-72. Marslen-Wilson, W, Brown, C. M., and Tyler, L. K. (1988). Lexical representation in spoken language comprehension. Language and Cognitive Processes, 3, 1-16. Meier, H. (1969). Deutsche Sprachstatistik. Hildesheim: Georg Olms. Nicol, J., Fodor, J. D., and Swinney, D. (1994). Using cross-modal lexical decision tasks to investigate sentence processing. Journal of Experimental Psychology: Language, Memory, and Cognition, 20, 1229-1238. Nicol, J., Swinney, D., Love, T, and Hald, L. (1997). The examination of sentence processing with continuous vs. interrupted presentation paradigms. Unpublished manuscript. Onifer, W, and Swinney, D. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory and Cognition, 9, 225-236. Ruoff, A. (1981). Haufigkeitsworterbuch gesprochener Sprache. Tubingen: Max Niemeyer. Schmauder, A. R. (1991). Argument structure frames. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 49-65. Schmauder, A. R., Kennison, S., and Clifton, C. (1991). On the conditions necessary for obtaining argument structure complexity effects. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17, 1188-1192. Schreuder, R. (1990). Lexical processing of verbs with separable particles. In A. Jongman and A. Lahiri (Eds.), Yearbook of morphology 3. Dordrecht: Foris. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., and Bienkowski, M. (1982). Auto-

Verb Processing in English and German

263

matic access of the meaning of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14,489-537. Shapiro, L.P., Grimshaw, J., and Zurif, E. (1987). Sentence processing and the mental representation of verbs. Cognition, 27, 219-246. Shapiro, L. P., Brookins, B., Gordon, B., and Nagel, N. (1991). Verb effects during sentence processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 17,983-996. Swinney, D. (1979). Lexical access during sentence comprehension: (Re-)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 15, 681-689. Swinney, D. A., Onifer, W., Prather, P., and Hreshkovits, M. (1979). Semantic facilitation across sensory modalities in the processing of individual words and sentences. Memory and Cognition, 1, 165-195. Tabossi, P., Colombo, L., and Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tanenhaus, M. K., Leiman, J. M., and Seidenberg, M. (1979). Evidence for multiple stages in the processing of ambiguous words in syntactic contexts. Journal of Verbal Learning and Verbal Behavior, 18, 427-440.

NOTES 1

Correspondence to: Dieter Hillert, University of California, San Diego, Department of Psychology, La Jolla, CA 92093-0109 ([email protected]).

This page intentionally left blank

THE DIMENSIONAL CONCEPTION OF SPACE AND THE USE OF DIMENSIONAL PREPOSITIONS IN DIFFERENT LANGUAGES THEO HERRMANN JOACHIM GRABOWSKI Department of Psychology University of Mannheim Mannheim, Germany

1. INTRODUCTION There are many communicative contexts in which a speaker's goal, or subgoal, is to emphasize a particular place where an object is (or has been, will be, shall be, has never been, etc.) located in order to point the hearer's attention in the right direction (Herrmann and Grabowski, 1994). Whenever these goals are to be achieved by verbal means and not, for example, by gestures (Levelt, Richardson, and La Heij, 1985), we typically employ prepositional phrases in which the (former, present, or future) position of an object is characterized by its spatial relation to a reference object. We restrict ourselves to the localization of individual objects; however, it is possible to localize object groups or complexly composed objects in the same way. As a result of linguistic analyses, it seems advantageous to assume that spatial prepositional phrases express relations between places (which may be occupied by objects) and not between objects themselves (cf. Grabowski, 1996; Habel, 1989; Klein, 1994). Thus, when we describe spatial relations between objects, we implicitly refer to relations between the places that these objects occupy (or are intended to occupy, even if they never reach there). The basis of a speaker's localizing utterance is his or her cognitive concepSyntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

265

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

266

Theo Herrmann and Joachim Grabowski

tion of a spatial relation. What kinds of spatial relations are conceived in Western cultures? (See Ameka, 1995; Brown, 1994; Levinson, 1991,1992; Svorou, 1994; for further types of spatial cognition.) First of all, topological relations are to be distinguished from dimensional relations (Landau and Jackendoff, 1993; RetzSchmidt, 1988). The essential characteristic of topological relations is that they are determinable solely from the spatial relation between the related objects. Thus, their formal description is a two-place relation with both related objects (or the places they respectively occupy) being the arguments. The basic topological relations are inclusion ('in,' 'inside'), contact ('at'), and proximity ('near'). In order to denote the spatial relation between two objects more specifically, a speaker can refer to the dimensions of space, which are, at this point in time, adopted from our everyday knowledge and which will subsequently be derived from an anthropocentric point of view: the vertical axis ('top-bottom'), the first horizontal axis, or sagittal axis ('front-back'), and the second horizontal axis, or transversal axis ('left-right'). Among the linguistic expressions that describe object relations along these dimensions are, aside from adjectives and adverbs, the dimensional prepositions or prepositional phrases 'over,' 'above,' and 'under'; 'in front of,' 'in back of,' and 'behind'; 'right of and 'left of.' For the following explanations, we restrict ourselves to the use of the dimensional prepositions that refer to the first horizontal axis (i.e., 'in front of and 'behind') and their counterparts in other languages. Given the first horizontal axis and its poles, the basis for the use of 'left' and 'right' can be derived regularly; except for brief references in the corresponding passages, localizing relations on the second horizontal axis will be given only implicit treatment. For the vertical axis, we go no further than the observation that, in most cases, it corresponds to the geocentric vertical line as set by gravity. (See Carlson-Radvandsky and Irwin, 1993; Grabowski, 1996; Levelt, 1984, 1986; for special reflections on vertical relations.) In the next section, it will be shown that 'in front of and 'behind' can be used within different frames of reference depending on the (real or imagined) point of view from which the speaker conceives an object relation; we will derive these uses from a psychological-anthropocentric approach. When interpreting the speaker's utterance, the hearer is faced with the problem of reconstructing the viewpoint that the speaker has taken, because it is, more often than not, not obvious from the linguistic surface of an utterance. In section 3, we will refer to some attempts to explain the conditions under which the prepositions are used in one way or another. These attempts appear insufficient, however, particularly when the prepositional usage is investigated in languages other than English. This will be proven by our experiments in German, Dutch, French, Italian, and (American) English, reported in section 4. In the final section, we will propose a psychological explanation of the observed crosslinguistic differences that relates to different patterns of information processing in a spreading activation network, at the same time rejecting apparent explanations in terms of linguistic rules.

The Use of Dimensional Prepositions

267

2. DIFFERENT CONCEPTIONS OF THE FIRST HORIZONTAL AXIS AND ITS POLES 'In front of and 'behind,' or German vor and hinter, respectively, describe object relations on the first horizontal axis. ('In back of,' which, according to Miller and Johnson-Laird, 1976, is considered a synonym of 'behind,' was never used by the participants of our English production experiments; cf. Grabowski and Miller, 1995.) For the meaningful use of these prepositions, it is necessary to conceive the underlying dimension or, in more general terms, to decide on an entity that defines the spatial dimensions and their poles in a given situation. In the tradition of Biihler (1934), it is the Origo that serves these space-dimensioning functions. In our case, the most important feature is its orientation (i.e., the directions it defines). If the Origo's spatial position is also considered—which at the same time provides the Origo's relative position to the object relation at issue—a viewpoint is defined that is the basis of all dimensional conceptions of spatial contexts (cf. Herrmann, 1996; Schweizer and Janzen, 1996). Thus, dimensional relations between objects are always three-place relations: An object is in a specific spatial relation to a reference object from the space-dimensioning viewpoint of an instantiated Origo. The various uses of dimensional expressions are described and classified in Herrmann's (1990) 6H-model (where H stands for main variants of dimensional localization). His taxonomy comprises six main variants that are again restricted to dimensional relations in the plane. These variants emerge from the combination of two aspects: (a) the speaker, the hearer, or a third object constitutes the Origo; (b) the Origo is, or is not, identical with the reference object (which must, in the first case, be intrinsically directed). This second distinction leads either to two-point localizations or to three-point localizations. According to this approach, for example, "the tree is in front of me (you; the car)" would be described as a speakerrelated (hearer-related; third object-related) two-point localization, whereas "from my point of view (from your point of view; from the station), the hotel is behind the market place" is an example for the corresponding three-point localizations. Obviously, the 6H-model offers a more precise and differentiated account of the usage of dimensional prepositions than the prevalent distinction between deictic and intrinsic (or observer-centered vs. object-centered, respectively) frames of reference (Carlson-Radvansky and Irwin, 1993; Carroll, 1993; Levelt, 1984, 1986). However, we go beyond this taxonomy here and propose a psychologicalanthropocentric approach to the dimensional conception of space, which allows for the derivation of the alternatives in Origo setting from one uniform principle (see Grabowski, 1996, for further details). To achieve this, we start from the basic assumption according to which the dimensions of space, as well as their poles, are generally derived from the body asymmetries of the space-conceiving human being (see as early as Stern, 1936). The initial definition of the vertical line is the axis from the head to the feet, where

268

Theo Herrmann and Joachim Grabowski

the headward direction marks the upper pole and the feetward direction the lower pole. (In case of the canonical erect position of the human body, the anthropocentric and the geocentric vertical lines coincide.) The first horizontal axis runs through chest and back, with the positive pole being in the direction of the privileged perceptive field of the human senses. The second horizontal axis runs perpendicular to both the vertical and the first horizontal line. It is controversial as to whether the poles of this third axis are also based on physical asymmetry. At least, left and right are distinguishable through the feature of handedness and through the location of the unpaired organs. With this body-centered derivation, the point of view on which the dimensional conception of an object relation is based is the half axis that points from the positioned anthropomorphous Origo towards the positive pole of the first horizontal line. First we consider the case of an actual Origo, in which the speaker (in a given situation) sets the Origo egocentrically. This is only possible in situations where the speaker, the localized object and the reference object (or these objects' places, respectively) are temporally and spatially copresent. Given an actual (egocentric) Origo, there are two manners of use of dimensional prepositions that depend on the Origo's relative position to the reference object: Either the Origo is in the place of the reference object, or it is outside the place of the reference object. We call the first case inside perspective, the second outside perspective. 2.1. The Inside Perspective The fundamental characteristic of the inside perspective is that the Origo— which is, for the moment, the speaker in his or her actual position and orientation—is in the same place as the reference object with respect to the relevant horizontal dimensions. This situation can occur when the reference object is a container in which the Origo is located (e.g., inside a car or inside a garment), or when the Origo's location differs from the place of the reference object only vertically (e.g., when sitting on a chair or on a tree stump). In other theoretical contexts, the inside perspective is described such that the Origo induces the reference object front and back (and with it right and left) according to the anthropomorphous pattern, thus considering the reference object as the Origo itself (cf. Herrmann's two-point localizations mentioned above). However, with respect to a general and uniform account of the conception of dimensional object relations, it seems advantageous to carefully separate the space-dimensioning function of the Origo from the spatial anchoring of the localized object by the reference object (see Figure 1). In localization utterances, however, only the localized object and the reference object (and, of course, the spatial relation between the two) is verbally expressed, but not the space-dimensioning entity that is fundamental to the dimensional interpretation of the spatial relation. If a speaker sits on a tree stump and uses his or her actual egocentric anthropomorphous orientation when dimen-

The Use of Dimensional Prepositions

269

Figure 1. The use of dimensional prepositions in the inside perspective.

sionally relating another object to the tree stump, the underlying spatial conception of the utterance "the picnic basket is behind the stump" is as follows: "The picnic basket is in the direction from the tree stump that leads off from my back (or: that is opposite to my actual line of vision)." If a child sits in a car and says, "The ball is in front of the car.", this utterance is based on the conception, "The ball is in the direction from the car that corresponds to my actual line of vision." Note that this case of defining the polar axis through the speaker's actual orientation does not depend on the child's relative orientation to the car: From the speaker's actual Origo in the inside perspective, 'in front of the car' is always in the Origo's line of vision, whether the speaker looks through the windshield or through the rear window. A special case of the inside perspective is when the speaker conceives him- or herself as the reference object of an object relation. Again, however, there is a functional distinction between the place that the speaker occupies as a physical object, and his or her body asymmetries that form the basis for the instantiation of a space-dimensioning Origo.

2.2. The Outside Perspective If the Origo is not in the place of the reference object, as was described above, then it is always turned towards the reference object so that, from the Origo's point of view, the reference object is towards the positive pole of the first horizontal axis. This is an indispensable requirement for the conception of dimensional object relations in all cases that do not come under the inside perspective. If necessary, the speaker must establish this prerequisite by a change of his or her own orientation, before an object relation is interpreted dimensionally. In other theoretical contexts, the outside perspective is described so that the Origo induces the reference object its front on the side that is facing the Origo; accordingly, the opposite side of the reference object is interpreted as its back. After that, the dimensioning of space and the polarization of these dimensions is put down to these

270

Theo Herrmann and Joachim Grabowski

acquired sides of the reference object. This principle is often justified with reference to Clark (1973), who describes the "canonical encounter" (i.e., the face-toface position of speaker and hearer) as the typical spatial pattern of interaction from which the other cases, in which people encounter objects, are derived. However, it appears questionable as to why from the hearer's facing position only the definition of front and back is derived, but not the definition of left and right. This is because the poles of the second horizontal axis correspond, under the outside perspective, again to the anthropomorphous pattern of the Origo, and not to the intrinsic left and right of the hearer on the other side (see Figure 2). These sideinduction approaches have to assume a transposition of the left and right poles with regard to the front-back axis (cf. Wunderlich and Herweg, 1991). (Compare the prepositional patterns on the left-hand sides of Figures 1 and 2.) With a uniform description of the conception of dimensional object relations in mind, we prefer to understand the difference between inside and outside perspective in a way that emphasizes as far as possible what both perspectives share in common: "right of the reference object" and "left of the reference object" are defined in both perspectives through the Origo's anthropomorphous poles on the second horizontal axis. The place "in front of the reference object" is characterized in both perspectives by the fact that it is part of the Origo's perceptual catchment area. (Remember that, in our approach, the Origo is derived from the anthropomorphous attributes of human beings.) In simplified terms, "in front of the reference object" is a place on the first horizontal axis that the Origo can "see"; "behind the reference object" is a place on the first horizontal axis that the Origo can't "see." So far, this pattern holds for both inside perspective and outside perspective. Both perspectives differ regarding the causes that account for the perceptual accessibility (or visibility) of the localized object. From the inside perspective, an object is either visible or (at least partially) invisible, because it is either in the

Figure 2. The use of dimensional prepositions in the outside perspective.

The Use of Dimensional Prepositions

271

Origo's line of vision or behind it (i.e., in the direction towards the negative pole of the first horizontal axis). From the outside perspective, an object is either visible or invisible, because the reference object either admits or obscures its perceptual accessibility. In the latter case, the perceptual overlap of the localized object and the reference object determines the Origo's perceptual access to the localized object. Thus, the main difference between inside and outside perspective is whether the sensory-perceptual equipment of the anthropomorphous Origo, or the position of the reference object compared to the localized object, is considered responsible for perceptual accessibility of the localized object, from the Origo's point of view. Our reconstruction of the usage of dimensional prepositions is solely derived from psychological conditions of human beings when conceiving an object relation dimensionally. The results we generated to reach our objective of providing a general basis for the uses of all four prepositions for relations in the plane under both perspectives seem to correspond to a goal of linguistic semantics which is, through the analysis of the usability of these prepositions, to attain core meanings that are as comprehensive and as extensive as possible (cf. Klein, 1991; Moilanen, 1979; Wunderlich and Herweg, 1991) or, in other theoretical approaches, to attain central prototypes of prepositional usability (Herskovits, 1986; Hottenroth, 1993). However, this coincidence is neither essential nor intended. (See section 5 for further discussion on appropriate psychological explanations.) So far we restricted ourselves to Origo setting via the actual position and orientation of the space-conceiving speaker. However, speakers and hearers are capable of cognitive processes called mental translation and mental rotation (see Graf, 1996; Shepard and Cooper, 1982) (i.e., they can imagine how the spatial world looks from a perspective other than their actual egocentric perspective). In other words, they can mentally project their egocentric Origo onto another position and/or orientation. For this reason, it is possible to dimension the spatial environment from a projected Origo's perspective. As with the actual egocentric Origo, projected Origos involve both position and orientation. Origo projections always follow the anthropomorphous principle, which means that the projections preserve the anthropomorphous pattern of spatial dimensions and their poles. The necessity of the human ability to perfom Origo projections follows from the very fact that the hearer, in order to comprehend the speaker's egocentric localization, must be able to imagine the speaker's point of view in all cases in which both are not co-oriented in space. No wonder the speaker, who is also a hearer in many situations, also employs this ability in the course of language production (Herrmann and Schweizer, 1998). In principle, the speaker can put him- or herself mentally in any orientation and position, thus instantiating a projected Origo. However, in communicative contexts three projection types are the most important: 1. In order to facilitate the hearer's understanding, speakers can conceptualize and describe an object relation from the hearer's point of view. This occurs mainly

272

Theo Herrmann and Joachim Grabowski

for two reasons: when the hearer apparently lacks cognitive or communicative competence, as is the case with children or with people from other language communities, or when the adoption of the hearer's perspective seems to be recommended due to politeness requirements or due to the hearer's social status. The communicative importance of hearer-related localizations is reflected by particular consideration within Herrmann's above-mentioned 6H-model (cf. Herrmann, 1989). 2. When a speaker describes an object relation that is not part of his or her actual surroundings, but is merely imagined or recalled, his or her actual orientation is of no use as a space-dimensioning Origo. In such cases, it is possible to explicitly introduce a fictitious Origo. In route directions, for example, we hear such utterances as, "when you leave central station, the hotel is right behind the conspicuous Art Nouveau building." By the phrase "when you leave central station," a space-dimensioning Origo is introduced with respect to both position and orientation. Again, fictitious Origos are always anthropomorphous Origos; an important means of Origo setting in imagined space is the generic wanderer, like the "you" in the above example, which means a generic person and not specifically the addressee (Ehrich, 1989; Herrmann, 1996). 3. Several objects are formed in such a way that people take a particular position and orientation towards these objects when using them. For such objects, asymmetry on at least one dimension is required (presupposed that the object is in its normal vertical orientation). The typical position and orientation of people when handling objects of this kind can serve to define a canonical Origo's point of view, which in turn serves as a space-dimensioning entity projected by the speaker. Again, an inside perspective and an outside perspective exist. Among reference objects that are typically used from the inside perspective are vehicles, garments, conference rooms, theaters, and seats. Among reference objects that are typically used from the outside perspective—where the canonical user is better referred to as a canonical observer—are grandfather clocks, closets, and TV sets (cf. the distinction between "vehicle objects" and "facing objects" in Herrmann, 1990.) In psycholinguistic approaches, these objects are frequently described as having intrinsic (or inherent) fronts, backs, lefts, and rights that define spatial dimensions and their poles by themselves. Among the features used to define an object's front are the typical direction of motion, the location of "eye-like" object parts, or the differentiatedness of the surface (Fillmore, 1982; Miller and Johnson-Laird, 1976; Retz-Schmidt, 1988). In these approaches, the relation between the localized object and the reference object may be conceptualized as a three-place relation where, however, the intrinsically directed reference object simultaneously plays the role of the Origo (Herrmann, 1995). In this view, the familiar problem arises that "facing objects" need the poles of the left-right axis to be interchanged, compared to the anthropomorphous pattern of poles (see Figures 1 and 2). More-

The Use of Dimensional Prepositions

273

over—as can easily be concluded from our previous argumentation—it appears, even in the case of intrinsically directed reference objects, advantageous with respect to a general description of the conception of dimensional object relations, to not disregard the canonical, anthropomorphous interactor who defines the prerequisites for the dimensional conception of the object relation at issue. From a psychological-anthropocentric starting point, the intrinsic fronts of objects are merely derived from the typical orientation of the canonical user or observer, respectively. The "true" space-dimensioning principle is not the object itself, but the canonical anthropomorphous Origo. (This opinion does not necessarily suspend the possibility that the canonical viewpoint towards an object can be a regular part of this object's mental representation; cf. Marr, 1982; Schweizer, 1997). For the sake of simplicity, and in order to more easily follow the elsewhere propagated view of objects for which a canonical viewpoint exists, we will call these objects (intrinsically) directed objects; we frequently pointed to the derivation of this intrinsic directedness from the anthropomorphous qualities of the canonical Origo. At the end of the day, the most important cases of the dimensional conception of object relations vary according to two parameters: actual (i.e., nonprojected, egocentric) Origo vs. canonical (or, more precisely, canonically projected) Origo, and inside perspective vs. outside perspective. Four cases of the dimensional conception of object relations emerge from the combination of values from both parameters: 1. Actual Origo in the inside perspective: The speaker and, with him or her, the actual Origo are in the place of the reference object (e.g., "The picnic basket is in front of the tree stump," when the speaker is sitting on the stump). 2. Actual Origo in the outside perspective: The speaker and, with him or her, the actual Origo are not in the place of the reference object, but facing it (e.g., "There is a mushroom in front of the tree stump," when the speaker looks at the stump from a forest path. This case is traditionally called the deictic perspective). 3. Canonical Origo in the inside perspective: The canonical position and orientation of the reference object's user, which is inside the object, becomes the projected Origo. Objects being typically used from inside are called vehicle objects. (This case is traditionally called the intrinsic perspective.) 4. Canonical Origo in the outside perspective: The canonical position and orientation of the reference object's user, or observer, which is outside the object, becomes the Origo. Objects being typically used from outside are called facing objects. A relevant difference between actual and canonical Origo settings is the fact that the canonical Origo's position can be inferred from the knowledge of the position and orientation of the reference object alone, whereas the actual Origo

274

Theo Herrmann and Joachim Grabowski

may be in any position relative to the position of the reference object (as long as the Origo is facing the reference object). The latter aspect does also hold for an Origo projected into the hearer's actual position and orientation. Thus, there is an advantage when using canonical Origos, because the hearer can know their point of view without knowing the speaker's actual (and transitory) position and orientation. Particularly when the speaker and the hearer do not have a common perceptual field in a communicative situation (i.e., when they can't see each other— e.g., on the telephone), the speaker will preferably conceive a dimensional object relation from the canonical Origo's point of view. The communicatively successful use of canonical Origos for the dimensional conception of object relations requires, however, two things: people in the respective language community must agree in their knowledge about the typical interaction with certain classes of objects, and for each object class, one canonical position of use alone may be defined (if a canonical Origo is defined at all). The latter criterion does not apply in general (e.g., for meeting places like churches and theaters, both a canonical inside perspective and a canonical outside perspective is defined). From the inside perspective, the canonical front is the stage (or the altar, respectively); from the outside perspective, the canonical front is the entrance or the portal. Thus, the insides's canonical 'back' is the outside's canonical 'front', and vice versa. This means that conflicts of interpretation when trying to comprehend dimensional localization utterances may arise even within the use of canonical Origos (Carroll, 1993). However, the most frequently investigated case up to now relates to the conflict of interpretation, or ambiguity, between the speaker's actual Origo in the outside perspective and the canonical inside perspective. This conflict may arise in cases in which a canonical inside perspective is defined for a reference object, and in which the speaker and the hearer are in a sensorily mutually accessible situation, where it would also be possible, and useful, to conceive the situation from the speaker's actual Origo. Localization utterances, as produced by the speaker, usually only express the relationship between the localized object and the reference object, but do not explain the underlying Origo (although this would be possible by adding a phrase like "from my point of view ...") Now, if the canonical Origo inside the reference object and the actual Origo outside the reference object are equally oriented (i.e., the positive poles of their first horizontal axis point in the same direction), the hearer will interpret "in front of" and "behind" exactly the other way around, according to which of both alternative Origos determines his or her understanding of the localization utterance. Figure 3 is an illustration of this interpretation conflict. We already mentioned that comparison of both cases is treated elsewhere in terms of the deictic versus intrinsic frame of reference (Levelt, 1984,1986), or the obsever-centered versus object-centered frame of reference (Carlson-Radvansky

The Use of Dimensional Prepositions

275

Figure 3. From the woman's actual point of view (= outside perspective), the panda is in front of the car, and the kangaroo is behind the car. From the canonical driver's point of view (= inside perspective), the kangaroo is in front of the car, and the panda is behind it. (Note that in this situation, "right" and "left" do not differ in both perspectives.)

and Irwin, 1993), where the deictic, or observer-centered, frame corresponds to the speaker's actual outside perspective, whereas the canonical perspectives are combined in the intrinsic, or object-centered, frame of reference. However, it should have become clear from the above considerations that the sole distinction between deictic and intrinsic localization is not sufficient for a differentiated description and treatment of the varieties occurring in dimensional localization (see Herrmann, 1990; Grabowski, 1996; for further critical arguments). It is also worth noting that the concept of deictic versus intrinsic frames of reference were developed from semantic analysis of linguistic expressions, whereas our approach relates to the analysis of human conception of dimensional object relations (which does not necessarily find itself expressed verbally); thus, both approaches appear in quite different theoretical contexts. However, interpretation alternatives with the use of "in front of" and "behind" are phenomena in the pretheoretical world; they may—despite their different explanation—occur in both contexts of derivation.

3. SOME DETERMINANTS OF ORIGO SETTING So far, we have discussed theoretical issues. We now turn to an empirical question: What are the conditions that make the hearer interpret localization utterances that contain "in front of" and "behind" from either an actual or a canonical point of view, given an intrinsically directed reference object? Without claiming completeness, we will outline four approaches that appear to be the most relevant to offer an answer to this question (see also Grabowski, 1994; Grabowski and WeiB, 1996a,b). For both actual and canonical Origo setting, there are linguistic means for its unambiguous indication. Expressions such as, "from my point of view" or "when you leave central station" almost exclusively involve an actual outside

276

Theo Herrmann and Joachim Grabowski

perspective. Phrases such as, "on the left hand side" or "at the front of" are unambiguously based on a canonical Origo's dimensional definition. However, in everyday communication (as well as in respective language production experiments) people refrain almost without exception from using these means (Grabowski and Miller, 1995). Most often, we find simple utterances like "X is in front of/behind Y," where alternative interpretations are possible. According to Miller and Johnson-Laird (1976; see also Abkarian, 1982), canonical perspective generally dominates the speaker's actual point of view, when the reference object is conceived as inherently directed, thus defining a canonical orientation. We will show in the next section that this assumption does not lead to accurate predictions, at least for German and Dutch, and does not sufficiently describe the situation in English. Ehrich (1985) examines room descriptions and shows that the integration of separate groups of objects into more extensive spatial arrangements is facilitated when a strategy of Origo setting is kept up through the entire discourse (cf. Linde and Labov, 1975). However, spatial expressions are not always embedded in an extensive spatial discourse, as is the case with room descriptions or directions; they sometimes appear separately, or in a nonspatial context, where the hearer cannot refer to a previously chosen Origo when interpreting a verbal localization. Wunderlich (1981) points out that the static versus dynamic character of a situation in which an object is localized—expressed by the use of static or dynamic verbs—influences the dimensional interpretation of spatial relations between objects. Interpretation from an actual Origo's point of view is preferred for spatial phrases that include motion verbs (e.g., "The red car drives behind the black one"), whereas a canonical interpretion is preferred for spatial expressions that include static verbs (e.g., "The red car is behind the black one."). However, there is no deterministic correspondence between the chosen Origo and the static versus dynamic factor; Wunderlich merely found correspondences that range from 70 to 80%. Levelt (1982) points to individual styles and preferences in the choice of one's space-dimensioning Origo. However, in our own data, inconsistent interpretations arise both between and within speakers, thus empirically rejecting the assumption of stable individual styles (Grabowski, 1996). There are further theoretical and methodical objections to the assumption of cognitive styles being responsible for the Origo setting, which we critically discussed elsewhere (Grabowski, 1994; Grabowski and WeiB, 1996a,b; Korek, 1997). Each of the mentioned approaches, as well as further studies that we will not mention here, consider one aspect that contributes to the prediction of the interpretation of localizing expressions from an actual versus canonical Origo's point of view. However, even taken together, they do not allow for a comprehensive and satisfactory answer of how dimensional prepositions are used, particularly when the space-dimensioning patterns are investigated crosslinguistically. In the next

The Use of Dimensional Prepositions

277

section, we will introduce the factors that we found to be decisive for the hearers' Origo setting in potentially ambiguous situations, and report on the results yielded through systematic experimentation.

4. EXPERIMENTS ON THE INTERPRETATION OF "IN FRONT OF" AND "BEHIND" AND THEIR COUNTERPARTS IN GERMAN, DUTCH, ITALIAN, AND FRENCH Consider the following example (Figure 4), also refered to by Vater (1996) and by Ruhrberg and Rutz (1990), in order to illustrate an alternatively conceivable situation; we used this example for the experiments reported below (see Grabowski, 1994,1996; Grabowski, Herrmann and WeiB, 1993; Grabowski and WeiB, 1996a,b; for further details.) A driver and a front-seat passenger are driving in a car in the right lane of the road. At a visible distance ahead, there are parking spots on the right-hand side of the road. In the center spot there is an object. The passenger asks the driver to stop the car either "in front of" or "behind" this object; thus, the object in the center spot serves as the reference object in the localizing utterance. The following explanations hold for all experimental variants that will be reported: Participants were positioned in front of the setting illustrated in Figure 4; this setting was approximately 60 x 40 cm in size and was constructed from toy material. Participants were orally instructed by the experimenter; then they moved the car that, according to the instructions, held the speaker (i.e., the passenger)

Figure 4. A scenario for the experimental investigation of the usage of "in front of" and "behind."

278

Theo Herrmann and Joachim Grabowski

and the hearer (i.e., the driver) to the spot that corresponds to their interpretation of the spatial utterance given in the instruction (i.e., either to spot #1 or to spot #3). Experimentation was conducted with one participant at a time; only one measurement (the chosen spot in one situation after one utterance) was obtained from each participant in order to avoid serial effects and to maintain the participants' behavioral naturalness; they thought they were participating in an experiment on traffic behavior. Which factors could, in this situation, determine the hearers' interpretation of "in front of" and "behind" on the basis of their dimensional conception of the situation? A first influential factor is, of course, the preposition used. From a naive understanding of the "meaning" of both prepositions alone, it is to expected that in any situation "in front of" is interpreted differently from "behind." The second assumed factor of influence is the reference object being intrinsically directed (D+) or not (D—). With a car as the reference object (= D+), a canonical Origo is defined that can serve as a space-dimensioning alternative to the actual Origo (which speaker and hearer share in common, as they are copositioned and cooriented in the driving car). In our experiments, the intrinsically directed reference object was instantiated by a yellow Volkswagen beetle. If spot #1 is chosen after hearing "in front of the yellow beetle," or, respectively, spot #3 after hearing "behind the yellow beetle," then the interpretation is based on the actual Origo's perspective. On the other hand, if spot #3 is chosen after hearing "in front of," or spot #1 after hearing "behind," we can infer that the canonical Origo defines the hearer's dimensional conception of the situation. With a tree as a reference object (= D—), no such canonical Origo is defined. In this case, nothing but the interpretation from the actual Origo's point of view should occur. Anticipating the results, it is also possible to interpret "in front of the tree" as spot #3 and "behind the tree" as spot #1 in the given traffic situation. This can be put down to an 'extrinsic' Origo projection, which is to say that, according to the overall directedness of the street in the direction of the traffic, all objects can be conceived as if they were lined up with the moving car. A third influential factor was detected in our previous experiments when we varied, with the traffic context remaining the same, the social situation in which the localizing utterance was produced and interpreted. Here, the crucial aspect appears to be the officiality of the speaker-hearer interaction, being either high (O+) or low (O—). In case of low officiality, we set a private, informal discourse situation (= O—), where the hearer was instructed as follows ((= -A) English version): "Imagine you are giving a friend a lift home. But you don't know exactly where he lives. He says, 'Could you please drop me off in front of/behind the yellow beetle/the tree?' Where would you stop your car?" In case of high officiality, we set an institutionalized, formal discourse situation (= O+), where the hearer was instructed that he or she was taking his or her driving test, receiving

The Use of Dimensional Prepositions

279

instructions by the examiner to park either in front of or behind the reference object. After a complete combination of these factors, we arrived at eight experimental conditions: used preposition (= +A) (in front of versus behind) X directedness of the reference object (yellow beetle versus tree) X officiality of the situation (lift home versus driving test). Starting with the German series, where the prepositions vor ('in front of) and hinter ('behind') were used in the localization phrases, Figure 5 shows the results that we received from an overall number of 239 participants. As can be seen from Figure 5, there are two communicative situations in which the hearers interpret the prepositions of the first horizontal dimension largely consistently. The left-most bars indicate that, when an intrinsically directed reference object is given in the formal driving test situation, interpretations are predominantly based on the canonical Origo's perspective, whereas the right-most

Figure 5. Percentage of spot #1 interpretations in four experimental situations, each being instructed with either vor ('in front of) or hinter ('behind'), for the German sample. D+ = directed reference object: yellow beetle; D- = nondirected reference object: tree; O+ = high officiality: driving test; O- = low officiality: lift home. You get the percentage of spot #3 interpretations by complementing the bars to 100%.

280

Theo Herrmann and Joachim Grabowski

bars indicate that, when a nondirected reference object is given in the informal lift home situation, interpretations are predominantly based on a dimensional conception of the situation that arises from the actual (or egocentric) Origo's point of view. The pattern of results for the two situations in the middle of Figure 5, however, show largely inconsistent interpretations; vor and hinter are assigned to both spots almost by chance. In order to guarantee communicative success, however, the two prepositions must differentiate between both parking spots in a given situation. In the language of the diagram in Figure 5, this means that for each pair of bars, one must be ideally at 100% while the other bar shows 0%, when speaker and hearer shall be enabled to perfectly communicate on the intended spot. (Note that the reported data are summarized across several experimental replications with different experimentors and independent samples, each of them yielding the same pattern. Thus, these results are highly reliable.) Altogether, we need a threefold interaction of the preposition used, the characteristics of the reference object, and the type of social interaction between speaker and hearer, in order to predict the hearers' Origo setting in two situations out of four. If we don't want to leave it at that, we need to find further determinants that explain the inconsistent interpretation patterns in the D+/O— and D— /O+ situations as well. To achieve this, we turned to a crosslinguistic perspective. The train of thought is as follows: In German, the preposition that denotes the positive pole of the first horizontal axis, vor, is morphologically identical with the preposition for the unmarked pole of the temporal axis. In other words, German vor translates both as 'in front of and as 'before.' Spatial dynamic situations, like the traffic situation at issue, may also be conceived temporally (cf. Feagens, 1980; Johnston and Slobin, 1979; Traugott, 1975). Spot #1 is, from an actual Origo's perspective, not only vor ('in front of) the reference object, but it also "comes" vor ('before') the reference object when driving along the road. Thus, at least some part of the observed inconsistencies could be put down to the fact that German hearers conceive the experimental situation temporally (although it is clearly a spatial comprehension task), which leads to an interpretation pattern that corresponds to the actual Origo's perspective when we look at it in terms of a spatial interpretation. Thus, the results for the situations D+/O— and D—/O+, where the prepositions' interpretations appear somewhat arbitrary, would reflect a conflict between a spatial and a temporal conception of the situation rather than between an actual and a canonical Origo. How can we test these assumptions? Languages differ with respect to their prepositional inventory (i.e., to the number of different prepositional forms they provide for expressing spatial relations on the first horizontal axis) and for temporal relations. More precisely, the interesting difference between languages is whether they have spatial prepositions that are not identical with temporal prepo-

The Use of Dimensional Prepositions

281

sitions. (Note that our aim is neither to analyze exhaustively the meaning potential of these prepositions, nor do we touch on the fact that temporal metaphors may generally be used when expressing spatial relations. In the context at issue, the only interesting feature of languages is whether they have spatial (dimensional) prepositions that cannot be interpreted in temporal sense.) English can be considered a four-preposition language, providing "in front of" and "behind" as well as "before" and "after." The same situation is given in French (devant—derriere as opposed to avant—apres) and in Italian (davanti— dietro for spatial relations, prima—dopo for temporal relations). In contrast, German can be called a three-preposition language, with its aforementioned prepositional antonyms vor and hinter, and the temporal prepositions vor and nach. The same prepositional structure holds for Dutch, where the prepositional pairs are voor—achter and voor—na. If the dimensional interpretation that we reported from the German experiments is anything to do with the prepositional inventory of a language, then we would expect similar results when the same experiments were conducted with Dutch speakers, whereas the results obtained from French, Italian, and English experiments should yield different patterns of interpretation. It is worth noting that this is not an issue of the different prepositional usage among individual languages, but involves the introduction of a—very specific—factor in the realm of linguistic typology. The results obtained from corresponding experiments in the above-mentioned languages are shown in Figures 6 (for Dutch; N = 102), 7 (for Italian; N = 74), 8 (for French; N = 88), and 9 (for American English; N = 100). Participants were native speakers who were tested at universities in the respective countries (Nijmegen, Netherlands; Verona, Italy; Dijon, France; Princeton, NJ, USA). It appears that within the three-preposition languages, as well as within the four-preposition languages examined, the obtained result patterns are comparatively similar, whereas they substantially differ between the two types of languages. (See Grabowski and WeiB, 1996a,b for a more detailed analysis.) The results can be described as follows: 1. In all five languages there are situations in which the two prepositions denoting spatial relations on the first horizontal axis do, or do not, differentiate spot #1 from spot #3. Thus, none of these languages is, on the whole, superior with respect to successful communication about dimensional relations. 2. In the given type of traffic situation, the actual (egocentric) Origo's space-dimensioning (and polarization) pattern, and the canonical Origo's spacedimensioning pattern show roughly the same probabilities for forming the basis of the hearers' interpretations of the two dimensional prepositions. This holds true as long as we refrain from considering at least two more specific factors and their

Figure 6. Percentage of spot #1 interpretations for the Dutch sample. The diagrams in Figures 6-9 are explained with Figure 5.

Figure 7.

Percentage of spot #1 interpretations for the Italian sample. 282

Figure 8. Percentage of spot #1 interpretations for the French sample.

Figure 9.

Percentage of spot #1 interpretations for the American English sample. 283

284

Theo Herrmann and Joachim Grabowski

interaction: intrinsic directedness of the reference object and officiality of the discourse situation. Again, this observation is valid for all investigated languages. 3. With four-preposition languages, the decisive factor appears to be whether the reference object is intrinsically directed or not. When the reference object is directed, prepositions are consistently interpreted from the canonical Origo's point of view, whereas hearers do not agree in their dimensional interpretations when the reference object lacks a canonical Origo. However, the officiality factor also comes into play, although with subordinate strength, as it changes the interpretation patterns when the reference object is a tree (compare the bars for D—/O+ and for D—/O— in the right half of Figures 7, 8, and 9). 4. With three-preposition languages, both the directedness factor and the officiality factor appear to influence the prepositions' interpretations equally strongly. With four-preposition languages, hearers almost perfectly agree upon a canonical interpretation when the reference object is directed, and almost completely lack consistent understanding when the reference object is nondirected. This, we need to consider the interaction of both influencing factors from the very beginning in order to describe the hearers' interpretation patterns in three-preposition languages. Here, it seems that a highly official discourse situation prefers a directed reference object that allows for a distant, "objective" point of view, namely, the canonical perspective, whereas a private, informal discourse situation is better managed from the actor's egocentric point of view, namely the actual Origo, which is without competition when the reference object is nondirected.

5. A PSYCHOLINGUISTIC EXPLANATION OF CROSSLINGUISTIC DIFFERENCES How can the results from section 4 be explained from a psychological perspective? Here, we will briefly go further afield, based on the above-mentioned findings, in order to comment on the problem of appropriate theories on human linguistic behavior within the psychology of language. For some parts of the recent psychology of language (in contrast to cognitively oriented trends in linguistics which are nevertheless aimed at the proper explanation of linguistic structures), a considerable theoretical shortcoming can be observed. These approaches are content with apparent explanations, subject to what we call the Structuralist Misconclusion (cf. Herrmann and Graf, 1996): Linguistic output, as produced by human speakers, can often be described by indication of structures of constitutive or regulatory rules (as they are given, for example, by grammars). However, it is concluded that the system that generated the observed linguistic output did so on the basis of the representation of these rules and their employment.This is to say that the fact that the output may be described by rule structures falsely implies

The Use of Dimensional Prepositions

285

that the output generation can be explained by processes operating on these rules, which are internally represented. Sometimes it is even argued that we need rulebased models of language comprehension and production because it is possible to descriptively reconstruct linguistic outputs by means of structural rules (e.g., a grammar) (cf. Hacker, 1987). Or it is assumed that grammar is "translated into action" through a person's linguistic behavior (Buhler, 1934). The structuralist misconclusion is characterized by the opinion that the processes underlying linguistic behavior are to be deductively concluded from the rules that decribe the behavioral output. Consider that this comment does not impact the empirical question as to whether the production and comprehension of linguistic utterances, given a pattern of empirical results, can be theoretically reconstructed as the employment of any kind of rules at all. However, such internally represented rules do not necessarily correspond to those rules employed for the description of the respective linguistic output. What do these remarks mean with respect to our crosslinguistic findings? In order to interpret our results, one might enlist the following kind of linguistic usage rules: "The spatial use of the preposition that denotes the positive pole of the first horizontal axis in informal situations with reference to non-intrinsic objects, is to be based on the speaker's egocentric point of view." This rule would then belong to the German and Dutch linguistic systems, but not to the Italian, French, and English ones. Falling into the trap of the structuralist misconclusion, one might even arrive at the following conclusion: This usage rule, which is valid in German (or invalid in English) determines the linguist behavior of German (and English) speakers and hearers. Thus, the differences between the German and the English hearers' comprehension patterns would appear to be explained. In contrast to this theoretically insufficient strategy of explanation, we can first assume that speakers of three-preposition languages, as opposed to speakers of four-preposition languages, have generated, during their acquisition of dimensional prepositions within the acquisition of language in general, different patterns of association strengths between certain concepts and certain words (or word forms). Thus, speakers of three-preposition languages (like German and Dutch) have learned, during ontogenesis, to name both the concept of a particular spatial relation, and the concept of a particular temporal relation by one and the same word, vor (or voor, respectively). Altogether they distinguish between only three prepositions for four concepts. On the other hand, speakers of four-preposition languages (like English, to give just one example) have learned to call the particular spatial concept "in front of" and the particular temporal concept "before." Altogether they distinguish four prepositions for the relations at issue. Then, it is to be assumed that, in languages that employ a twofold use of vor (or its counterparts), the hearing (or reading) of vor triggers space-related as well as time-related comprehension processes, whereas the perception of "in front of" only triggers space-related, but not time-related comprehension processes.

286

Theo Herrmann and Joachim Grabowski

If it is further assumed that the two influential factors "directedness" and "officiality" interact differently in both prepositional language types, the processes underlying our results can be illustrated, in simplified terms, in a diagram that employs different strengths of association between nodes in (cognitive) networks, without requiring a rule-based acccount of information processing. Figure 10 shows such diagrams for German (left) and English (right) speakers, representing the Dutch and, respectively, the Italian and French conditions as well. The differences between speakers of three-preposition and four-preposition languages that relate to our results for the D+/O+ and D+/O— situations (represented on the left-hand side of Figures 5 to 9) may be reconstructed more precisely than is achieved by our diagrams, by assuming different activation patterns between the input nodes that stand for directedness (D+) and officiality (O+). Thus, the influence of the directedness factor dominates the influence of the officiality factor more strongly, and in a different manner in four-preposition languages compared to three-preposition languages, as indicated by the inhibition arrow from the D+ onto the O+ node. The more interesting differences, however, relate to the results in the D—/O+ and D—/O— situations (represented on the right-hand side of Figures 5 to 9). These results are probably only explicable when assuming different association patterns between the learned prepositions. Let us start with three-preposition languages, represented in the left diagram of Figure 10. If a German hearer perceives the preposition vor, both the spatial (S) and the temporal (T) path of information

Figure 10. An activation network for the interpretation of dimensional prepositions. Left diagram: three-preposition languages, exemplified by German; right diagram: four-preposition languages, exemplified by English. S = spatial information-processing pathway; T = temporal informationprocessing pathway; " + " = excitatory influence; "-" = inhibitory influence.

The Use of Dimensional Prepositions

287

processing are triggered. According to Figure 10, the path for spatial processing, however, effects the same excitatory change of activation on both spot #1 and spot #3; neither of the two spots receives stronger activation than the other. The temporal path is only connected to spot #1, because in temporal respect, vor can be interpreted as exclusively egocentric. Thus, spot #1 and not spot #3 receives additional activation from the temporal path. Because both spots are connected to one another through lateral inhibition, the stronger activated spot, which is now spot #1, de-activates (inhibits) spot #3. Therefore, it becomes clear that in the situation D—/O—, where neither D+ nor O+ exerts influence on the process, spot #1 receives the highest activation from the processing of the preposition vor, thus entering further processes of behavior planning (e.g., moving the car to a particular place). With speakers of four-preposition languages (right diagram of Figure 10), "in front of" only triggers the spatial pathway. Through this both spots receive equal activation. The temporal path would only have been activated if "before" had been received. Neither spot is preferred through stronger activation, to enter the subsequent processes; the individual behavior becomes arbitrary, leading to a lack of interindividual consent. Similar explanations for the processing of hinter and "behind" may easily be derived. Finally, we want to point to the possibility of empirically testing our proposed interpretation. According to our activation-related assumptions, the word nach should be a good prime for the target hinter when studying native speakers of German, whereas a much weaker priming effect is to be expected from the prime word after onto the target behind when studying English speakers.

ACKNOWLEDGMENTS Institutional support for the experiments in languages other than German was provided by Wolfgang Klein and Ursula Brinkmann (Nijmegen, Netherlands), Bianca di Bernardi (Verona, Italy), Michel Fayol and Beatrice Bourdin (Dijon, France), and George Miller and Christiane Fellbaum (Princeton, USA). We are grateful to Irene ter Haar, Emanuela Antolini, Alexa Heintze, Shari Landes, and Petra WeiB, who conducted the experiments in their mother tongues.

REFERENCES Abkarian, G. G. (1982). Comprehension of deictic locatives: The object "behind" it. Journal of Psycholinguistic Research, 11, 229-245.

288

Theo Herrmann and Joachim Grabowski

Ameka, F. K. (1995). The linguistic construction of space in Ewe. Cognitive Linguistics, 6,139-181. Brown, P. (1994). The INs and ONs of Tzeltal locative expressions: The semantics of static descriptions of location. Linguistics, 32,743-790. Biihler, K. (1934). Sprachtheorie: Die Darstellungsfunktion der Sprache. [Linguistic theory: The symbol function of language.] Jena: Fischer. Carlson-Radvansky, L. A., and Irwin, D. E. (1993). Frames of reference in vision and language: Where is above? Cognition, 46, 223-244. Carroll, M. (1993). Deictic and intrinsic orientation in spatial descriptions: A comparison between English and German. In J. Altarriba (Ed.), Cognition and culture: A crosscultural approach to Cognitive Psychology (pp. 23-44). Amsterdam: Elsevier. Clark, H. H. (1973). Space, time, semantics, and the child. In T. E. Morre (Ed.), Cognitive development and the acquisition of language (pp. 23-63). New York: Academic Press. Ehrich, V. (1985). Zur Linguistik und Psycholinguistik der sekunddren Raumdeixis. [On the linguistics and psycholinguistics of secondary spatial deixis.] In H. Schweizer (Ed.), Sprache und Raum: Psychologische und linguistische Aspekte der Aneignung und Verarbeitung von Rdumlichkeit (pp. 130-161). Stuttgart: Metzler. Ehrich, V. (1989). Die temporale Festlegung lokaler Referenz. [The temporal establishing of spatial reference.] In Ch. Habel, M. Herweg, and K. Rehkamper (Eds.), Raumkonzepte in Verstehensprozessen: Interdisziplindre Beitrage zu Sprache und Raum (pp. 1-16). Tubingen: Niemeyer. Feagans, L. (1980). How to make sense of temporal/spatial "before" and "after." Journal of Child Language, 7, 529-537. Fillmore, C. J. (1982). Towards a descriptive framework for spatial deixis. In R. Jarvella, and W. Klein (Eds.), Speech, place, and action. Studies in deixis and related topics (pp. 31-59). Chichester: Wiley. Grabowski, J. (1994). Kommunikative Unschdrfen—Zur Rezeption und Produktion von Richtungsprdpositionen am Beispiel von "vor" und "hinter." [Communicative fuzziness—on the comprehension and production of dimensional prepositions, illustrated by German vor and hinter.] In H.-J. Kornadt, J. Grabowski, and R. Mangold-Allwinn (Eds.), Sprache und Kognition: Perspektiven moderner Sprachpsychologie (pp. 183208). Heidelberg: Spektrum Akademischer Verlag. Grabowski, J. (1996). Die Raumauffassung des Menschen und ihrAusdruck in der Sprache: Determinanten der Verwendung dimensionaler Prapositionen. [Human spatial conception and its linguistic expression: Determinants of the use of dimensional prepositions.] Unpublished habilitation thesis, University of Mannheim. Grabowski, J., Herrmann, T, and WeiB, P. (1993). Wenn "vor" gleich "hinter" ist— zur multiplen Determination des Verstehens von Richtungsprdpositionen. [When vor equals hinter—on the multiple determination of the comprehension of dimensional prepositions.] Kognitionswissenschaft, 3, 171-183. Grabowski, J., and Miller, G. A. (1995). Factors affecting production and comprehension of spatial prepositions: A comparison of German and American English (Research Group on Language and Cognition, report no. 56). Mannheim: University, Chair for Psychology III.

The Use of Dimensional Prepositions

289

Grabowski, J., and WeiB, P. (1996a). The prepositional inventory of languages: A factor that affects comprehension of spatial prepositions. Language Sciences, 18, 19-35. Grabowski, J., and WeiB, P. (1996b). Determinanten der Interpretation dimensionaler Lokalisationsauflerungen: Experimente in funf Sprachen. [Determinants of the interpretation of dimensional localizing utterances: Experiments in five languages.] Sprache und Kognition, 15,234-250. Graf, R. (1996). Mentale Rotation und Blickpunkttransformation. [Mental rotation and viewpoint transformation.] Sprache und Kognition, 15, 178-202. Habel, C. (1989). 'zwischen'-Bericht. [A report on zwischen.] In C. Habel, M. Herweg, and K. Rehkamper (Eds.), Raumkonzepte in Verstehensprozessen: Interdisziplindre Beitrdge zu Sprache undRaum (pp. 37-69). Tubingen: Niemeyer. Hacker, P. (1987). Language, minds, and brains. In C. Blakemore, and S. Greenfield (Eds.), Mindwaves: Thoughts on intelligence, identity and consciousness (pp. 122-127). Oxford: Oxford University Press. Herrmann, T. (1989). Sprachpsychologische Beitrage zur Partnerbezogenheit des Sprechens. [Psycholinguistic contributions to the partner-relatedness of speech.] In H. Scherer (Ed.), Sprache in Situation. Eine Zwischenbilanz (pp. 179-204). Bonn: Romanistischer Verlag. Herrmann, T. (1990). Vor, hinter, rechts und links: Das 6H-Modell. Psychologische Studien zum sprachlichen Lokalisieren. [In front of, behind, right and left: The 6H-model. Psychological studies into verbal localization.] Zeitschrift fur Literaturwissenschaft und Linguistik, 78, 117-140. Herrmann, T. (1995). Geddchtnis und Sprache. Zum Problem des Blickpunktes. [Memory and language. On the viewpoint problem.] In D. Dorner and E. van der Meer (Eds.), Das Geddchtnis. Problems—Trends—Perspektiven (pp. 199-226). Gottingen: Hogrefe. Herrmann, T. (1996). Blickpunkte und Blickpunktsequenzen. [Points of view and point of view sequences.] Sprache und Kognition, 15, 159-177. Herrmann, T., and Grabowski, J. (1994). Sprechen—Psychologic der Sprachproduktion. [Speaking—The psychology of language production.] Heidelberg: Spektrum Akademischer Verlag. Herrmann, T., and Graf, R. (1996). Konzeptuelles und semantisches Wissen aus psychologischer Sicht [Conceptual and semantic knowledge from a psychological perspective] (Research Group on Language and Cognition, report no. 61.) Mannheim: University, Chair for Psychology III. Herrmann, T., and Schweizer, K. (1998). Sprechen uber Raum. Das sprachliche Lokalisieren und seine kognitiven Grundlagen. [Speaking about space. Verbal localization and its cognitive basis.] Bern: Huber. Herskovits, A. (1986). Language and spatial cognition. Cambridge, MA: Cambridge University Press. Hottenroth, P.-M. (1993). Prepositions and object concepts: A contribution to cognitive semantics. In C. Zelinsky-Wibbelt (Ed.), The semantics of prepositions (pp. 179220). Berlin: Mouton de Gruyter. Johnston, J. R., and Slobin, D. I. (1979). The development of locative expressions in English, Italian, Serbo-Croatian, and Turkish. Journal of Child Language, 6, 529-545.

290

Theo Herrmann and Joachim Grabowski

Klein, W. (1991). Raumausdriicke. [Spatial expressions.] Linguistische Berichte, 132, 77-114. Klein, W. (1994). Keine Kanguruhs zur Unken—Uber die Variabilitdt von Raumvorstellungen und ihren Ausdruck in der Sprache. [No kangaroos at the left—On the variability of spatial conceptions and their expression in language.] In H.-J. Kornadt, J. Grabowski, and R. Mangold-Allwinn (Eds.), Sprache und Kognition: Perspektiven modemer Sprachpsychologie (pp. 163-182). Heidelberg: Spektrum Akademischer Verlag. Korek, D. (1997). Raumliche Intelligenz als Determinante der Origo-Setzung. [Spatial intelligence as a determinant of Origo setting.] Unpublished diploma thesis, University of Mannheim. Landau, B., and Jackendoff, R. S. (1993). "What" and "where" in spatial language and spatial cognition. Behavioral and Brain Sciences, 16, 217-265. Levelt, W. J. M. (1982). Cognitive styles in the use of spatial direction terms. In R. J. Jarvella, and W. Klein (Eds.), Speech, place, and action. Studies in deixis and related topics (pp. 251-268). Chichester: Wiley. Levelt, W. J. M. (1984). Some perceptual limitations on talking about space. In A. J. van Doom, W. A. van de Grind, and J. J. Koenderink (Eds.), Limits in perception. Essays in honour of Maarten A. Bouman (pp. 323-358). Utrecht: VNU Science Press. Levelt, W. J. M. (1986). Zur sprachlichen Abbildung des Raumes: Deiktische und intrinsische Perspektive. [On the linguistic representation of space: Deictic and intrinsic perspective.] In H.-G. Bosshardt (Ed.), Perspektiven auf Sprache (pp. 187-211). Berlin: de Gruyter. Levelt, W. J. M., Richardson, G., and La Heij, W. (1985). Pointing and voicing in deictic expressions. Journal of Memory and Language, 24, 133-164. Levinson, S. C. (1991). Relativity in spatial conception and description (Working paper No. 1). Nijmegen (NL): Research Group on Cognitive Anthropology at the Max Planck Institute for Psycholinguistics. Levinson, S. C. (1992). Primer for the field investigation of spatial description and conception. Pragmatics, 2, 5-47. Linde, C., and Labov, W. (1975). Spatial networks as a site for the study of language and thought. Language, 51, 924-939. Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. San Francisco: Freeman. Miller, G. A., and Johnson-Laird, P. N. (1976). Language and perception. Cambridge, MA: Cambridge University Press. Moilanen, M. (1979). Statische lokative Prapositionen im heutigen Deutsch: Wahrheitsund Gebrauchsbedingungen. [Static locative prepositions in contemporary German: Truth conditions and usage.] Tubingen: Niemeyer. Retz-Schmidt, G. (1988). Various views on spatial prepositions. AIMagazine, 9, 95-105. Ruhrberg, P., and Rutz, H. (1990). Rdumliches Wissen und Semantik im Kontext der Generierung von Wegbeschreibungen. [Spatial knowledge and semantics in the context of the generation of route directions.] In C. Freksa, and C. Habel (Eds.), Reprdsentation und Verarbeitung raumlichen Wissens (pp. 235-249). Berlin: Springer. Schweizer, K. (1997). Raumliche oder zeitliche Wissensorganisation? Zur mentalen Reprdsentation der Blickpunktsequenz bei rdumlichen Anordnungen. [Spatial or tern-

The Use of Dimensional Prepositions

291

poral organization of knowledge? On the mental representation of point of view sequences in spatial arrangements.] Lengerich: Pabst Science Publisher. Schweizer, K., and Janzen, G. (1996). Zum EinfluB der Erwerbssituation aufdie Raumkognition: Mentale Reprasentation der Blickpunktsequenz bei rdumlichen Anordnungen. [How does encountering space influence spatial cognition? Mental representation of point of view sequences of spatial arrangements.] Sprache und Kognition, 15, 217-233. Shepard, R. N., and Cooper, L. A. (1982). Mental images and their transformations. Cambridge, MA: MIT Press/Bradford Books. Stern, W. (1936). Raum und Zeit als personale Dimensionen. [Space and time as personal dimensions.] Acta Psychologica, 1, 220-232. Svorou, S. (1994). The grammar of space. Amsterdam: Benjamins. Traugott, E. C. (1975). Spatial expressions of tense and temporal sequencing: A contribution to the study of semantic fields. Semiotica, 15, 207-230. Vater, H. (1996). Einfuhrung in die Raum-Linguistik [Introduction to the linguistics of space] (Kolner Linguistische Arbeiten—Germanistik; 24). Hurth-Efferen: Gabel. Wunderlich, D. (1981). Linguistic strategies. In F. Coulmas (Ed.), A Festschrift for native speaker (pp. 279-296). The Hague: Mouton. Wunderlich, D., and Herweg, M. (1991). Lokale und Direktionale. [Locatives and directional expressions.] In A. von Stechow, and D. Wunderlich (Eds.), Semantics. An international handbook of contemporary research (pp. 758-785). Berlin: de Gruyter.

This page intentionally left blank

SYNTACTIC AMBIGUITY RESOLUTION IN GERMAN BARBARA HEMFORTH* LARS KONIECZNY CHRISTOPH SCHEEPERS GERHARD STRUBE* * Center for Cognitive Science Institute of Computer Science and Social Research Freiburg University Freiburg, Germany Computational Linguistics University of the Saarland Saarbrucken, Germany Department of Psychology University of Glasgow Glasgow, Scotland

1. INTRODUCTION What makes a language an interesting subject for psycholinguistic research? Two answers can be given to this question: the first answer is that whatever makes a language different from the languages studied so far is interesting because new phenomena can then be seen from a new perspective. The second answer concentrates on the similarities more than the differences between languages: if parsing complexity or parsing preferences result from the general language independent architecture and mechanisms of the human language processor, comparable constructions in different languages should show the same pattern of preferences. In this paper, we will discuss sentence processing research from both perspectives. Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

293

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

294

B. Hemforth et al.

We will show in which ways the peculiarities of German give us the opportunity to provide a new test ground for theories of sentence processing in general. Furthermore, we will discuss evidence on constructions which are highly comparable, not only in German and English, but also many other languages (e.g., Spanish, French, Italian, Dutch, Japanese). We will discuss constraints on models of human sentence processing in general which can be derived from experiments on German. Our own account of attachment preferences, the principle of parameterized head attachment will be discussed based on experiments on PP- and NPattachment as well as extensions of the models which have proven to be necessary to explain relative clause attachment preferences.

2. THE DIFFERENTIAL PERSPECTIVE English is the language most thoroughly investigated in psycholinguistics. Many of the central garden-path phenomena as for example the garden-path sentences in (1), however, are mainly possible in English so that it is not clear whether the parsing principles derived from these phenomena really apply to all languages in the world. (1) a. The children visited by their grandparents had a lot of fun. b. While the psychologists discussed the evidence was disproved. English is known to have a rather fixed ordering of constituents. The functional role of a constituent can mostly be derived from its surface position whereas morphology (in particular case marking) is not very telling. German is certainly worth being looked at from the differential view since it differs from English in decisive ways: constituent ordering in German is relatively free (as e.g., in Dutch or Finnish, among others), and German has a rather rich morphology (as e.g., Finnish or Turkish, among others) that can be exploited for psycholinguistic experiments. In the remainder of this section, we will spell out some of the peculiarities of German that make it particularly interesting for psycholinguistics. These basics of German syntax will form the background for the experiments to be presented in later sections. In German, subjects may appear before objects or vice versa (2a,b). (2) a. Der Arzt besuchte den Patienten. 'The doctor[NOM, masc] visited the patient[ACC, masc].' b. Den Patienten besuchte der Arzt. 'The patient[ACC, masc] visited the doctor[NOM, masc]. This feature allows us to separate aspects of functional/thematic role assignment from aspects of surface ordering. In languages with fixed constituent ordering this is obviously not possible because surface position and functional role are unavoid-

Syntactic Ambiguity in German

295

ably confounded. For singular masculine noun phrases as in (2), the functional role of subject and direct object is unambiguously marked by grammatical case, nominative or accusative, respectively. Nicely enough, the German case marking system is far from perfect: singular feminine and neuter noun phrases as well as all plural noun phrases are ambiguous with respect to nominative or accusative case marking (3). (3) Die Arztin besuchte die Patientin. 'The doctor[NOM or ACC, fem] visited the patient[NOM or ACC, fem].' Example (3) is fully ambiguous with respect to the assignment of functional roles. Nonetheless, there is a strong preference to interpret the first NP as the subject of the sentence and the second NP as the object. The nominative/accusative syncretism gives us the opportunity to disambiguate structural ambiguities like constituent ordering ambiguities as in (3) or all sorts of attachment ambiguities at various positions in the sentence. Subject-object ordering may be disambiguated on the first NP as in (2), on the second NP as in (4a,b), by number agreement on the verb (4c,d), or it may not be disambiguated at all as in (3). (4) a. Die Arztin besuchte der Patient. 'The doctor[NOM or ACC, fem] visited the patient[NOM, masc].' 'The patient visited the doctor.' b. Die Arztin besuchte den Patienten. 'The doctor[NOM or ACC, fem] visited the patient[ACC, masc].' c. Die Arztin besuchten die Patientinnen. 'The doctor[NOM or ACC, sing] visited[plural] the patients[NOM or ACC, plural].' 'The patients visited the doctor.' d. Die Arztin besuchte die Patienten. 'The doctor[NOM or ACC, sing] visited [singular] the patients [NOM or ACC, plural].' Another grammatical feature that makes German a valuable language to work on is the variability in verb positioning. Finite verbs in German (as in Dutch) either precede their arguments (5a), follow them (5b), or have one argument preceding and the others following (5c). Nonfinite verbs always appear clause-final (5d,e) (5) a. Gestern schenkte der Mann dem Neffen ein Buch. 'Yesterday gave the man the nephew a book.' 'Yesterday, the man gave the nephew a book.' b. DaB der Mann dem Neffen ein Buch schenkte. "That the man the nephew a book gave.' 'That the man gave the nephew a book.' c. Der Mann gab dem Neffen ein Buch. "The man gave the nephew a book.'

296

B. Hemforth et al.

d. Der Mann versprach, dem Neffen das Buck zu schenken. 'The man promised the nephew the book to give.' 'The man promised to give the nephew the book.' e. Der Mann hat dem Neffen das Buck gegeben. 'The man has the nephew the book given.' 'The man has given the nephew the book.' We will not expand on the syntactic structures and derivations assumed for these variants. We only want to emphasize that this feature allows us to disentangle aspects of surface ordering, of functional roles like subject or object, and of thematic roles like agent, patient etc. Why is this feature of German syntax particularly interesting? A considerable amount of psycholinguistic research has been devoted to the role of verb frame information in ambiguity resolution (Abney, 1987; Britt, 1994; Clifton, Frazier, and Connine, 1984; Clifton, Speer, and Abney, 1991; Ferreira and Henderson, 1990; Ford, Bresnan, and Kaplan, 1982; Frazier, 1989; MacDonald, Pearlmutter, and Seidenberg, 1994; Mitchell, 1987,1989). The main question in this kind of research is whether or not verb frame information influences first analyses in ambiguity resolution. Varying the position of the verb with respect to its arguments allows us not only to vary the strength of a particular kind of information but the availability of this information. If the verb always appears in the same position, either preceding (as in English) or following (as in Japanese) its complements, we are not able to decide which role the availability of verb frame information plays within a single language. In sections 6 and 7, we will present results from a series of experiments, exploiting these features for the analysis of word order ambiguities as well as attachment ambiguities.

3. PARALLEL CONSTRUCTIONS As Mitchell (1996) pointed out, examining different kinds of ambiguities in different languages does not necessarily give us a convincing insight into the universal foundations of the human language processor. Since languages differ with respect to the possible syntactic structures and the grammatical cues available during sentence processing, parsing strategies may be fully dependent on the particular language, and nothing about the basic cognitive system may be derivable from this kind of experiment. Hence, it is certainly useful to try to find structures as closely matched as possible, to set up experiments with materials as closely matched as possible, and to use the same kind of experimental technology. Starting with Cuetos and Mitchell (1988), this enterprise is now being pursued by an increasingly large group of researchers (e.g. Brysbaert and Mitchell, 1996, Frazier and Clifton, 1996; Gilboy, Sopena, Frazier, and Clifton, 1995; Hemforth,

Syntactic Ambiguity in German

297

Konieczny, and Scheepers, 1994, in press; Kamide and Mitchell, 1997; Mitchell and Cuetos, 1991; de Vincenzi and Job, 1995). This research has been based on a particular kind of ambiguity that occurs in many languages in the world: the choice of an attachment site for a relative clause within a complex noun phrase (6a,b). (6) a. the daughter of the teacher who came from Germany b. the fence near the watering place at the door that had to be repaired Evidence on ambiguity resolution in this kind of construction will be presented in a later section. In order to acknowledge the full range of insight available from cross-linguistic studies, we will have to look at languages from two perspectives: the differential perspective where different questions can be tackled in different languages, and the universal perspective where we look for highly comparable phenomena in order to get closer to the underlying cognitive mechanisms. In the following sections, we will present evidence from experiments under both perspectives.

4. CONSTITUENT ORDERING The preferred ordering of subject and object NPs is currently one of the best studied topics in German sentence processing (e.g. Bader, 1994; Bader, Bayer, Hopf, and Meng, 1996; Gorrell, 1996; Hemforth, 1993; Hemforth, Konieczny, and Strube, 1993; Konieczny, 1996; Mecklinger, Schriefers, Steinhauer, andFriederici, 1995; Meng and Bader, 1996; Scheepers, 1996; Schlesewski, Fanselow, Kliegl, and Krems, 1997). The main finding is that of whatever construction you look at, be it simple main clauses (7a), which-questions (7b), embedded questions (7c), relative clauses (7d), or subordinate clauses (7e), among others, you always find a preference to interpret the first noun phrase as the nominative subject of the sentence. (7) a. Die Arztin besuchte der Patient. 'The doctor[NOM or ACC] visited the patient[NOM].' b. Welche Arztin besuchte der Patient? 'Which doctor[NOM or ACC] visited the patient[NOM]?' c. Ich weiB nicht, welche Arztin der Patient besuchte. 'I don't know which doctor[NOM or ACC] the patient[NOM] visited.' d. Die Arztin, die der Patient besuchte. 'The doctor who[NOM or ACC] the patient[NOM] visited.' e. DaB die Arztin der Patient besuchte. 'That the doctor[NOM or ACC] the patient[NOM] visited.'

298

B. Hemforth et al.

However, the strength of the garden-path elicited by structures like these differs considerably, being weakest for which-questions and relative clauses and strongest for subordinate clauses (main clauses like 7a lie somewhere in between, at least intuitively). One of the explanations given for this difference (e.g. Gorrell, 1995; Meng and Bader, 1996) is based on the assumption that different kinds of ambiguity are involved in these kinds of constructions. Within the framework of the Principles and Parameters approach (e.g. Chomsky, 1986), a filler gap ambiguity occurs in sentences with fronted which-phrases or relative pronouns. A filler is identified in the canonical topic position and it has to be bound to a gap in its base position (8ab, 9ab), i.e. the unmarked position it would appear in if it had not been fronted. In standard accounts this is assumed to be true for subject which-phrases as well as object which-phrases and all sorts of relative pronouns (but see Gorrell, 1996, for a different perspective). (8) a. Ich weiB nicht, [top welche Arztin] i der Patient — i besuchte. 'I don't know [top which doctor]i the patient[NOM]—i visited.' b. Ich weiB nicht, [top welche Arztin] i—i den Patienten besuchte. 'I don't know [top which doctor] i—i the patient[ACC] visited.' (9) a. Die Arztin, [top die]i der Patient—i besuchte. 'The doctor [top who]i the patient[NOM] — i visited.' b. Die Arztin, [top die]i — i den Patienten besuchte. 'The doctor [top who]i the patient[ACC] — i visited.' Main clauses (7a) and subordinate clause (7e) with fronted objects differ from these examples in several ways. Whereas object-first which questions and relative clauses are not restricted to specific pragmatic conditions, object first main clauses and subordinate clauses are pragmatically only licit in a constrained set of contexts, whereas subject first sentences are much less constrained (see Kaan, 1996, for a detailed discussion of pragmatic constraints based on similar constructions in Dutch). This pragmatic difference may account for the increased processing load found for main clauses and subordinate clauses. Additionally, object fronting in subordinate clauses and main clauses is asssumed to differ in structural complexity: object fronting in main clauses is usually assumed to result from moving the object NP to the pre-specified topic position similar to which questions and relative clauses. For subordinate clauses, object fronting cannot easily result from such a movement because the topic position is supposed to be to the left of the complementizer ("daB"). Additional structure building may have to take place to provide an attachment site preceding the subject (Gorrell, 1996). The increased processing load found for subordinate object-subject sentences compared to main clauses (Scheepers, 1996) may be due to the additional structural complexity of the unpreferred object-before-subject reading. These examples of subject-object ordering ambiguities which differ with respect to the strength of the subject-first

Syntactic Ambiguity in German

299

preference give us the opportunity to get a clearer picture of underlying representations and processes.

5. CONSTITUENT ORDERING AND THE TIME COURSE OF PHRASE STRUCTURE BUILDING A question central to psycholinguistic research concerns when items are attached to phrase markers during the processing of the input stream, i.e. when syntactic structure is built. It is widely acknowledged that parsing takes place as an interaction of bottom-up and top-down processes. Although several algorithms have been proposed on this subject in the literature, only three of them (or variants of these) are still central to psycholinguistic research1: (1.) Phrase structure is projected from the head of a phrase; no structure is built until the head of a phrase has been encountered. (2.) Items are attached to the phrase marker of the sentence as soon as they are processed without any delay (e.g. left-corner arc-eager parsing; see Abney and Johnson, 1991). (3.) A phrase can be projected as soon as its leftmost constituent has been processed. Following constituents can then be predicted and attached fully incrementally (left-corner arc-standard parsing, Aho, Hopcroft, and Ullman, 1974). German constituent ordering preferences can be exploited quite easily to get a clearer picture of the processes going on: functional roles can only be assigned to phrases in the context of the phrase marker. Therefore, a preference to interpret a phrase in terms of a particular functional role implies that this phrase has already been integrated into the phrase marker of the sentence. Since the finite verb of a sentence is assumed to be its syntactic head, no signs of functional role preferences should be found before the verb if approach (1.) is right. This is surely incompatible with the data. In experiments using various experimental techniques, a subject-before-object preference was consistently established at positions long before the finite verb was encountered. Increased processing load was found on the unambiguously nominative marked noun phrase which forces an object-first reading in sentences like (7c,d,e), both in self-paced reading and eyetrakking experiments (Scheepers, 1996; Scheepers, Hemforth, and Konieczny, 1997), as well as in event related potentials time locked to this phrase (Mecklinger et al., 1995; Friederici, in press). Even more telling are eye-tracking experiments on German verb second clauses like (10a,b). (10) a. Den bekannten Arzt besuchte der Patient. 'The well-known doctor[ACC] visited the patient[NOM].' "The patient visited the well-known doctor.'

300

B. Hemforth et al.

b. Der bekannte Arzt besuchte den Patienten. 'The well-known doctor[NOM] visited the patient[ACC].' In (l0a) the first NP is unambiguously marked as the direct object of the sentence, in (l0b), it is unambiguously marked as subject. Case-marking is evident on each constituent of the NP, i.e., on the determiner, the adjective, and the noun. Objectfirst sentences even with an unambiguously case-marked first NP are slightly more difficult to process than subject-first sentences (Hemforth, 1993). The question we have to ask is exactly when this increased processing load can be noticed. If no attachment took place before the head of the sentence, i.e. the finite verb "besuchte", no object-first penalty should be detectable before the verb. If the phrase marker of the sentence can be projected as soon as its left-most constituent has been processed, increases in processing load may be found by the end of the first NP. Effects found earlier in the first NP, i.e. on the determiner or the adjective, are only compatible with incremental attachment. In word-by-word self-paced reading experiments (Hemforth, 1993; Hemforth, Konieczny, and Strube, 1993), we established increased reading times on the noun of the first NP if it was unambiguously marked as the direct object of the sentence.2 This pattern of results is incompatible with head corner parsing, but surely compatible with left-corner parsing. More recent results form eye-tracking experiments, however, show that signs of an object-first penalty can be found even before the noun of the first NP, namely on the case-marked adjective (Konieczny, 1996).3 From these results, we conclude that the integration of items into the phrase marker of the sentence takes place fully incrementally starting with the first word encountered, a view highly compatible with serial immediate attachment theories; i.e. the garden-path family of sentence processing theories (Frazier, 1979, 1987; Mitchell, 1987, 1994; but see Schlesewski et al., 1996, for an approach based on specifier and head projection that is compatible with some of the data). Note, that the time course of attachment processes even within the first NP could only be investigated because of the rich morphological case-marking system where determiners, adjectives, and nouns are marked for case.

6. CONSTITUENT ORDER AND THEMATIC PROMINENCE The question still remains whether the subject-before-object preference is due to purely syntactic principles, or whether it is a result from "higher order" principles based, for example, on default knowledge about how thematic roles are functionally realized. Recent accounts in theoretical linguistics (e.g. Grimshaw, 1990; Haider, 1993) suggest that the unmarked linearization of subject and object at the sentence surface directly follows a prominence hierarchy of the thematic roles specified in the theta grid of the verb: the most prominent "agent"- or

Syntactic Ambiguity in German

301

"experiencer"-role, which is typically realized as the subject, will precede the less prominent "theme"-argument, which is typically marked as the direct object. In this view, the subject-before-object preference emerges from a tendency to interpret the first NP as "agent" or "experiencer" (cf. MacDonald et al. 1994), which is rendered unlikely (and hence is accompanied with processing difficulty) when the first NP is unambiguously marked as the direct object. Psychological verbs like "furchten" ("to fear") and "angstigen" ("to frighten") are particularly interesting cases for testing this hypothesis because they show a systematic alternation in the realization of their thematic roles (see Grimshaw, 1990, for a discussion). Verbs like "furchten" realize their " experiencer" argument as subject and their "theme" as direct object. The opposite rolefunction-assignments hold for "angstigen". To vary the availability of verb specific thematic hierarchy information, we investigated clauses with the verb preceding its arguments (11, 13) as well as verb-final clauses (12, 14). According to the thematic prominence hypothesis, different word-order preferences should show up in accordance with the type of verb involved: a subject-before-object preference is predicted for verbs like "furchten", whereas an object-beforesubject preference should be observed for verbs like "angstigen", at least when the verb is read before its arguments are encountered, as in (11) and (13).4 (11) Offenbar furchtete die fahige Polizistin {a. den beruchtigten Verbrecher / b. der beriichtigte Verbrecher} ein wenig. 'Obviously feared the capable police woman the notorious criminal {a.ACC/b. NOM} a bit.' (12) Es war offensichtlich, daB die fahige Polizistin {a. den beruchtigten Verbrecher / b. der beruchtigte Verbrecher} ein wenig furchtete. 'It was obvious that the capable policewoman the notorious criminal {a. ACC / b. NOM} a bit feared.' (13)

Offenbar angstigte die fahige Polizistin (a. den beruchtigten Verbrecher / b. der beruchtigte Verbrecher} ein wenig. 'Obviously frightened the capable policewoman the notorious criminal {a. ACC/b. NOM} a bit.'

(14)

Es war offensichtlich, daB die fahige Polizistin {a. den beruchtigten Verbrecher / b. der beruchtigte Verbrecher} ein wenig angstigte. 'It was obvious that the capable policewoman the notorious criminal {a. ACC / b. NOM} a bit frightened.'

A series of questionnaire-studies as well as an eye-tracking experiment were carried out to test this prediction (Scheepers, 1996; Scheepers, Hemforth, and Konieczny, 1997). It was found out that the type of verb influenced the acceptability as well as the processing difficulty of sentences like (11, 13). Irrespective of the verb position, however, this influence of verb type was much weaker

302

B. Hemforth et al.

than was predicted by the thematic prominence hypothesis: although the subjectbefore-object preference was significantly less pronounced in sentences with "angstigen" (13) than in sentences with "fiirchten" (11), it was not completely overwritten by verb type information, i.e. the subject-before-object preference still showed up as a reliable effect in (13). Most interestingly, the eye-tracking data indicated that different stages of processing are involved with respect to functional constraints on the one hand, and thematic constraints on the other hand. Even if the verb preceded its arguments (11, 13), a verb-independent subject-before-object preference was observed immediately on the disambiguating second NP (note that the first NP is always ambiguous with respect to case marking). Effects of verb type, however, showed up much later, namely during regressive saccades from the last region (i.e. the adverbial phrase "ein wenig" or the verb in verb-final clauses) to earlier passages of the sentence. Thus, even if the verb is read before its argument-NPs are encountered, thematic prominence is not considered during initial processing of the critical region, but during a re-checking stage of processing where the links between functional and thematic roles are evaluated. The observed word order preferences can therefore not be reduced to thematic prominence, indicating that verb-independent differences in syntactic complexity are responsible for the subject-before-object preference during first analysis.

7. ATTACHMENT AMBIGUITIES The role of verb frame information for attachment ambiguities can be investigated in a similar way to constituent ordering. Some years ago we started investigating attachment preferences in sentences like (15), with well-known PP-attachment ambiguities (Hemforth, Konieczny, Scheepers, and Strube, 1992; Konieczny, Hemforth, and Strube, 1991; Konieczny, Hemforth, Scheepers, and Strube, 1997). (15) Susan schlug das Madchen mit dem Buch. 'Susan hit the girl with the book.' Since we were mainly interested in the role of lexical information and world knowledge in parsing, we used two sets of verbs, which differed with respect to the verbal bias for an instrumental PP: one set of verbs, like "to watch" and "to hit" showed a strong preference for an instrumental PP, whereas the others, like "noticed" and "to catch sight of", did not, which had been checked in pre-studies. Secondly, the PP was varied so that its content forced either the attachment to the VP (16) or to the NP (17). (16) Marion (a. beobachtete, b. erblickte] das Pferd mit dem neuen Fernglas. 'Marion {a. watched, b. caught sight of) the horse with the new binoculars.'

Syntactic Ambiguity in German

303

(17) Marion (a. beobachtete, b. erblickte} das Pferd mil dem weiBen Fleck. 'Marion (a. watched, b. caught sight of} the horse with the white fleck.' In self-paced reading and eye-tracking experiments (Konieczny et al., 1995; Konieczny et al., 1997), we found a verb-frame dependent attachment preference of the PP: Reading times were increased when there was a conflict between verbframe preferences and plausibility (16b, 17a) Additionally, we took advantage of the variation of the verb-placement possibilities in German sentences, here verb-placement in perfect-tensed sentences, like (18) and (19), and in subclauses, like (20) and (21). (18) Marion hat das Pferd mit dem neuen Fernglas {a. beobachtet, b. erblickt}. 'Marion has the horse with the new binoculars {a. watched, b. caught sight of}.' (19) Marion hat das Pferd mit dem weifien Fleck {a. beobachtet, b. erblickt}. 'Marion has the horse with the white fleck {a. watched, b. caught sight of}.' (20) Ich habe gehort, daB Marion das Pferd mit dem neuen Fernglas {a. beobachtete, b. erblickte}. 'I have heard, that Marion the horse with the new binoculars (a. watched, b. caught sight of}.' (21) Ich habe gehort, daft Marion das Pferd mit dem weiBen Fleck {a. beobachtete, b. erblickte}. 'I have heard, that Marion the horse with the white fleck {a. watched, b. caught sight of}.' The important thing to notice here is that the main verb appears at the clause-end in both of these cases. Thus, when the ambiguous PP is read, the verb, of which it might be an argument, is still absent. In our self-paced reading and eye-tracking studies, processing times at the noun of the PP increased more when the PP was semantically biased towards VP-attachment than when it was biased to fit to the preceding direct-object-NP (Konieczny et al., 1997). These data suggest that there has been a preferred initial parsing decision before the first source of semantic information became available, namely the semantic content of the PP. In verbfinal clauses, the PPs were preferentially attached to the preceding NP. In a further series we investigated attachment preferences in sentences like (22-25) (22) DaB der Arzt der Sangerin ein Medikament gegeben hat, wuBite niemand. 'That the doctor [the singer[fem, {GEN/DAT}]] a medicine given has, knew nobody.' 'Nobody knew that the doctor has given a medicine to the singer.'

304

B. Hemforth et al.

(23) Daft der Arzt der Sangerin ein Medikament entdeckt hat, wuftte niemand. 'That the doctor [the singer[fem, {GEN/DAT}]] a medicine discovered has, knew nobody.' 'Nobody knew that the doctor of the singer has discovered a medicine.' In these constructions, which were first tested by Bader (1990), we took advantage of case-marking of NPs. In the case of (22) and (23), case marking on the critical region is ambiguous. Locally, the [NP der Sangerin] can either be interpreted as a dative-case complement of the verb, or as a genitive-case modifier of the subject [NP der Arzt], which would have been expressed in English by the PP [of the singer]. This ambiguity is only resolved when the verb-subcategorizer at the end of the subordinate clause is encountered: in the case of (22), geben (to give) requires a dative-marked object, and thus demands the critical NP to be integrated into the VP. Entdecken (to discover) in sentence (23), on the other hand, is a strictly transitive verb, ruling out a dative-marked object. Therefore, the critical NP has to be attached to the subject-NP as a genitive-marked modifier. Subject-paced experiments as well as an eye-tracking experiment indicated that the ambiguous NP was preferentially attached to the preceding subject, indicated by lower reading times at the verb in sentences like (23), although the difference was only marginally reliable here. In sentences (24) and (25), the NP was unambiguously morphologically case marked, either as the dative-cased "dem Sanger", which could only be a verbcomplement, or as genitive-cased "des Sangers", which translates to "of the singer" in English. (24) DaB der Arzt dem Sanger ein Medikament gegeben hat, wuBte niemand. 'That the doctor [the singer[masc, DAT]] a medicine given has, knew nobody.' 'Nobody knew that the doctor has given a medicine to the singer.' (25) Daft der Arzt des Sangers ein Medikament entdeckt hat, wuBte niemand. 'That the doctor [the singer[masc, GEN]] a medicine discovered has, knew nobody.' 'Nobody knew that the doctor of the singer has discovered a medicine.' Thus, the critical NP was disambiguated as early as possible, i.e. at the determiner of the NP itself. Nevertheless, the pattern of results was similar: with word-byword subject-paced presentation as well as eye-tracking, subjects spent reliably more time reading the dative-case determiner "dem" than reading the genitivecase determiner "des", which strongly forces the "non-minimal" attachment to the preceding subject-NP. In an attempt to account more thoroughly for these and many other data, we came up with an account of sentence processing, (which we called parametrized head attachment), consisting of three rather simple principles:

Syntactic Ambiguity in German

305

(26) head attachment Attach a constituent to a phrase whose lexical head has already been encountered. The head attachment principle applies to all those cases, where an attachment ambiguity can be resolved by either attaching to a preceding head, or to a lexical head yet to come. In many other cases, such as the verb-second sentences in (15), there are two or more heads that are potential attachment sites for an ambiguous item. Our data suggests that the decision now depends on the lexical preferences of either of the heads, as expressed in (27). (27) preferred-role attachment Attach constituent i to a phrase within the current clause whose head highlights a theta-role for i. This is to an extent the reincarnation of Ford et al.'s (1982) principle of lexical preference, which was built into their LFG-based parsing model. It is also highly compatible with lexicalized constraint-based approaches like Trueswell et al. (1994) or MacDonald et al. (1994). Of course, there are cases where two (or more) preceding heads do not differ in their preference to bind a constituent. In these cases, a decision is supposed to be based upon recency (28). (28) most recent head attachment Attach the constituent to the phrase whose head was encountered most recently.5 Parametrized head attachment integrates these three principles: (29) parametrized head attachment, PHA (Konieczny, Hemforth, Scheepers, and Strube, 1997) (Attempt to) apply head attachment (26) before preferred-role attachment (27) before most-recent head attachment (28).

8. AN INTERMEDIATE SUMMARY What can we learn about sentence processing from the consideration of the experiments experiments presented so far? The results show a rather clear picture of the human sentence processor. The fine grained analysis of attachment processes even within noun phrases shows that the human parsing mechanism •

works highly incrementally, i.e. it attaches each incoming item as it is encountered (Frazier, 1987a), see also "linear parsing" (Konieczny and Strube, 1995), (as opposed to head-driven or head licensing parsing),

306

B. Hemforth et a).

•

commits itself to a structural alternative very early (as opposed to minimal commitment or to a "wait and see" strategy).

By varying the position of the verb and hence the availability of verb-specific information we learned that the human sentence processor •

resolves structural ambiguities by identifying the presence or absence of the lexical head of the attachment sites, identifying the lexical properties of such heads, and the respective distance between the item to be attached and such lexical heads. Word order preferences, however, cannot be overridden by thematic hierarchy information.

But as the following section will show, recent cross-linguistic data have proven this picture to be somewhat simplistic.

9. LIMITS OF PARAMETRIZED HEAD ATTACHMENT The verb-second PP-attachment examples described earlier on are already an example for constructions which are at least superficially very similar to structures in various languages like English, Spanish, or French, among many others. The constructions most investigated in cross-linguistic studies, however, involve relative clause modifiers (30). (30) Someone shot the daughter of the colonel who was on the balcony. The relative clause in this kind of construction may be attached to either of the preceding two NPs, the head of the object noun phrase ("the daughter"; Nl or "high" attachment) or the modifying NP ("the colonel"; N2 or "low attachment"). The predictions given by principles like most recent head attachment or late closure are rather clear cut: there should be a preference to attach the relative clause to the most recent noun phrase, i.e. the modifying NP "the colonel". Unfortunately, this prediction only holds for English (e.g. Cuetos and Mitchell, 1988; Frazier and Clifton, 1996; Henstra, 1996), and even there not consistently so (Carreiras and Clifton, 1993). In most (if not all) other languages investigated so far, at least the final interpretation of pragmatically ambiguous relative clause constructions showed a clear preference for high attachment (Italian: de Vincenzi and Job, 1995; French, Pynte and Zagar, in press, Japanese: Kamide and Mitchell, 1997; Dutch: Brysbaert and Mitchell, 1996; Mitchell and Brysbeart, this volume; Spanish: Cuetos and Mitchell, 1988; Carreiras and Clifton, 1993; Gilboy, Sopena, Frazier, and Clifton, 1995). German matches these languages well (Hemforth, Konieczny, and Scheepers, 1994, in press). In written or auditory presentation, in self-paced reading or eye-tracking experiments, using syntactic as well as pragmatic disambiguation, relative clauses consistently show a preference for high

Syntactic Ambiguity in German

307

attachment. The pattern of results found in different languages makes relative clauses odd for two reasons: first, they do not behave properly with respect to any kind of universal recency-based parsing principle. This is bad, because recency works very nicely for a huge amount of structures (see Frazier and Clifton, 1996; Gibson, 1991; Gibson et al., 1996; Konieczny, 1996), and it is cognitively plausible. The second problem is that relative clauses do not behave alike in all languages, presenting problems for any kind of universal principle. But which mechanisms make relative clauses behave so oddly? The first thing to do is to see whether the problems found here are just problems of relative clauses or problems of modifiers per se. Consequently, we looked at structurally and pragmatically similar relative clauses and prepositional phrases (31a,b; Hemforth et al., in press). (31) a. die Tochter der Lehrerin, die aus Deutschland kam 'the daughter of the teacher who was from Germany' b. die Tochter der Lehrerin aus Deutschland 'the daughter of the teacher from Germany' In on-line and off-line studies, we found a strong interaction between modifier type and attachment preference. Whereas relative clauses were preferentially attached high, prepositional phrases showed a clear preference for low attachment. This modifier dependent preference shift was also established for English recently (Traxler, Pickering, and Clifton, 1996). Contrasting to relative clauses, prepositional phrases obviously obey recency-based parsing principles in English as well as in German (native informants indicate that this may also be true for Spanish, Italian, and French). The two major questions to be answered now are: (1.) What makes relative clauses different with respect to the general human sentence processor, and (2.), what makes English different with respect to the processing of relative clauses. Due to space limitations, we are not able to review the theoretical accounts presented in the past few years, accounts particularly concerned with the second question (see e.g. Cuetos and Mitchell, 1988; Frazier and Clifton, 1996; Gibson, Pearlmutter, Canseco-Gonzalez, and Hickock, 1996; de Vincenzi and Job, 1995). Many potential differences based on pragmatic aspects like restrictiveness, the prosodic differences involved, or syntactic aspects may contribute to the differences between relative clauses and PPs. We will concentrate here on one difference that we have investigated in several experiments: relative clauses in German and many other languages are headed by a relative pronoun. Binding this relative pronoun to its antecedent is an anaphoric process. Thus, finding the entity to be modified by a relative clause is not only a question of syntactic attachment but also one of anaphor resolution. This is obviously not true for PPs. What preferences do we expect and find for anaphoric pronouns in these constructions?

308

B. Hemforth et al.

Anaphors tend to go for salient, focused, main discourse entities (Sanford and Garrod; 1981; Garrod and Sanford, 1985) and not for modifiers. Accordingly, preferences found for relative clause attachment and anaphoric binding in constructions like (32a,b) go exactly along the same lines (Hemforth, Konieczny, and Scheepers, in press): a strong high attachment/binding preference can be established in both cases. (32) a. Jemand sah den Studenten des Professors, der im Labor war. 'Someone saw the student of the professor who was in the lab.' b. Jemand sah den Studenten des Professors, als er im Labor war. 'Someone saw the student of the professor when he was in the lab.' If ambiguity resolution involving relative clauses is determined by a combination of syntactic and anaphoric processes, this can explain why these constructions do not behave the way a purely syntax based mechanism would predict. But why do we find language specific differences? Again, there are various potential reasons on offer (see Cuetos and Mitchell, 1988; Mitchell and Brysbaert, this volume; Frazier and Clifton, 1996; Gibson et al., 1996). We can only speculate on one of them here: the use of an overt relative pronoun in a language may influence the degree to which anaphoric processes determine the attachment preference. In English, relative pronouns may be omitted, resulting in constructions like (33a,b). (33) a. the director of the movie (who was) accused of murder b. the wife of the major (who was) wearing pink lipstick According to native informants, the low attachment preference appears to be particulary strong for these constructions, which we are currently investigating in cooperation with Martin Corley and Simon Liversedge. The reduced reliability of anaphoric information in English may increase the role of syntactic attachment processes, thus leading to a low attachment preference even in cases where a pronoun is available. Other potential sources of differences between the languages will clearly have to be looked at in the future, including differences in prosodic structures and in attachment sites for relative clauses.

10. CONCLUSIONS The cross-linguistic approach has proven very helpful in research on human parsing, and natural language processing in general. German is a language with some peculiarities (as Mark Twain once noted, when he presented his fellow Americans at home with a caricature of German, of which he nevertheless had an excellent command6). As we have shown, these peculiarities can help us to dis-

Syntactic Ambiguity in German

309

entangle word order, syntactic roles, and thematic roles, and hence, to gain a clearer overall picture of human sentence processing.

ACKNOWLEDGMENTS This research was supported by the German National Research Foundation (Deutsche Forschungsgemeinschaft, DFG, He 2310/2-1). We would like to thank Don Mitchell for many helpful discussions and critical comments on an earlier version of this manuscript, and Nick Ketley for his very helpful comments.

REFERENCES Abney, S. (1987). Licensing and parsing. Proceedings of NELS 17. Amherst: University of Massachusetts. Abney, S., and Johnson, M. (1991). Memory requirements and local ambiguities of parsing strategies. Journal of Psycholinguistic Research, 20(3), 233-250. Aho, A. V, Hopcroft, J. E., and Ullman, J. D. (1974). The design and analysis of computer algorithms. Reading, MA: Addison-Wesley Publishing Company. Bader, M. (1990). Syntaktische Prozesse beim Sprachverstehen: Theoretische Uberlegungen und experimentelle Untersuchungen. Magisterarbeit an der Albert-LudwigsUniversitat Freiburg. Bader, M. (1994). Syntaktische Prozesse beim Sprachverstehen. Stuttgart: Unveroffentlichte Dissertation. Bader, M., Bayer, J., Hopf, J. M., and Meng, M. (1996). Case-assignment in processing German verb-final clauses. Proceedings of the NELS Workshop on Sentence Processing. Cambridge: MIT, Occasional papers in linguistics, 9, 1-25. Britt, M. A. (1994). The interaction of referential ambiguity and argument structure in the parsing of prepositional phrases. Journal of Memory and Language, 33, 251-283. Brysbaert, M., and Mitchell, D. (1996). Modifier attachment in sentence parsing: evidence from Dutch. Quarterly Journal of Experimental Psychology, 49A, 664-695. Carreiras, M., and Clifton, Jr. C. (1993). Relative clause interpretation preferences in Spanish and English. Language and Speech. 36(4), 353-372. Chomsky, N. (1986). Knowledge of language: Its nature, origin, and use. New York: Praeger. Clifton, C., Frazier, L., and Connine, C. (1984). Lexical expectations in sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 23, 696-708. Clifton, C., Jr., Speer, S., and Abney, S. P. (1991). Parsing arguments: phrase structure and argument structure as determinants of initial parsing decisions. Journal of Memory and Language, 30, 251-271. Cuetos, F,, and Mitchell, D. (1988). Cross linguistic differences in parsing: Restrictions on the issue of the Late Closure strategy in Spanish. Cognition, 30, 73-105. De Vincenzi, M., and Job, R. (1995). An investigation of late closure: the role of syntax,

310

B. Hemforth et al.

thematic structure and pragmatics in initial and final interpretation. Journal of Experimental Psychology: Learning, Memory, & Cognition, 21(5), 1303-1321. Ferreira, F., and Henderson, J. M. (1990). Use of Verb Information in Syntactic Parsing. Journal of Experimental Psychology: Learning Memory and Cognition, 16(4), 555-568. Ford, M., Bresnan, J., and Kaplan, R. M. (1982). A comptence-based theory of syntactic closure. In J. Bresnan (Ed.), The mental representation of grammatical relations (pp. 727-796). Cambridge, MA: MIT Press. Frazier, L. (1979). On comprehending sentences: Syntactic parsing strategies. Bloomington, IN: IULC. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), The psychology of reading (pp. 559-586). Hove/London/Hillsdale, NJ: Lawrence Erlbaum. Frazier, L. (1989). Against lexical generation of syntax. In W. Marslen-Wilson (Ed.), Lexical representation and process (pp. 506-528). Cambridge, MA: MIT Press (Bradford). Frazier, L., and Clifton, C. (1996). Construal. Cambridge, MA: MIT Press. Friederici, A. (in press). Diagnosis and Reanalysis: two processing steps in the brain may differentiate. In: J. D. Fodor and F. Ferreira (Eds.), Reanalysis in sentence processing. Dordrecht: Kluwer Academic Press. Garrod, S., and Sanford, A. J. (1985). On the real-time character of interpretation during reading. Language and Cognitive Processes, 1,43-59. Gibson, E. (1991). A computational theory of human linguistic processing: memory limitations and processing breakdown. Unpublished Ph.D. dissertation. Carnegie Mellon University. Gibson, E., Pearlmutter, N., Canseco-Gonzalez, E., and Hickock, G. (1996). Recency preference in the human sentence processing mechanism. Cognition, 59, 23-59. Gilboy, E., Sopena, J., Frazier, L., and Clifton, C. (1995). Argument structure and association preferences in Spanish and English complex NPs. Cognition, 54, 131-167. Gorrell, P. (1995). Syntax and parsing. Cambridge: Cambridge University Press. Gorrell, P. (1996). Diagnosis and determinism. New York: Paper presented at the 9th annual CUNY conference on human sentence processing. Grimshaw, J. (1990). Argument structure. Cambridge, MA: MIT Press. Haider, H. (1993). Deutsche Syntax—generativ: Vorstudien zur Theorie einer projektiven Grammatik. Tubingen: Narr. Hemforth, B. (1993). Kognitives Parsing: Reprasentation und Verarbeitung sprachlichen Wissens. Sankt Augustin: Infix. Hemforth, B., Konieczny, C., and Konieczny, L. (in press). Syntactic attachment and anaphor resolution: two sides of relative clause attachment. In M. Crocker, M. Pickering, and C. Clifton (Eds.), Architectures and mechanisms for language processing. Cambridge: Cambridge University Press. Hemforth, B., Konieczny, L., Scheepers, C., and Strube, G. (1992). SOUL-Processing: Semantik-orientierte Prinzipien menschlicher Sprachverarbeitung. In G. Gorz (Ed.), KONVENS-92 (pp. 198-208). Berlin: Springer. Hemforth, B., Konieczny, L., and Scheepers, L. (1994). Probabilistic or universal approaches to sentence processing: how universal is the human language processor? In H. Trost (Ed.), KONVENS94 (pp. 161-170). Berlin: Springer. Hemforth, B., Konieczny, L., and Strube, G. (1993). Incremental syntax processing and

Syntactic Ambiguity in German

311

parsing strategies. Proceedings of the 15th annual conference of the Cognitive Science Society (pp. 539-545). Hillsdale, NJ: Erlbaum. Henstra, J. (1996). Relative clause attachment in English: Eye-tracking versus self-paced reading. Poster presented at the 2nd AMLaP, Turino, Italy, September, 1996. Kaan, E. (1996). Processing subject-object ambiguities in Dutch. Rijksuniversiteit Groningen: Unpublished doctoral dissertation. Kamide, Y., and Mitchell, D. (1997). Relative clause attachment: non-determinism in Japanese parsing. Journal of Psycholinguistic Research, 26, 247-254. Konieczny, L. (1996). Human sentence processing: A semantics-oriented parsing approach. IIG-Berichte 3/96. Konieczny, L., Hemforth, B., and Strube, G. (1991). Psychologisch fundierte Prinzipien der Satzverarbeitung jenseits von Minimal Attachment. Kognitionswissenschaft, 2, 58-70. Konieczny, L., Hemforth, B., Scheepers, C., and Strube, G. (1997). The role of lexical heads in parsing: evidence from German. Language and Cognitive Processes, 12, 307-348. Konieczny, L., and Strube, G. (1995). SOUL: A cognitive parser. In J. D. Moore and J. F. Lehman (Eds.), Proceedings of the Seventeenth Annual Conference of the Cognitive Science Society (pp. 631-636). Mahwah, NJ: Lawrence Erlbaum Associates. MacDonald, M., Pearlmutter, N., and Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101,4, 676-703. Mecklinger, A., Schriefers, H., Steinhauer, K., and Friederici, A. (1995). Processing relative clauses varying on syntactic and semantic dimensions: An analysis with eventrelated potentials. Memory and Cognition, 23, 477-494. Meng, M., and Bader, M. (1996, September). Syntax and morphology in parsing. Paper presented at the 2nd AMLaP at Turino, Italy. Mitchel, D. (1996). Empirical facts on human parsing: findings to be explained in viable models of the process. (Ed.), Talk presented at the Workshop on Computational Psycholinguistics, MAS, Wassenaar, NL. Mitchell, D. (1994). Sentence Parsing. In M. A. Gernsbacher (Ed.), Handbook of Psycholinguistics (pp. 375-410). San Diego: Academic Press. Mitchell, D. C. (1987). Lexical guidance in human parsing: locus and processing characteristics. In M. Coltheart (Ed.), Attention and Performance XII: The psychology of reading (pp. 601-618). London: Lawrence Erlbaum Associates Ltd. Mitchell, D. C. (1989). Verb-guidance and other lexical effects in parsing. Language and Cognitive Processes, 4, 123-155. Mitchell, D. C., and Cuetos, F. (1991). The origins of parsing strategies. In C. Smith (Ed.), Current Issues in Natural Language Processing. Austin: Center for Cognitive Science, University of Texas. Sanford, T, and Garrod, S. (1981). Understanding written language. Chicester: Wiley. Scheepers, C. (1996). Menschliche Satzverarbeitung: Syntaktische und thematische Aspekte der Wortstellung im Deutschen. Universitat Freiburg: Unveroffentlichte Dissertation. Scheepers, C., Hemforth, B., and Konieczny, L. (1997). Linking syntactic functions with thematic roles: psych-verbs and the resolution of subject-object ambiguity. Manuscript to appear in B. Hemforth and L. Konieczny (Eds.), Cognitive parsing in German.

312

B. Hemforth et al.

Schlesewski, M, Fanselow, G., Kliegl, R., and Krems, J. (1996). Locally ambiguous wh-questions in German. Manuscript to appear in B. Hemforth and L. Konieczny (Eds.), Cognitive parsing in German. Traxler, M. J., Pickering, M. J., and Clifton, C. (1996). Architectures and mechanisms that process prepositional phrases and relative clauses. Paper presented at the 2nd AMLaP, Turino, Italy, September, 1996. Trueswell, J. C., Tanenhaus, M. K., & Garusey, S. M. (1994). Semantic influences on pausing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33, 285-318.

NOTES J

We will not discuss the further possibility that no syntactic structure building takes place at all. 2 The preference for a nominative marked NP is position dependent. In positions following the main verb, accusative marked NPs are easier to process than nominative NPs if a case-ambiguous NP preceded the verb. 3 Since the determiner was the first word in the sentence, potential effects at the very beginning of this NP may have been obscured by noise. 4 Here a special kind of verb-second sentences in German was examined where the "Vorfeld"-position (roughly the so called topic position before the verb in verb-second sentences) is occupied by an adverbial and both arguments remain in the "Mittelfeld" (roughly the area following the verb in these cases). 5 The predictions of most-recent head attachment compare to those of late closure in the garden-path model. However, it is only applied if the other principles fail to provide a decision. This has been expressed in the unified parametrized-head attachment principle, PHA (29), which furthermore serves the purpose of emphasizing the fact that attachment ambiguities are resolved on the basis of certain parameters of lexical heads, such as relative position and lexical preferences. 6 From the essay: "The awful German language". In Mark Twain: a tramp abroad. London, Chatto and Windus, ca. 1910.

CHALLENGES TO RECENT THEORIES OF CROSSLINGUISTIC VARIATION IN PARSING: EVIDENCE FROM DUTCH DON C. MITCHELL* MARC BRYSBAERt * Department of Psychology University of Exeter Exeter, England t

Department of Psychology University of Ghent Ghent, Belgium

1. INTRODUCTION The last quarter century has seen a burgeoning of systematic and controlled experimental studies of language processing. In this program of studies, 90% of the work has been based on a single language (English) with just a tiny portion of the research effort being devoted to the other 6000 or so languages of the world. Given these circumstances, it is sensible to ask whether the outcomes of this enterprise offer insights into language processing in general or whether researchers have just been formulating dubious generalizations based on idiosyncratic features of individual languages. The present chapter provides a case study of an area of psycholinguistic research in which investigators may have been in danger of overgeneralizing from a narrow evidence base. We shall examine an aspect of parsing in which widely accepted generalizations have turned out to be inaccurate or incomplete and in which languages other than English appear to operate according to principles previously unexplored in mainstream research. The work raises questions about the possibility that different languages are analyzed by Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

313

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

314

Don C. Mitchell and Marc Brysbaert

means of qualitatively different procedures. If this turns out to be the case, then a full account of language processing will have to cover more than a description of the general procedures that might be shared by all language-processing systems. It will also have to say something about the special-purpose operations that occur only in individual languages or in subclasses of languages.

2. BACKGROUND Our discussion will focus on the aspects of syntactic analysis that are responsible for deciding whether a word, phrase, or clause should be associated with one of two or more competing "attachment sites" within a sentence. Classic examples include (1) and (2). (1) The man saw the spy with binoculars. (2) Barbara said the politician died yesterday. In both cases an ambiguous constituent (italicized) can be linked to either of two attachment sites in the sentence (marked in bold). A central challenge for parsing theories is to set out the principles which form the basis for resolving ambiguities of this kind (together, of course, with numerous other examples). Our treatment of the issue will concentrate exclusively on examples like (2), in which the two potential hosts are both from the same word-class. Although both are verbs in this particular illustration, the two attachment sites will both be nouns in all of the examples considered after this. A decade ago there was, in essence, a complete consensus about the broad principles underlying attachment decisions of this kind. The operating principles were based uniformly on the relative positions of the as-yet-unattached constituent and the competing head sites, and all accounts posited that the new constituent would be attached to the nearest or most recent site. The principle underlying this choice was given a variety of sobriquets, but the claim was essentially the same. In (2) the proposal was that 'yesterday' is attached to the nearby 'died' rather than the more distant 'told' following principles dubbed Late Closure, Right Association, Local Association, and numerous other terms. Thus, in the terminology of the most influential parsing model of the time, Frazier's (1979, 1987) garden-path model, the relevant operating rule (Late Closure) was "attach new items . . . into the phrase or clause postulated most recently." Attach 'yesterday' to 'died' in preference to 'told.' Given the theoretical statements current in the mid-1980s it seemed perfectly clear that the same operating principles should apply to sentences like (3). (3) Someone shot the servant of the actress who was on the balcony, As before, the ambiguous constituent can be attached to one of (at least) two different sites—again represented by words of the same class (both nouns). Here, as before, the prevailing "recency" or "locality" principles stipulated that the new

Parsing: Recent Challenges in Dutch

315

constituent should be preferentially attached to the closer site (i.e., to 'actress'). Indeed, at the time one of us (DCM) conducted a questionnaire study at the University of Exeter confirming that this was true (in the study reported as Experiment IB in Cuetos and Mitchell, 1988). Thus, the evidence seemed to be entirely compatible with the prevailing theoretical frameworks. There the story might have ended had Fernando Cuetos not been on an extended research visit to Exeter at the time (Fall, 1986). It was his clear intuition that in a Spanish equivalent of Sentence (3) the relative clause is preferentially attached to 'el criado' ('the servant') rather than actrice. As we elaborate below, this observation has subsequently been confirmed not only in numerous studies in Spanish, but also in several other languages as well. Much of the subsequent work has tried to explain why it is that the attachment biases of language processors seem to vary from language to language. In the course of this work, researchers have proposed a wide variety of new parsing principles in their attempts to capture the data. Arguably, none of these developments would have come about if the focus of research had not shifted away from the one language that (it turns out) is highly idiosyncratic when it comes to relative clause attachment. Scientific developments of this kind make a strong case for studying processes in different languages. However, the evidence on crosslinguistic differences raises numerous new problems at the same time. It seems that the machinery for sentence processing and parsing must vary from language to language. This raises several important questions. Does the machinery alter in systematic ways depending on the specific "niche" within which attachment processes have to operate in different languages? Are the fine details of site selection influenced by subtle variations in grammar from language to language? Are they influenced by statistical variations in the frequencies with which different "solutions" turn up in different languages? If we had full answers to all of these questions we could be confident that we have a good understanding of the way in which these parsing choices are determined in different languages. In this chapter we outline evidence showing how the characteristics of site selection vary from language to language, and we review a variety of theoretical proposals that have been put forward in attempts to explain these differences. We then proceed to examine these proposals and discuss their status in the light of recent psycholinguistic work. In particular, we outline a number of new studies in Dutch and spell out the theoretical implications of these data.

3. A CLOSER LOOK AT THE CROSSLINGUISTIC DIFFERENCES As we have already indicated, the core questions for present purposes are: (a) What are the processes by which a modifier like a relative clause is attached to an appropriate noun-head? and (b) Do these processes vary from language to language and if so why?

316

Don C. Mitchell and Marc Brysbaert

As we have already noted, intuitive judgments by native speakers suggest that the relative clause is linked to the earlier of the two competing noun phrases in Spanish and to the later one in English. This was confirmed in questionnaire studies carried out by Cuetos and Mitchell (1988). Informants in Spain and England were asked to read sentences like (3) in their own language. Each sentence was followed by a question designed to probe whether the Relative Clause (RC) had been understood as being attached to the first Noun Phrase (NP1) (i.e., 'the servant' or el criado) or the second (NP2: 'the actress' or la actrice). In Example (3) the question was simply "Who was on the balcony?" or its Spanish equivalent. The results confirmed that NP2 attachment was preferred in English while NP1 attachment predominated in Spanish. The attachment preference in Spanish was examined more closely by conducting a series of self-paced reading studies in which an extra phrase or clause was added to the RC to force NP1 or NP2 attachment. Confirming an existing bias in favor of NP1 attachment, the results showed that people took longer to read phrases that forced the alternative NP2 attachment than those resolved in favor of NP1. There have been at least five different kinds of theoretical explanations of this crosslinguistic difference, and these will be set out at length in the next section. However, before embarking on this, it is important to establish that the crosslinguistic differences are genuine. Since the Cuetos and Mitchell study, NP1 preference has been reported in numerous other Spanish studies (e.g., Carreiras, 1992; Carreiras and Clifton, 1993; Gilboy, Sopena, Clifton and Frazier, 1995; Gibson, Pearlmutter, and Torrens, 1997; see Cuetos, Mitchell, and Corley, 1996, for a review). Comparable findings have also been obtained in French (Mitchell, Cuetos, and Zagar, 1990; Pynte and Frenck-Mestre, 1996; Pynte, this volume; Zagar, Pynte and Rativeau, 1997); German (Hemforth, Konieczny, and Scheepers, 1994, in press; Hemforth, Konieczny, Scheepers, and Strube, this volume; Konieczny and Hemforth, 1996); Dutch (Brysbaert and Mitchell, 1996a); Russian (V. Kempe and R. Radach, personal communication, 1993); Afrikaans (P. Swanepoel, personal communication, 1995); and Thai (V. Robertson, personal communication, 1996). In contrast with these findings (and in line with the original Cuetos and Mitchell results), English appears to show a bias against attaching the RC to NP1. Apart from an early study reported informally by Clifton (1988), all questionnaire studies with monolingual English speakers have shown a bias in favor of NP2 attachment (e.g., Corley, 1996; Cuetos and Mitchell, 1988; Mitchell and Cuetos, 1991b; Gilboy et al., 1995, cf. Type B Kinship Relations; J. Ganger, personal communication, July 1996). Clifton's (1988) questionnaire showed a small NP1 bias—an outlying and arguably aberrant result that has never been fully explained. One possible explanation for this finding could be that bilingual subjects were not systematically excluded from the study. Fernandez (1995) has recently reported evidence that, even when tested only in English, the attachment preferences shown by bilingual Spanish-English readers is different from that shown by monolingual

Parsing: Recent Challenges in Dutch

317

English readers. The normal English NP2 preference seems to be swamped by NP1 biases that presumably come from the competing language. These findings stress the importance of either screening bilinguals from studies or using populations in which bilinguals are rare. Where these conditions are met, questionnaire results show that English readers display attachment biases that are different from those that show up in all other languages tested to date. The evidence considered so far has concerned attachments that end up prevailing once the sentence has been read. It is less certain that these effects exert their influence while the person is still reading the RC under examination. Support for an early NP2 bias of this kind seems to be confirmed in some on-line studies (e.g., Clifton, 1988; Corley, 1996, Section 5.3; Frazier and Clifton, 1996, pp. 79-80). However, the NP2 advantage sometimes fails to reach significance (e.g., Carreiras and Clifton, 1993; Mitchell and Cuetos, 1991b; see Cuetos et al., 1996, for a review). In another recent study Henstra (1996) has suggested that these differences may be due to variations in the sensitivity of different experimental tasks. She reports two eyetracking studies showing clear NP2 attachment effects, whereas the third experiment, using self-paced reading, showed no reliable bias. Perhaps the best interpretation of these results is that for English the initial NP2 bias is real, but relatively weak and therefore difficult to pick up reliably. Overall, these results seem to confirm that there is a genuine crosslinguistic difference between the attachment preferences in English and (it seems) almost every other language that has been investigated to date. Before moving on to the theoretical accounts of these findings it may be worth commenting on one or two possible exceptions to the apparent generalization that all languages other than English favor NP1 attachment. The first concerns the pattern of findings in Japanese. Kamide and Mitchell (1997) provided evidence that in this language the relative clause is briefly attached to NP1 as the sentence is being processed but that an NP2 preference eventually prevails by the end of the sentence. However, there are problems in interpreting this finding. Japanese is not directly comparable with any of the other languages studied to date. The RC appears before either of its potential host sites. It is followed by NP1, which for a brief period remains the only legitimate attachment site. The subsequent arrival of NP2 then provides an alternative (and, eventually, preferred) host site. Thus, Japanese contrasts markedly with languages in which two or more attachment sites are available as soon as the relative clause arrives. The analysis is also complicated by the fact that syntactic (phrase-marker) analysis for Japanese places NP2 higher than NP1 (rather than the reverse, as for the other languages discussed so far). Given the differential status of the two potential hosts in Japanese, it is not at all surprising that attachment preferences differ. There has also been a good deal of debate about the possibility that NP2 attachment might prevail for a brief period during sentence processing in other languages—particularly Italian. Vincenzi and Job (1993, 1995) have carried out

318

Don C. Mitchell and Marc Brysbaert

several self-paced reading experiments that appear to show that early reanalyses times were longer in sentences forcing attachment to NP1 than to NP2 (suggesting early NP2 linkage). Questions at the ends of sentences, however, indicated that there must subsequently have been a shift to NP1 preference. At present it is not entirely clear that these results provide solid support for an initial NP2 attachment preference. Pynte and Frenck-Mestre (1996) have recently argued that this nonstandard pattern of results may be due to methodological idiosyncrasies of the de Vincenzi and Job studies (see also Pynte, this volume). In particular, the Italian experiments differ from most others in the field in that complex NPs other than genitives (e.g., NP-con-NP; i.e., NP-with-NP) were included in the same on-line sessions as the genitive forms that are the subject of the present crosslinguistic comparisons. Pynte and Frenck-Mestre (1996) present evidence that the presence of alternative complex forms in the experimental materials can shift the initial bias from its normal NP1 preference. The role of such artifacts will have to be examined before it is safe to argue that attachment biases alter over time in languages like Italian. Before moving on to consider the theoretical interpretations of the crosslinguistic findings, it is perhaps worth noting that studies of modifier attachment have not been restricted to genitive NPs. Nor is it true that RCs are the only modifiers that have been examined. However, the present discussion focuses largely on the attachment of RCs to complex NPs consisting of a NP followed by a genitive prepositional phrase. Readers who are interested in finding about the findings with other types of NP and modifier are referred to other studies and surveys (e.g., Cuetos et al., 1996; de Vincenzi and Job, 1995, 1993; Frazier and Clifton, 1996; Gibson, Pearlmutter, Canseco-Gonzalez, and Hickok, 1996; Gilboy et al., 1995; Hemforth, et al.,1994; in press; Konieczny and Hemforth, 1996). What lies behind the differences in RC attachment? If they result from different parsing procedures, what does this say about the constraints and pressures that shape parsing mechanisms in different languages?

4. THEORETICAL INTERPRETATIONS Before moving on to consider possible explanations for these crosslinguistic differences, it should be pointed out that the empirical findings on RC attachment are not handled well by all mainstream accounts of parsing. Several proposals billed as relatively comprehensive models of human parsing simply offer no explicit account of RC ambiguity resolution (e.g., Crocker, 1992; Lewis, 1993; McClelland, St. John and Taraban, 1989; see also St. John and McClelland, 1990; Pritchett, 1992; Tanenhaus, Carlson, and Trueswell, 1989). Such accounts are perhaps best viewed as theories of verb-centered aspects of sentence processing

Parsing: Recent Challenges in Dutch

319

rather than full accounts of parsing processes more generally. Other models make incorrect predictions (i.e., NP2 preference) for the majority of the languages investigated so far (e.g., Frazier, 1987; Kempen and Vosse, 1989; see also Kempen, 1996; Konieczny, Hemforth, and Scheepers, 1994). The following paragraphs will be restricted to theories that offer explicit accounts of RC attachment and do a reasonable job of accounting for the basic findings. Five main theoretical accounts will be considered in the following paragraphs. These are (a) an account based on the Tuning Hypothesis (Mitchell and Cuetos, 199la; Cuetos et al., 1996); (b) an account based on parameter setting (Gibson et al., 1996); (c) a constraint-satisfaction lexicalist account (MacDonald, Pearlmutter, and Seidenberg, 1994; Thornton, Gil, and MacDonald, 1995); (d) an account based on Construal Theory (Frazier and Clifton, 1996); and (e) a dualmechanism explanation proposed by Hemforth and Konieczny (1996). 4.1. The Tuning Hypothesis Arguably the simplest account of crosslinguistic differences in attachment is that the preference in any language is determined by the frequency with which the alternative attachments are used in comparable structures in that language. For example, the Linguistic Tuning hypothesis maintains that initial parsing choices are made exclusively on the basis of the relative frequencies with which the reader or listener has resolved an ambiguity each way in the past (Brysbaert and Mitchell, 1996a; Cuetos et al, 1996; Mitchell and Cuetos, 1991a; Mitchell et al., 1995). According to the Tuning account, there is some kind of mechanism for keeping track of statistical frequencies (see Mitchell et al., 1995, for further discussions of this procedure). These statistical records are then accessed and used whenever the parser encounters a related structural ambiguity. The crosslinguistic differences arise because the statistical records have different entries for different languages (as determined by the exposure to differing language samples). In other words, NP2 attachment dominates in English because the relative clause is most often attached to this site when NPl-of-NP2-RC structures are used in the language. NP1-attachment prevails in other languages for precisely the same reason— because RCs are predominantly linked to the corresponding site in the languages in question. 4.2. Predicate Proximity/ Recency Theory A second theory that offers an explanation of the crosslinguistic differences is the Predicate Proximity/Recency Theory put forward by Gibson et al. (1996). The basic proposal here is that two competing factors play a role in selecting an attachment site for a modifier. These are (a) the structural proximity of each potential host site to the head of the entire predicate phrase (Predicate Proximity), and

320

Don C. Mitchell and Marc Brysbaert

(b) the relative distances between the modifier and each of the potential heads. The suggestion is that there are processing tendencies favoring attachments that are close to the head of the predicate phrase, together with biases pushing for attachment to closer or more recent sites (Recency Preference). The model generates its predictions by combining these tendencies in a series of simple formulae to give a numerical measure of "Processing Load" for each of the potential attachment sites. The site associated with the lowest processing load is the one selected as the preferred host. In a sentence like (3), the Recency mechanism would favor attaching the RC to the nearest potential head ('actress,' in this case), whereas PredProx would push to attach the clause to the competing site ('servant'). The preference for attachment to 'actress' is explained by Recency being stronger than PredProx (in English). To account for crosslinguistic differences, the model assumes that the PredProx (but not Recency) weights vary from language to language. In particular, it is assumed that the PredProx values are high enough to outweigh Recency in most languages, but that in English the weights are low enough to be swamped by Recency. The model proposes that the value of the PredProx weights is set as a result of exposure to the language in question (in a manner similar to that proposed in tuning accounts). Gibson et al. (1996) speculate that a possible determinant of this measure is the "average distance from the head of a predicate (verb) to its arguments" (p. 49). A useful feature of the model is that it can easily be adapted to provide explanations for the pattern of attachment preferences in three-site ambiguities like (4). (4) The lamp near the painting of the house that was damaged in the flood. Recency favors attaching the RC to 'house' rather than 'painting' and means that both are preferred to interpretations in which the word 'lamp' is the one modified by the clause. PredProx produces exactly the opposite pattern of preferences, and Gibson et al. (1996) have used the theory to account for findings with sentences of this kind. However to return to the crosslinguistic findings, the crucial point is that the theory accounts for the findings by parametrically adjusting the PredProx weights in the model's calculations. Beyond the precise details of the model, it seems that the parameter is assumed to be set by some kind of averaging mechanism that integrates head-argument distances on the basis of exposure to sentences of the language.

4.3. Constraint-Satisfaction Lexicalist Accounts In the PredProx/Recency model the degree of competition between different tendencies is rather circumscribed. However, other models propose much more open-ended competition. For example, the crucial feature of the Constraint Satis-

Parsing: Recent Challenges in Dutch

321

faction model (e.g., Boland, Tanenhaus, and Garnsey, 1990; MacDonald et al., 1994; Tanenhaus et al., 1989) is that information of any type can be used to influence the course of decision making. Models of this kind have only given very limited attention to the problems of modifier attachment and (as far as we are aware) no attention at all to crosslinguistic variations in attachment preferences. However, there is one version of the model that has specifically addressed the problem of resolving modifier attachment ambiguities. According to this account (MacDonald et al., 1994, pp. 697-698), the attachment of a modifier to NP1 or NP2 within a complex head is determined predominantly by the lexical properties of the nouns occupying to two competing slots. Thus if the noun 'actress' attracts a modifier more than the noun 'servant' it will "win" the competition for the attachment and the ambiguity will be resolved in favor of the link with 'actress.' Presumably, on this account an overall preference for NP1 (or NP2) attachment is that the nouns occupying the positions in question have a stronger lexical bias in favor of appearing alongside a modifier. Any variation from one language to another could potentially be explained in terms of the lexical properties of the individual nouns used in the experimental studies. Even where there was close translation (e.g., the Spanish actrice for the English 'actress'), there could never be any guarantee that the modifier-attracting properties of the corresponding nouns remained equivalent in the different languages. And so crosslinguistic differences could be explained away in terms of variations in lexical biases. Undoubtedly, less head-driven or lexically driven versions of the Constraint satisfaction model could be formulated to offer alternative ways of accounting for the data. However, at the time of writing we are unable to locate any specific models of this kind. 4.4. Construal Theory In contrast with the theories outlined above, the two final accounts assign no role at all to adjustments that might be made over time as the parser is exposed to language. Construal Theory is a refinement of the highly influential gardenpath theory developed by Frazier and her colleagues (e.g., Frazier, 1979, 1987; Frazier and Rayner, 1982). It distinguishes between primary and nonprimary attachments—roughly verb-argument-based and nonargument-based, respectively (though the formal definition is more technical than this—see Frazier and Clifton, 1996, p. 41). Attachments of the kind involved in linking an RC to its head are classified as being nonprimary. It is beyond the scope of this chapter to provide a full description of the theory. However it is maintained that nonprimary relations (including RC-attachment) are subject to a range of influences including discourse, semantic, and syntactic factors. Of these, one of the strongest is a discourse principle referred to as Relativized Relevance. This states that the parser should "preferentially construe a

322

Don C. Mitchell and Marc Brysbaert

phrase as being relevant to the main assertion of the current sentence" (Frazier, 1990, p. 321). With Sentences like (3), it is argued (Frazier and Clifton, 1996, pp. 71-83) that this and related discourse principles introduce a pressure in all languages to attach the RC to NP1 as opposed to NP2 (for other versions of the argument, see De Vincenzi and Job, 1993, 1995; and Gilboy et al., 1996). This provides an explanation of the NP1 bias found in most languages. To account for the nonstandard pattern in English, Frazier and Clifton posit a role for a rather different kind of mechanism—one based indirectly on the existence of different genitive forms in English (Frazier and Clifton, 1996, p. 74 and p. 92; see also Frazier, 1990, p. 324, for earlier discussion). The starting point of the argument is that although the Spanish language has just one grammatical device for expressing a complex genitive NP (i.e., the Norman genitive—el criado de la actrice), English has two: the Norman form ('the servant of the actress') and the Saxon form ('the actress's servant'). The next point to note is that RCs following the Saxon genitive are not subject to the attachment ambiguity that occurs in sentences using the Norman form. Consider sentence (5)—the Saxon version of our earlier example. (5) Someone shot the actress's servant who was on the balcony. In this case the relative clause can only modify the noun 'servant.' Frazier and Clifton go on to argue that readers take account of the alternative grammatical forms that the writer might have used. Assuming that writers try to follow the Gricean maxim of clarity, they argue that they would always display the unambiguous Saxon form whenever they wanted to express statements in which the RC modifies 'servant.' The Norman form would therefore be left for situations in which it was not so vital to convey the fact that clause was attached to 'servant,' and the use of this ambiguous form would therefore be taken to imply that the RC was intended to be interpreted as modifying the alternative NP ('the actress'). Playing by the rules of a Gricean contract, the reader would therefore interpret Norman sentences like (3) as signalling that the relative clause should be interpreted as being attached to 'actress' (i.e., to NP2). Frazier and Clifton (1996) assumed that this tendency is strong enough to outweigh the competing pressure for high attachment (originating from the effects of relativized Relevance), and this is therefore the essence of their explanation of the NP2 attachment preference in English. From here, just one further step in the argument is needed to explain crosslinguistic differences. In Spanish, the absence of an alternative to the Norman genitive means that there is no basis for this kind of "Gricean NP2 preference" to swamp or reverse the prevailing discourse preference for NP1 attachment. The outcome, according to this argument (henceforth dubbed the "Gricean argument" or the "Gricean hypothesis"), is that NP1 attachment should dominate in Spanish (as the empirical evidence confirms), whereas NP2 attachment should prevail in English (provided that the Gricean effect is stronger than the discourse influence).

Parsing: Recent Challenges in Dutch

323

Essentially, then, Construal's explanation of crosslinguistic RC attachment differences in Norman sentences like (3) is that they occur (rather indirectly) because the grammars of some languages (like English) offer alternative, unambiguous versions of the sentence, while others do not. 4.5. A Referential Account The final account to be considered here is based on a recent suggestion put forward by Konieczny and Hemforth (1996; Hemforth et al., in press; this volume). These authors speculate that relative clause attachment may be handled by a set of procedures that are different from those used to establish the attachment sites for other kinds of modifiers (such as prepositional phrases). Their conjecture is that these procedures draw in part on some of the processes that are used to resolve ambiguities of anaphoric reference, and they present a case for assuming that it is these operations that are particularly prone to NP1 attachment biases. Like the Gibson et al. model, this account sees attachment choice as determined by the competition between two tendencies—one favoring NP1 attachment and the other favoring NP2 attachment. To account for the fact that attachment biases in English appear to differ from those in other languages, Hemforth and her colleagues draw attention to the fact that relative pronouns are frequently omitted in English relative clauses, whereas deletion of this kind is much rarer in the other head-first languages in which modifier attachment has been investigated. They hypothesize that the optional usage of relative pronouns reduces the salience of these entities in sentence processing, and that this in turn reduces the chance that head searching is taken over by anaphoric rather than nonanaphoric machinery. The consequence is that in English the process of finding a host for an RC is less likely to be taken over by NP1-favoring anaphoric procedures than it is in any other language. Since the account is based on the readiness with which relative pronouns can be eliminated we will dub this the "RelPro drop model." 4.6. Recap According to these different proposals, the attachment of RCs to competing sites (NP1, NP2, etc.) is either determined by reference to a system that records the relative preponderance of comparable attachments in the different languages under consideration (Tuning) or somehow computes the "average distance" between heads and their arguments and uses this figure to adjust the setting of the PredProx parameter (PredProx/Recency). According to another class of exposure-based models (Constraint-Based lexicalist models), the biases are determined solely by the lexical properties of the head nouns. In contrast with these accounts, there are proposals that crosslinguistic variations in RC-attachment are produced (rather indirectly) by subtle differences in the detailed grammatical devices available in different languages. Specifically, Construal theory proposes that the NP2

324

Don C. Mitchell and Marc Brysbaert

preference in English can be traced back ultimately to the fact that this language has a Saxon as well as a Norman genitive form for expressing possessive relationships. Stressing a rather different grammatical difference, Hemforth and her colleagues base their case on the legitimacy of dropping the relative pronoun from English RCs, and they use this particular phenomenon to explain crosslinguistic differences. The rest of the chapter will examine which of these accounts is currently most viable. We will start with a general assessment of each of the proposals before moving on to a more detailed considerations of recent evidence from Dutch.

5. A PRELIMINARY EVALUATION OF THESE PROPOSALS 5.1. Tuning The Tuning hypothesis maintains that statistical factors are the sole determinants of initial RC attachment (and, indeed, the only consideration in the initial stages of any other form of ambiguity resolution). In support of this there is some evidence that NP1 attachment prevails in languages where this bias is evident in extensive corpora of the language, and that the reverse occurs where NP2 attachment dominates in corpus samples. Specifically, NP1 attachment prevails both in psycholinguistic preference studies and in corpora in Spanish (Cuetos et al., 1996), and in French (Baltazart and Kister, 1995; Mitchell et al., 1995; Zagar et al., 1997), while NP2 attachment predominates in English corpora and experiments (Corley, 1996; Cuetos et al., 1996). However, Gibson and his colleagues have recently presented evidence that suggests that Tuning on its own may not be able to account for all RC-attachment effects (Gibson and Schutze, 1996; Gibson, Schiitze and Salomon, 1996). Detailed corpus analyses of the preferences found in three-site RC-attachment ambiguities like example (4) indicate that the middle is systematically preferred to NP1 (and detailed analyses argue strongly that these biases survive whatever grain of analysis is used). However, both off-line (Gibson et al., 1996) and on-line studies (Gibson and Schutze, 1996) show that readers prefer to attach modifiers to NP1 rather than to the middle site. The indications are that they make this choice despite the fact that they are likely to have experienced a predominance of the reverse pattern in the past, which suggests at the very least that "pure" Tuning needs to be supplemented by the acknowledgment that there are at least some nonstatistical influences in determining the initial attachments of RCs to their heads. 5.2. PredProx/Recency According to the Gibson et al. (1996) model, modifier attachment is guided by just two parameters: Predicate-Proximity and Recency. The features of Recency

Parsing: Recent Challenges in Dutch

325

are assumed to be fixed across languages, leaving Predicate Proximity variation to account for any crosslinguistic variation in attachment preferences. In the current version of the model, the relative weight assigned to Predicate Proximity is assumed to be adjusted from language to language using a mechanism similar to linguistic tuning. The strength of the parameter is assumed to be high where the average distance between a verb and its arguments is high, and low in languages where these two entities typically occur in close succession. In the absence of corpus data, Gibson et al. (1996, p. 49) simply assume that English falls toward the low end of the scale, whereas the NP1-attaching languages presumably lie at the opposite extreme. Unfortunately, in its current form the proposal makes no commitment to a specific definition of verb/argument distance. This is an important shortcoming, especially considering that the "position" of an argument is uncertain in a free-order language. Nor is there any suggestion concerning the formula that relates "distance" to parameter weight. As a result, there is no way of using the framework to predict the preference order for attachment sites in any new language. To this extent the model falls short of being a fully predictive account of attachment phenomena. On a more positive note, however, it is fair to say that the model has been quite effective in predicting three-site attachments in English and Spanish (Gibson et al., 1995,1997; Gibson and Schutze, 1996; Gibson, Schutze, and Salomon, 1996).

5.3. Constraint Satisfaction Lexicalist Models In its simplest form (e.g., MacDonald et al., 1994, pp. 697-698), attachment biases are attributed to the particular lexical items that appear in the two alternative slots. Thus a constraint-satisfaction system for resolving the ambiguity might settle for NP2 attachment simply because the word 'actress' has stronger "modifier-attracting properties" than those associated with its competitor. If this were a complete account of the biases either within a language or across languages, a simple prediction would be confirmed. Namely, the attachment biases would be reversed if the nouns in the two slots were exchanged. However, it turns out that the preference pattern remains the same when the noun in slot NP1 is moved to NP2 and vice versa (Corley, 1996; see also Mitchell et al., 1995, p. 479). The model also predicts that there should be no bias at all in studies that are carefully counterbalanced so that each individual noun appears equally often in the two competing slots. In such circumstances, the more clause-attracting noun will draw the RC to the NP1 position just as often as it does to the NP2 position, leaving no residual imbalance. It turns out, however, that carefully counterbalanced studies of this kind do still show the biases that are characteristic of the languages in question (e.g., Corley, 1996; Gibson et al., 1997; Hemforth et al., in press). In short, the head-driven version of the lexicalist model does not seem to provide a satisfactory account of the data.

326

Don C. Mitchell and Marc Brysbaert

5.4. Construal Theory This theory faces problems in accounting for the bilingual data presented by Fernandez (1995). The fact that Spanish-English bilinguals systematically resolved English RC-attachment ambiguities in a different way from English monolinguals indicates emphatically that the answer to the questions of RC attachment cannot lie exclusively in any linguistic analysis of the sentences themselves. The preferences vary as a function of the linguistic background of the people reading the sentences—a clear indication that RC-attachment preferences are influenced by experience. Furthermore, evidence is accumulating that the presence of an unambiguous Saxon genitive form in a language does not necessarily result in a NP2 attachment preference in that language. On the basis of the Gricean argument, we hypothesized that Afrikaans should show an NP2-attachment preference comparable with that in English (on the grounds that it has an unambiguous Saxon genitive form that is widely used in both written and spoken text). However, an off-line study using sentences translated directly from the Cuetos/Mitchell questionnaire produced a 58% NP1-attachment preference (Swanepoel, personal communication, 1995). Further problems for Construal account will be set out shortly when we consider the implications of several recent studies carried out in Dutch. 5.5. The RelPro Drop Account Like Construal this account has some difficulty in explaining the findings reported by Fernandez (1995). If attachment were determined only by the salience of Relative Pronouns in the language under consideration, then Spanish-English bilinguals should not handle English any differently than English monolinguals. The evidence that they do suggests that readers' patterns of behavior are determined in part by their exposure to different languages in the past and not only by the characteristics of the language they are dealing with at the time of testing. The Hemforth-Konieczny account clearly needs to be elaborated to account for findings of this kind.

6. FURTHER EVIDENCE FROM DUTCH Studies in any new language raise the possibility of uncovering evidence that may bear upon both exposure-based and grammar-based accounts of crosslinguistic differences. Comparisons of on-line and corpus biases can be used to check whether the two kinds of data are compatible with Tuning and other exposurebased accounts. Equally, grammar-based accounts can be put through their paces by providing new opportunities to examine the detailed effects of grammatical variation (e.g., the existence of competing genitive forms and RelPro deletion).

Parsing: Recent Challenges in Dutch

327

We have previously reported several studies in which we have examined RCattachment biases in Dutch (Brysbaert and Mitchell, 1993; 1996a,b; Mitchell, Cuetos, Corley, and Brysbaert, 1995). As already indicated, these studies show a consistent bias in favor of NP1 attachment. This has been shown by using a direct translation of the original Cuetos/Mitchell Spanish-English questionnaire (Brysbaert and Mitchell, 1996a, Experiments I and III) and also in several eyetracking and self-paced reading experiments that feature a disambiguation region forcing attachment to one of the two competing sites (Brysbaert and Mitchell, 1996a; Experiments II and III). The type of material used in the on-line studies is illustrated in 5(a,b). (5) a. De gangsters schoten op de won van de actrice die op het balkon zat met zijn arm in het gips. 'The terrorists shot the son of the actress who was on the balcony with his arm in a cast.' b. De gangsters schoten op de won van de actrice die op het balkon zat met haar arm in het gips. 'The terrorists shot the son of the actress who was on the balcony with her arm in a cast.' The underlined continuation in (5a) forces NP1 attachment (to zoon—'son') while the corresponding portion of (5b) resolves the ambiguity in favor of NP2 attachment. The results of the studies consistently revealed that sentences containing regions forcing NP2 attachment took longer to read than those resolved in favor of NP1 attachment. We took this as evidence that the clause was attached to NP1 by the time people read the second part of the RC. Since the publication of this work (in Brysbaert and Mitchell, 1996a) we have conducted a new series of studies to test the various theoretical positions outlined above. First, we showed that the NP1 attachment preference is undiminished when the disambiguating material turns up just a word or two after the beginning of the RC, undermining accounts that maintain that the RC is initially linked to NP2 before eventually being attached to NP1 (cf. De Vincenzi and Job, 1995, as outlined above). Second, to test Tuning and other exposure-based hypotheses, we conducted corpus searches to ascertain whether our experimental NP1 attachment bias also shows up in normal language usage (Brysbaert and Mitchell, 1996b). We extracted a random sample of 675 sentences including the substructure NP-vanNP-RC from corpora based on four different Dutch newspapers and magazines (Knack, Trends, Style, and Volkskrani). Attachment to NP1 (or NP2) were rated by two independent Dutch-speaking judges. Of the sample, 469 yielded a clear attachment to one of the two sites, with a strong and consistent preference for NP2 attachment (see Table 1).

328

Don C. Mitchell and Marc Brysbaert TABLE 1 FREQUENCIES OF CLEAR ATTACHMENTS TO NP1 AND NP2 IN CORPUS SAMPLES OF NPl-van-NP2-RC STRUCTURES SAMPLED FROM FOUR DUTCH NEWSPAPERS

NP1 NP2

Knack

Trends

Style

Volkskrant

Total

60 168

30 44

36 81

18 32

144 325

Needless to say, this finding was completely unexpected on the basis of the Tuning hypothesis. This is the first study in which two-site corpus biases have failed to coincide with on-line patterns of preferences. (Recall that the discrepancies reported earlier by Gibson, Schiitze, and colleagues were based on threesite NPs) One possible way of refining the Tuning hypothesis and reconciling it with these data would be to show that there are grains of analysis for which an NP1 preference prevails in the face of the more general bias in favor of NP2 attachment. If such subclasses could be identified within the general populations of NPvaAz-NP-RC structures, then it might be argued that the Tuning device was capable of keeping a using records based on the classification features identified in this analysis (cf. Mitchell et al., 1995). Unfortunately, up to now we have not been successful in our efforts, despite the fact that we have examined an extensive range of subclassifications, generating separate tabulations, among several others, for (a) Relative pronoun—die or dat; (b) head type—human versus nonhuman; (c) corpus type—spoken versus written. In every case, we have been unable to find any subclassification that systematically produces a bias in favor of NP1 attachment. This is clearly problematic for the Tuning hypothesis. It is not easy to explain an NP1 attachment bias if samples of the language consistently show that NP2 attachments dominate in normal usage. Given our failure to define a class of substructures in which NP1 attachment dominates, a possible concern is that the materials we used in our on-line studies may have somehow been unrepresentative of the sentences that occur in "normal" texts. On this argument, the apparent discrepancy between parsing and corpus biases might be due not to the fact that exposure-based models are misguided, but to the fact that there were qualitative differences between the materials used in the experimental studies and those that turned up in the corpora. To test this possibility we are currently running an on-line study using materials sampled randomly from our corpora. If the parsing biases with these materials switch over to showing an NP2 preference, this will indicate that the materials used in our Dutch studies to date must have been subject to uncontrolled or extraneous variation. If the corpus materials continue to show an NP1 preference, then it will become difficult to

Parsing: Recent Challenges in Dutch

329

sustain the argument that attachment is determined exclusively by exposure to language. Evidence from Dutch has also thrown some light on grammar-based accounts of attachment preference. We deal first with the RelPro drop account because the argument here is relatively straightforward. The Dutch language does not allow the relative pronoun to be omitted in normal circumstances, and so, according to the Hemforth/Konieczny account, Dutch attachment preferences should pattern with those of Spanish, French, and other languages where reduction is prohibited. As indicated above, this is consistent with the on-line evidence available to date, and to that extent the Dutch data are compatible with this account. As a cautionary note, however, it is perhaps worth noting that, according to the current evidence, the salience boost for discourse processes in Dutch (relative to English) only shows up in on-line parsing biases (where the NP1 attachment preference prevails). Our corpus evidence suggests that the putative influence does not extend as far as causing newspaper authors to conform to the same biases in their written composition. When it comes to Construal theory, the findings are much less encouraging. The Gricean argument predicts that NP2 attachment should prevail in any language that shares with English the basic characteristics of Norman and Saxon genitive forms outlined above. Brysbaert and Mitchell (1996a) argued that Dutch falls into this category. The Dutch language has three genitive forms: (a) the Norman form (e.g., de hoed van vader ('the hat of father'); (b) the Saxon form (e.g., voders hoed ('father's hat') and (c) a possessive pronoun form (e.g., vader zijn hoed ('father his hat'). RCs following complex NPs of the first type are ambiguous in their attachment (as with English, Spanish, and most other languages). In the other two cases such an RC can only be attached to the second site (hoed). In other words, like English, Dutch has in addition to the ambiguous form under examination alternatives that force attachment to the site occupied by NP1 in the Norman form. It follows, therefore, that the Gricean hypothesis must predict an NP2 attachment preference in Dutch (in an argument that essentially parallels that for English). As already indicated, however, Dutch shows an NP1 attachment even with sentences that are exact translations of English materials showing the opposite bias (see Brysbaert and Mitchell, 1996a, for details). This implies that the mere presence of an unambiguous Saxon form in a language is not sufficient to cause the hypothesized Gricean influences to outweigh the competing discourse effects. The implication is that theories that draw on Gricean accounts of this kind need to be elaborated yet again before they can account for the existing data. Possible moves (suggested by Clifton, personal communication, 1996), are to argue that the Gricean shift to NP2 attachment is triggered not by the mere existence of unambiguous Saxon forms in the language, but either by the acceptability of these forms in the language in general or, alternatively,

330

Don C. Mitchell and Marc Brysbaert

by more specific information about the particular NP-complex that precedes the RC under analysis. According to the former elaboration, the Gricean influence may be countermanded because the Saxon genitive form (just like the possessive pronoun form) is used relatively rarely in written Dutch. According to the latter extension, the Gricean shift should only occur if the particular Norman genitive used in the sentence could have been expressed felicitously in an equivalent Saxon form. As we have argued before (Brysbaert and Mitchell, 1996a), the first elaboration becomes indistinguishable from the Tuning account: The Saxon genitive form has an effect only if it is frequent enough to cause a predominance of NP2 attachment in constructions with the ambiguous Norman genitive form. The second extension of Construal theory predicts a negative correlation between the acceptability of the Saxon genitive for the Norman NPs used in the individual sentences, and the likelihood of NP1 attachment. The proposal here is that the "Gricean Reasoner" does not act indiscriminately to shift the attachment from NP1 to NP2 in each and every language that happens to have an ambiguous Saxon form. Rather, it is "smart" enough to examine the particular NP-complex under consideration, and determine whether the Norman genitive in question could potentially have been expressed in the Saxon form (Clifton, personal communication, 1996). In this case alone, the Gricean shift would be set in motion. In two different experiments based on eye-tracking and self-paced reading, we have failed to find any evidence of the predicted negative correlation between acceptability of the Saxon form (as rated by a separate sample of informants) and the likelihood of NP1 attachment. These findings appear to have been corroborated independently by Frazier and Vonk (1997), who summarize the corresponding part of their own study in the following terms: "There was no greater tendency to attach to the lower NP in a complex NP . . . for which the [Saxon] alternative exists . . . than in a complex NP where there is no such alternative available" (p. 81). Therefore, on the basis of these different studies, it seems unlikely that the "smart" form of the Gricean argument will fare any better than the version that maintains that NP2 preferences prevail in any language that resembles the English "grammar of genitive forms" (see also the preliminary data for Afrikaans referred to above).

7. CONCLUSIONS FROM DUTCH The evidence from our new Dutch experiments leaves the RelPro drop account relatively unscathed, but presents problems for the other (Gricean) account, which maintains that attachment variation is produced indirectly by grammatical variation across languages. In relation to statistical accounts of the differences, the data

Parsing: Recent Challenges in Dutch

331

remain to be reconciled with the simplest exposure-based accounts (which explain the differences in terms of the prevalence of the different attachments in corpora). More complex explanations of this kind, such as PredProx/Recency theory could not be addressed because the theory is not formulated with enough precision to derive predictions about the parameter weights, and therefore attachment biases in new languages.

8. OVERALL CONCLUSIONS The work on languages other than English has made it abundantly clear that any viable theory of syntactic analysis has to postulate something over and above a proximity or recency principle to account for RC-attachment findings in most languages. To this extent it is clear that crosslinguistic work in this field has been instrumental in demonstrating the shortcomings of parsing theories that prevailed a decade or so ago. If research had confined itself to investigations of English it is possible that none of these problems would have come to light. The findings from languages other than English have prompted researchers to make a variety of interesting proposals that might replace or supplement the locality-recency principles, creating the pressure that replaces NP2 with NP1 attachment. At present, it is not clear which, if any, of these proposals provides the best basis for an account of the ways in which attachment decisions are made in different languages. The detailed explanations of crosslinguistic difference have proved in every case to be either incomplete or incorrect. In almost every case, the existing accounts would be improved if more precision or detail were added to the theories. Tuning accounts need to be more explicit about the grain of recordkeeping they rely upon. The PredProx/Recency account needs to offer an explicit specification of its parameter-setting procedures before it can be applied to new languages. Constraint-based theories need to shift away from the current emphasis on constraints imposed by heads or attachment sites, and explore the potential of nonlexicalist, structure-based systems. The Gricean argument within Construal theory needs to be reformulated so that it does not indiscriminately predict NP2 attachment for any language that has a non-Norman genitive forms, but at the same time doesn't make the wrong predictions about attachment biases in individual sentences in a language like Dutch. The RelPro Drop needs to be more explicit about exactly what it is that determines the relative influence of its two competing mechanisms for linking RCs to their heads. In addition to all of these theoretical developments, it would obviously be helpful to have a great deal more empirical data about the attachment preferences in languages that haven't been examined so far. It begins to look as if most languages

332

Don C. Mitchell and Marc Brysbaert

will show an NP1 attachment bias, but there might also be other NP2 languages like English waiting to be discovered and tested. For example, there is some informal evidence that Swedish and other Scandinavian languages may fall into this category. If further examples could be found, this would help us to place tighter limits on the range of explanations for attachment phenomena. Thus, the answers to many of the questions raised in this chapter may lie waiting to be discovered in the processing patterns that characterize the 5990 languages that have yet to be subjected to close experimental scrutiny.

ACKNOWLEDGMENTS We are grateful to Stefan Grondelaers for assistance and collaboration on the corpus studies and to Chuck Clifton and Barbara Hemforth for their stimulating comments at various stages of this work.

REFERENCES Baltazart, D., and Kister, L. (1995). Correlation entre determination et selection anaphorise dans une structure N de N. Paper presented at the seminar on "Anaphore et Reference." Nancy (CRIN), Sept 20-22. Boland, J. E., Tanenhaus, M. K., and Garnsey, S. M. (1990). Evidence for the immediate use of verb control information in sentence processing. Journal of Memory and Language, 29,413-432. Brysbaert, M., and Mitchell, D. C. (1993). Sentence parsing in Dutch: Another departure from Late Closure. Paper presented to the Dutch Psychonomic Society, Noordwijkerhout, December 18-19. Brysbaert, M., and Mitchell, D. C. (1996a). Modifier attachment in sentence processing: Evidence from Dutch. The Quarterly Journal of Experimental Psychology, 49A, 664-695. Brysbaert, M., and Mitchell, D. C. (1996b). Modifier attachment in Dutch: Deciding between Garden Path, Construal and statistical tuning accounts of parsing. Paper presented at the MAS Workshop on Computational Models of Human Syntactic Processing. Wassenaar, Netherlands, June 5-8. Carreiras, M. (1992). Estrategias de analisis sintactico en el procesamiento de frases: Cierre temprano versus cierre tardio. Cognitiva, 4, 3-27. Carreiras, M. and Clifton, C. (1993). Relative clause interpretation preferences in Spanish and English. Language and Speech, 36, 353-372. Clifton, C. Jr. (1988). Restrictions on late closure: Appearance and reality. Paper presented at the 6th Australian Language and Speech Conference. University of South Wales, Sydney, August, 1921.

Parsing: Recent Challenges in Dutch

333

Corley, M. M. B. (1996). The role of statistics in human sentence processing. Unpublished Ph.D thesis, University of Exeter. Crocker, M. W. (1992). A logical model of competence and performance in the human sentence processor. PhD Thesis, University of Edinburgh, Scotland. Cuetos, F., and Mitchell, D. C. (1988). Cross-linguistic differences in parsing: Restrictions on the use of the Late Closure strategy in Spanish. Cognition, 30, 73-05. Cuetos, F., Mitchell, D. C., and Corley, M. M. B. (1996). Parsing in different languages. In M. Carreiras, J. Garcia-Albea, and N. Sabastian-Galles (Eds.), Language processing in Spanish (pp. 145-187). Hillsdale, NJ: Erlbaum. De Vincenzi, M. and Job, R. (1993). Some observations on the universality of the lateclosure strategy. Journal of Psycholinguistic Research, 22, 189-206. De Vincenzi, M. and Job, R. (1995). An investigation of late closure: The role of syntax, thematic structure and pragmatics in initial and final interpretation. Journal of Experimental Psychology: Learning, Memory and Cognition, 21, 1303-1321. Fernandez, E. M. (1995). Processing strategies in second language acquisition: Some preliminary results. Paper presented at GAS LA, May 4-5. Frazier, L. (1979). On comprehending sentences: Syntactic parsing strategies. Ph.D Dissertation, University of Connecticut. Bloomington: Indiana University Linguistics Club. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance XII (pp. 601-681). Hillsdale, N.J.: Erlbaum. Frazier, L. (1990). Parsing modifiers: Special purpose routines in the human sentence processing mechanism? In D. A. Balota, G.B. Flores d'Arcais, and K. Rayner (Eds.), Comprehension processes in reading (pp. 303-330). Hillsdale, NJ: Erlbaum. Frazier, L., and Clifton, C. Jr. (1996). Construal. Boston: MIT Press. Frazier, L., and Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178-210. Frazier, L., and Vonk, W. (1997) Modifier attachment. In J. P. de Ruiter and D. Wilkins (Eds.) Max-Planck-Institut fur Psycholinguistik: Annual Report, 1996 (p.81). Nijmegen, Netherlands: Max-Planck-Institut fur Psycholinguistik Gibson, E., Pearlmutter, N., Canseco-Gonzalez, E., and Hickok, G. (1996). Cross-linguistic attachment preferences: Evidence from English and Spanish. Cognition, 59, 23-59. Gibson, E., Pearlmutter, N. J., and Torrens, V. (1997). Recency and predicate proximity in sentence comprehension. Poster presented at the Tenth Annual CUNY Conference on Human Sentence Processing, Santa Monica, March 20 -22. Gibson, E., and Schutze, C. T. (1996). The relationship between the frequency and the perceived complexity of conjunction attachments: On-line evidence. Poster presented at the Ninth Annual Cuny Conference on Human Sentence Processing, New York, March 21-23. Gibson, E., Schutze, C., and and Salomon, A. (1996). The relationship between the frequency and the complexity of linguistic structure. Journal of Psycholinguistic Research, 25, 59-92. Gilboy, E., Sopena, J. M., Clifton, C. Jr., and Frazier, L. (1995). Argument structure and association preferences in Spanish and English complex NPs. Cognition, 54,131-167. Hemforth, B., and Konieczny, L. (1996). Preferences in the parser and in the discourse model. Paper presented at AMLaP-96, Turino, Italy, Sept 20-21.

334

Don C. Mitchell and Marc Brysbaert

Hemforth, B, Konieczny, L., and Scheepers, C. (1994). Principle-based and probabilistic approaches to human parsing: How universal is the human language processor? In H. Trost (Ed.), Tagungsband KONVENS '94 (pp. 161-170). Berlin: Springer. Hemforth, B, Konieczny., L., and Scheepers, C. (in press). Syntactic attachment and anaphor resolution: Two sides of relative clause attachment. In M. Crocker, M. Pickering and C. Clifton, (Eds.), Architectures and Mechanisms for Language Processing. Henstra, J. (1996). Relative clause attachment in English: Eye-tracking versus self-paced reading. Poster presented at AMLaP-96, Turino, Italy, Sept 20-21. Kamide, Y., and Mitchell, D. C. (1997). Relative clause attachment: Non-determinism in Japanese parsing. Journal of Psycholinguistic Research, 26, 247-254. Konieczny, L., and Hemforth, B. (1996). A visibility-based model of human sentence processing. Paper presented at the MAS Workshop on Computational Models of Human Syntactic Processing. Wassenaar, Netherlands, June 5-8. Konieczny, L., Hemforth, B., and Scheepers (1994). Reanalysis vs. internal repairs: Nonmonotonic processes in sentence perception. In B. Hemforth, L. Konieczny, C. Scheepers, and G. Strube (Eds.), First analysis, reanalysis, and repair (pp. 1-22). IIGBerichte 8/94. Kempen, G. (1996). Computational models of syntactic processing in human language comprehension. In A. Dijkstra and K. de Smedt (Eds.), Computational Psycholinguistics: Symbolic and subsymbolic models of language processing. Hemel Hempstead, UK: Harvester Wheatsheaf. Kempen, G., and Vosse, Th. (1989) Incremental syntactic tree formation in human sentence processing: A cognitive architecture based on activation decay and simulated annealing. Connection Science, 1, 273-290. Lewis, R. (1993). An architecturally-based theory of human sentence comprehension. PhD Thesis, Carnegie Mellan University, Pittsburgh, PA. MacDonald, M. C., Pearlmutter, N. J., and Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101, 676-703. MacDonald, M. C., and Thornton, R. B. (1996). Constraint-based models and modification ambiguities. Poster presented at the Ninth Annual Cuny Conference on Human Sentence Processing, New York, March 21-23. McClelland, J. L., St. John, M., and Taraban, R. (1989). Sentence comprehansion: A parallel distributed processing approach. Language and Cognitive Processes, 4, 287335. Mitchell, D. C. (1994a). Sentence parsing. In M. Gernsbacher (Ed.) Handbook of psycholinguistics (pp 375-409). San Diego: Academic Press. Mitchell, D. C., and Cuetos, F. (1991a) The origins of parsing strategies. In C. Smith (Ed.), Current Issues in Natural Language Processing. Center for Cognitive Science, University of Texas, Austin. Mitchell, D. C., and Cuetos, F. (1991b). Restrictions on late closure: The computational underpinnings of parsing strategies in Spanish and Engligh. Unpublished manuscript, University of Exeter, UK. Mitchell, D. C., Cuetos, F., and Zagar, D. (1990). Reading in different languages: Is there a universal mechanism for parsing sentences? In D. Balota, G. B. Flores d'Arcais, and K. Rayner, (Eds.), Comprehension processes in reading (pp. 285-302). Hillsdale, N.J.: Erlbaum.

Parsing: Recent Challenges in Dutch

335

Mitchell, D. C., Cuetos, F., Corley, M. M. B., and Brysbaert, M. (1995). Exposure-based models of human parsing: Evidence for the use of coarse-grained (nonlexical) statistical records. Journal of Psycholinguistic Research, 24, 469-488. Pritchett, B. L. (1992). Grammatical competence and parsing performance. Chicago: University of Chicago Press. Pynte, J., and Frenck-Mestre, C. (1996). Evidence for early-closure attachments on firstpass reading times in French: A replication. Poster presented at AMLaP-96, Turino, Italy, Sept 20-21. St. John, M. F, and McClelland, J. L. (1990). Learning and applying contextual constraints in sentence comprehension. Artificial Intelligence, 46, 217-257. Tanenhaus, M. K., Carlson, G., and Trueswell, J. C. (1989). The role of thematic structures in interpretation and parsing. Language and Cognitive Processes, 4, 211-234. Thornton, R., Gil, M., and MacDonald, M. C. (1995). Prepositional phrase attachment in English and Spanish. Poster presented at the Eight Annual Cuny Conference on Human Sentence Processing, Tucson, Az, March 16-18. Zagar, D., Pynte, J. and Rativeau, S. (1997). Evidence for early closure attachment on firstpass reading times in French. Quarterly Journal of Experimental Psychology, 50A, 421-438.

This page intentionally left blank

SYNTACTICALLY BASED PARSING STRATEGIES: EVIDENCE FROM TYPOLOGICALLY DIFFERENT LANGUAGES MARICA DE VINCENZI Institute of Psychology of the National Research Council Rome, Italy

1. INTRODUCTION Crosslinguistic investigation is a very popular topic in sentence processing (see the volumes of Slobin, 1985, and MacWhinney and Bates, 1989). This is certainly a welcome feature in the parsing studies, at least if the goal of the theories is to have a theory of human sentence processing and not merely the description of processing a single language. In this chapter I would like to illustrate how crosslinguistic studies can motivate the formulation of a parsing strategy, and second, how methodological considerations can affect the comparisons among languages. The parsing strategies I consider are syntactically based strategies: with this term I mean strategies that are defined over the geometry of the phrase marker of a sentence, regardless of what particular phrase types and languages are involved. Examples of such strategies are the Right Association (Kimball, 1973), Minimal Attachment and the Late Closure strategy (Frazier and Fodor, 1978), the Superstrategy (Fodor, 1979), the Recent Filler Strategy (Frazier, Clifton, and Randall, 1983), the Active Filler Strategy (Frazier, 1987). The basic idea in all these strategies is that they are directly derived from a simple principle: Choose to do whatever costs the least effort in terms of computation. This choice is derived by a Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

337

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

338

Marica De Vincenzi

basic cognitive reason, namely, the restrictions on short-term memory (STM) in terms of memory and computational space and the fact that the more structured the material to be stored, the smaller the demand it makes on the STM space (Miller, 1956). Given that these principles are based on cognitive needs and that they are not tied to any language-specific aspect, we expect them to be universal. For these kinds of strategies, therefore, a crosslinguistic test is crucial to assess their validity, even though they have been initially formulated and tested in English. In the following I would like to present a syntactically based strategy, the Minimal Chain Principal (MCP), which has been developed to account also for the processing of the so-called null-subject languages, like Italian, and I will show how it accounts for parsing in different languages.

2. THE MINIMAL CHAIN PRINCIPLE When we extend the application of a processing strategy to a new language, there can be the need to further specify the strategy to include the parsing of structures that are not present in the previously studied languages (cf. De Vincenzi, in press, for a discussion). For example, a processing strategy could be reconducted to a more abstract level of generalization, so that the principle applies to different structures, regardless of the form of the superficial string but according, say, to the underlying level of syntactic representation. Italian, and some other natural languages, allow phonetically null subjects in tensed clauses, as shown in (1). Another property systematically correlates with the null-subject property (Perlmutter, 1978; Rizzi, 1982): a free process of subject inversion, that is, the subject can appear after the verb (2). In the following examples, "pro" indicates the lexically null pronominal subjects. English glosses are given below the Italian examples; the English translation is in quotes. (1) pro telefonera' pro will telephone 'he/she will telephone' (2) proi telefonera' Giannii prOi will telephone Giannii 'Gianni will telephone' This means that in Italian (and in the other languages that adopt this linguistic parameter) there are linguistic dependencies not present in English. Therefore, the parsing principles referring to movement dependencies and empty elements originally formulated on the basis of English had to be revised to include the types of

Syntactically Based Parsing Strategies

339

structures mentioned above. To this aim, De Vincenzi (1991) proposes the MCP, paraphrased below. Minimal Chain Principle: To postulate required chain members at the earliest point grammatically possible but to postulate no potentially unnecessary chain members. The MCP manages to unify the processing of distinct types of empty categories: "pro"—the null pronominal that occurs in case-marked positions; "NP-trace"— the null anaphor bound by a phrase in argument position; and "wh-trace"—the null category (variable) bound by a phrase in an operator position. This permits the processing of null subject languages like Italian to be unified with the processing of overt subject languages like English.1

3. THE MINIMAL CHAIN PRINCIPLE AND THE PROCESSING OF DECLARATIVE SENTENCES One prediction of the MCP is that the parser prefers postulating a singleton chain to postulating multimember chains. It predicts that in Italian in cases of ambiguities of the null-subject position the parser will prefer to postulate a "pro" that is in a singleton chain, as in (3a), to a "pro" that is in a longer chain, as in (3b). (3)

Ha chiamato Gianni a. pro Will call Gianni 'he/she/it will call Gianni' b. pro. Will call Giannii. 'Gianni will call'

Notice that the preference to analyze an element as being in a one-member chain amounts to saying that the parser prefers to analyze an element as being in its deep-structure position, that is, in the position where it directly receives a thematic role. This part of the MCP is not a new idea in parsing strategies: in fact, this was Fodor's (1979) "Superstrategy" proposal that the parser "processes a word sequence as if it were the terminal string of a well-formed deep structure" (Fodor, 1979, p. 249). This part of the MCP has been confirmed by experimental evidence in both Italian and Russian. In Italian, sentences like (4) are ambiguous between having the interpretation (4a) with a transitive use of the verb, and the noun 'the seller' is the direct object, or the interpretation (4b) with an intransitive use of the verb, and therefore 'the seller' is the Inverted subject. The control condition was (4c): the verb is intransitive, so the postverbal NP can only be an inverted subject. The

340

Marica De Vincenzi

predictions were that if the parser follows the MCP, then the seller should be initially taken as direct object in (4a and 4b). Therefore, when subsequent information disambiguates the sentence as inverted subject, as in (4b), there should be longer reading time in (4b) than in (4a) due to a reanalysis process. The results confirmed the predictions, showing that the reading times on the dependent clause (to ask/offer a discount), which disambiguates the object or subject analysis of the postverbal noun, are significantly longer only in (4b) (the inverted subject condition) than in (a) or (c). (4) a. Ha richiamato il venditore per chiedere uno sconto. b. Ha richiamato il venditore per offrire uno sconto. c. Ha insistito il venditore per offrire uno sconto. a. Called the seller to ask for a discount, 'he/she called the seller to ask for a discount' b. Called the seller to offer a discount. "The seller called to offer a discount' c. Insisted the seller to offer a discount. 'The seller insisted to offer a discount' In Russian, as in Italian, it is possible to have inverted subject. Sekerina (1995) has shown the same preferences in processing declaratives with a subject-object ambiguity: namely, reading times were faster in (5a), where the preverbal noun is subject, than in (5b), where the preverbal noun has to be analyzed as an inverted subject. (5)

Trolleybus obonal avtobus Trolleybus-NOM/ACC passed bus-NOM/ACC a. 'The trolleybus passed the bus.' b. 'The bus passed the trolleybus.'

Notice that the Russian nouns used by Sekerina (trolleybus and bus) have ambiguous case marking and therefore sentences like (5) were ambiguous. It would be interesting to use Russian to test whether in cases where a declarative sentence is disambiguated by case-marking information, there is still a penalty for the structures with more complex chains, namely, the inverted subject construction. The interest of this experiment would consist not only in knowing something more about processing Russian, but in testing whether, regardless of a temporary ambiguity, the human sentence processor is sensitive to the complexity of an underlying syntactic representation. The goal of the experiment would not be to test if the parser prefers to build the shortest chain, given more than one possibility, but if, given two structures with unambiguously warranted nodes, the structure with a shorter chain is parsed faster than the one with a longer chain. We can find exactly this latter prediction tested in German (Hemforth, Konieczny, and Strube, 1993) and in Italian (De Vincenzi, 1991), using quite different

Syntactically Based Parsing Strategies

341

structures, given that the two languages differ in case marking. German, similar to Russian, has case marking. Depending on the gender of the noun, the determiner can be unambiguously nominative or accusative.2 In a self-paced reading experiment, Hemforth et al. used declarative sentences of the form Noun-Verb-Noun with an unambiguously case-marked subject and object, as in (6). (6) a. Der (Nom) hungrige Fuchs bemerkte den (Acc)fetten Hahn. 'The hungry fox (nom) noticed the fat chicken, (acc)' b. Den (Acc) hungrigen Fuchs bemerkte der (Nom) fetten Hahn. The hungry fox (acc) noticed the fat chicken (nom) The fat chicken noticed the hungry fox' The results showed that despite the unambiguous case marking, there are faster reading time for the preverbal subject structure (6a), which has less complex chain than the preverbal object structure (6b). The chain complexity hypothesis has been tested in unambiguous sentences also in Italian (De Vincenzi, 1991), but in a different structure given that Italian does not have case marking. We used declarative sentences with intransitive verbs. The manipulation involved the type of verb: namely, two classes of intransitive verbs with different structural representations: unaccusative verbs that have the subject originating in postverbal position, or unergative verbs that have the subject originating in preverbal position (Burzio, 1986; Perlmutter, 1978). The result is that different syntactic representations correspond to two similar strings of words. The experiment capitalized on the fact that intransitive verbs in Italian are easily identified as belonging to the class of either unaccusative or unergative verbs due to the selection of different auxiliary verbs in compound tenses. (7) illustrates the structural difference among the unaccusative verb arrivare ('to arrive') and the unergative verb esitare ('to hesitate'): (7) a. proi E' arrivato ei. Is arrived, 'she/he/it arrived' b. pro Ha esitato. Has hesitated, 'she/he/it hesitated' Therefore, a postverbal subject of an unaccusative verb (arrive) may receive case and thematic role in its deep-structure position. By contrast, subjects of unergative verbs (Hesitate) must be related to the preverbal subject position to receive case and thematic role. Hence, with an indefinite postverbal subject, there is a singleton chain only with the unaccusatives. Therefore, if the human sentence processor is sensitive to the syntactic complexity, we should find that Italian sen-

342

Marica De Vincenzi

tences with postverbal subjects should be processed faster when the verb is unaccusative than when it is unergative. To test this prediction we set up an experiment that used sentences as in (8). As shown by the figures, the conditions in (b), unergatives, have a more complex chains than the conditions in (a), with unaccusative verbs. (8) a. e' arrivato un amico per salutare Luciana. 'arrived a friend to greet Luciana' b. ha telefonato un amico per salutare Luciana. 'telephoned a friend to greet Luciana'

The results confirmed the predictions, showing that unaccusatives were read much faster than unergatives. Given that all the other relevant factors, such as frequency and length of the verb, were controlled for, we must conclude that the specific increased difficulty for the unergatives was due to the structural representation that has a more complex chain than unaccusatives. Therefore, the results support the idea that chain formation is a costly process for the human parser.

4.4. CONCLUSIONS In this brief exposition, I hope to have illustrated how a syntactically based parsing strategy can be tested not only within a single language but across languages. A parsing strategy that wants to be, at least in principle, a good candidate

Syntactically Based Parsing Strategies

343

to be universal should be formulated at a level such that it can apply to languages that differ typologically. This was the driving motivation to formulate the MCP, to account for parsing preferences in languages with or without the null-subject property. We have seen that the MCP can account for the preferred interpretation of null subject and postverbal subject sentences in Italian. The second point I wanted to make is more methodological: how to test the universality of a syntactic parsing principle. The easiest way to test a specific parsing principle is to study the same structure in two languages that express it similarly. However, it is rarely the case that a given structure does not interact with some other relevant linguistic properties in which the two languages differ. This should caution the researcher when testing the same structure crosslinguistically (see the debate on the Late Closure strategy: cf. Baccino, De Vincenzi, and Job, (in press) for a discussion of the methodological differences). However, if properly taken into account, these interactions can be extremely insightful. For example, either Russian, German, .or Italian have inverted subjects. However, only Russian and German have case marking. This feature has allowed to test the abstract linguistic property at the base of the MCP, using unambiguously case-marked structures. Therefore, the MCP is supported not only as a principle driving the parser decisions at points of ambiguity, but also as a principle accounting for processing complexity.

REFERENCES Baccino, T., De Vincenzi, M., and Job, R. (in press). Cross-linguistic studies of the Late Closure strategy: French and Italian. In Proceeding of the AMLap96 (Architectures and Mechanisms for Language Processing) Conference., Torino, September 1996. Burzio L. (1986). Italian syntax. Dordrecht: Reidel. De Vincenzi, M. (1991). Syntactic Parsing Strategies in Italian. Dordrecht, Holland: Kluwer Academic Publishers. De Vincenzi, M. (1996). Syntactic Analysis in Sentence Comprehension: Effects of Dependency Types and Grammatical Constraints. Journal of Psycholinguistics Research, 25,1,117-133. De Vincenzi, M. (in press). Reanalysis aspects of movement. In J. Fodor and F. Ferreira (Eds.), Reanalysis in Sentence Processing. Dordrecht, Holland: Kluwer Academic Publishers. Fodor, J. D. (1979). Superstrategy. In W. E. Cooper and E. C. T. Walker (Eds.), Sentence processing. Hillsdale, N.J.: Erlbaum Press. Frazier, L. (1987). Processing syntactic structures: Evidence from Dutch. Natural Language and Linguistic Theory, 5, 519-559. Frazier, L., and Fodor, J. D. (1978). The sausage machine: A new two-stage parsing model. Cognition, 6, 291-326.

344

Marica De Vincenzi

Frazier, L., Clifton, C., and Randall, J., (1983). Filling gaps: Decision principles and structure in sentence comprehension. Cognition, 13, 187-222. Hemforth, B., Konieczy, L., and Strube, G. (1993). Incremental syntax processing and parsing strategies. Proceedings of the XVth Annual Conference of the Cognitive Science Society, (pp. 539-544) Hillsdale, NJ: Lawrence Erlbaum. Kimball, J. (1973). Seven principles of surface structure parsing in natural language, Cognition, 2, 15-47. MacWhinney, B., and Bates, E. (1989). The cross-linguistic study of sentence processing. Cambridge, UK: Cambridge University Press. Perlmutter, D. (1978). Impersonal passive and the unaccusative hypothesis. Proceedings of the Berkeley Linguistics Society, 4. Rizzi, L. (1982). Issues in Italian syntax. Dordrecht: Foris. Sekerina, I. (1995). Ambiguity and scrambling in Russian syntactic processing. Paper presented at 8th CUNY Conference on Human Sentence Processing. Slobin, D. I. (1985). The cross-linguistic study of language acquisition. Volume 1: The data. Hillsdale, NJ: Lawrence Erlbaum.

NOTES 1

Given that the MCP abstracts away from the type of elements involved (a wh-element, a NP element, a null pro element), it unifies the processing of declarative and interrogative sentences. However, in the present work I will focus only on declarative sentences, and I refer the reader to De Vincenzi, (1996, in press) for the processing of wh-questions. 2 The case marker on the determiner becomes unambiguous when the full noun phrase is read, because only at that point is the gender of the noun known. For example, the determiner der in isolation can be ambiguous between nonminative case for a masculine noun and dative case for a feminine noun. So the ambiguity in (6a) is resolved by the suffix on the adjective: hungrige can be only nominative.

THE DEVELOPMENT OF SENTENCE COMPREHENSION IN ITALIAN AND SERBOCROATIAN: LOCAL VERSUS DISTRIBUTED CUES ANTONELLA DEVESCOVI* SIMONETTA D'AMICO* STAN SMITH* IVO MIMICA ELIZABETH BATES * Dipartimento di Psicologia dei Processi di Sviluppo e Socializzazione Universita La Sapienza Roma, Italy Wheeler Cognitive Neuropsychology Laboratory Good Samaritan Hospital and Medical Center Portland, Oregon * Center for Research on Language University of California, San Diego La Jolla, California

Crosslinguistic research on language development has focused primarily on one major question: what factors account for the order in which elements of grammar are acquired, within and across languages? This ordinal question presupposes two further interval questions. First, how much and what kind of evidence do we need to determine when a given element has been acquired (i.e., the problem of productivity)? Second, what factors control the size of the interval between ordered acquisitions (i.e., what constitutes "very early" and/or "very late") ? In this chapter, we will build on previous crosslinguistic studies of the ordinal question, concentrating on two richly inflected languages, Italian and Serbo-Croatian. We will Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

345

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

346

A. Devescovi et al.

focus on the use of morphological agreement in sentence comprehension in an attempt to account for some puzzling disparities in timing that are germane to the two interval questions just described: why are some aspects of grammatical comprehension relatively late to develop, and when are we justified in concluding that those forms are "really there" ?

1. BACKGROUND During the 1970s, crosslinguistic research was inspired primarily by a search for universal sequences of language development. Many of these studies were also heavily influenced by the facts of development in English, reflected in the hypothesis that inflectional morphology must be acquired after basic word order patterns are laid down (McNeill, 1975). For example, consider the assumption that all children pass through a "telegraphic stage": they produce strings of uninflected words in the basic and/or most frequent word order in their language, and they rely primarily on word order cues (as opposed to inflectional morphology) to interpret adult speech (Pinker, 1982). Indeed, some scholars have gone so far as to claim that certain word orders (especially SVO and SOV) are more "natural" for small children, either because these word orders reproduce the "natural" order of events in thought (Osgood and Tanz, 1977), or because these word orders stand as initial hypotheses in an innate Language Bioprogram (McNeill, 1975; Bickerton, 1984). Hence SVO and SOV should be the first word orders to appear in spontaneous speech, and languages with a different "canonical" word order should prove more difficult to learn. More recent work on the acquisition of highly inflected languages has forced a reconsideration of this "basic word order" hypothesis. The first doubts were raised early on by Bowerman (1973) and Braine (1976). Evidence against the primacy of word order accumulated rapidly across the 1980s (Frankel and Arbel, 1981; Hakuta, 1982; MacWhinney, Pleh, and Bates, 1985; Sinclair and Bronkart, 1972; Slobin, 1985; Sokolov, 1988; Weist, 1983). In a recantation of his earlier position, McNeill (1979, p. 245) now states that "the idea of an inherent word order loses credibility." The most remarkable data in this respect are those concerning Turkish, an inflectional language whose case morphemes are regular (i.e., without morphophonological subclasses and/or irregular forms), phonologically salient (carrying stress), semantically transparent (e.g., there are no arbitrary gender contrasts), and highly frequent (they are applied not only to nouns but also to pronouns, demonstratives, nominalizations, and interrogatives). As noted by Slobin and Bever (1982), although SOV is the unmarked word order in Turkish, all other combinations are possible; for example, in a corpus of 500 sentences addressed to Turkish children, all word orders except VOS occurred, and the sub-

Sentence Comprehension in Italian and Serbo-Croatian

347

ject preceded the object only 47% of the time. Turkish children learn quite early in life how to use these different word orders: in a study on spontaneous production carried on 14 children from 2.2 to 3.8 years of age, all word orders of adults' language occurred and with the same frequency rate (see also Bowerman, 1973, on the acquisition of Finnish, a language that is closely related to Turkish). However, they also begin to mark case productively very early in their multiword speech (i.e., there is no "telegraphic stage"), and they appear to rely heavily on case cues rather than word order in a test of sentence interpretation. In fact, Slobin and Bever have shown that 2-year-old Turkish children have already reached adult levels in the use of the nominative/accusative contrast as a cue to sentence meaning. Faced with results of this kind, Slobin (1982) concludes that we should take into greater consideration "the diversity of human languages and the varying acquisition tasks posed by languages of different sorts" (p. 129). In fact "language, in itself, constitutes a complex body of knowledge which must be discovered and structured on its own terms" (p. 129). In this vein, Slobin (1985) proposed that language acquisition is directed not by universal content (e.g., the word-order-first hypothesis) but by universal strategies or "operating principles" that direct children to the most regular and consistent morphosyntactic information in their mother tongue. The order in which grammatical elements are acquired will therefore depend on the clarity, regularity, and information value of individual aspects of morphology and syntax in the child's target language. Bates and MacWhinney (1982, 1987) have taken these recommendations several steps further in an attempt to quantify some of the ideas that are latent in Slobin's list of operating principles. Specifically, they suggest that the order in which aspects of grammar are acquired in a given language depends primarily on the "cue validity" or information value of each cue as a determinant of underlying sentence meaning. This is the central organizing principle within a more general functionalistic account of language acquisition and performance called the Competition Model, in which two levels of representation are specified: a functional level (where the meanings and intentions expressed in a sentence are represented) and a formal level (which contains the surface forms that best express those meanings). Cue validity is a measure of the quantitative relationship between these two linguistic levels. For any given form (e.g., nominative case) and its corresponding function (e.g., the agent role), cue validity is calculated as the product of availability (e.g., the proportion of sentences in which a nominative inflection is available as a cue to the agent role) and reliability (e.g., the number of instances in which a nominative case marking correctly assigns the agent role, divided by all those instances in which a nominative inflection is available somewhere in the sentence). Within comprehension, the Competition Model assumes a dynamic process of form-function mapping, in which alternative interpretations of the input are activated in parallel; the ensuing competition between these mappings is

348

A. Devescovi et al.

resolved through a computational process in which the cue validities for each form-function mapping are combined (working from left to right) until a "winner" emerges. This model is similar to proposals by Just and Carpenter (1980) and McClelland and Rumelhart (1981) for word recognition, and to more recent connectionist models of morphosyntactic processing (St. John and McClelland, in press; Elman, 1989). The Competition Model also makes a strong prediction about the order in which cues to sentence meaning will be acquired: specifically, children will acquire elements of grammar in an order that directly reflects the relative validity or information value of each element in their native language. In a volume summarizing research within this model, MacWhinney, Bates, and their colleagues have provided substantial evidence in favor of this hypothesis (MacWhinney and Bates, 1989). However, there are also a number of interesting exceptions to predictions based on cue validity. In most cases, these exceptions can be attributed to limits on the efficiency of information processing that Bates and MacWhinney (1987, 1989) summarized under the term cue cost. For instance, MacWhinney, Pleh, and Bates (1985) have shown that Hungarian children are affected by variations in the relative phonetic salience or perceivability of case cues: They first reach adult levels in the use of accusative marking to assign sentence meaning only for those case cues that follow a strong vowel (e.g., nominative = PIPA; accusative = PIPAT); there is a delay of approximately 1 year before these children can make reliable use of accusative markers that occur at the end of a consonant cluster (e.g., nominative = MOKUS; accusative = MOKUST). A different kind of cue cost factor is suggested by Kail (1989), based on a study by Kail and Charvillat (1988) showing that French children cannot or do not make consistent use of morphological agreement between a noun and a clitic pronoun to assign sentence roles before 6 years of age. Because this form of agreement is high in cue validity and serves as a powerful cue to sentence meaning for French adults, a marked delay in the use of agreement cues by French children constitutes an important exception to the principle of cue validity. To account for this exception, Kail makes a distinction between "local" cues (which can be interpreted immediately after encoding) and "global" cues (which cannot be used to make a sentence interpretation until a great deal of distributed information is received, stored, and compared). In line with an earlier proposal by Ammon and Slobin (1979; see also Johnston and Slobin, 1979), Kail suggests that "global" or distributed cues may be particularly difficult for young children because of the demands they place on short-term memory. Bates and MacWhinney (1987, 1989) have incorporated this proposal into the cue-cost component of the Competition Model, referred to as the problem of assignability. The purpose of the present study is to verify the hypothesis that morphological cues that require "local" processing, such as the case system, are learned more rapidly than those that require the "global" processing of agreement cues that

Sentence Comprehension in Italian and Serbo-Croatian

349

are distributed across several different elements in a sentence or clause. To this end, we have chosen to study the development of sentence comprehension in two languages (Italian and Serbo-Croatian), with morphological characteristics that permit a clear comparison between these two kinds of processing. Experiment 1 (in Serbo-Croatian) will demonstrate that local cues are acquired before global cues in a language that makes extensive use of both forms of morphology. Experiment 2 (in Italian) will show that there are great delays in the use of subjectverb agreement, despite the fact that this is the single most valid and important cue to sentence meaning in the language. Finally, Experiment 3 (in Italian) will show that these marked delays in the use of agreement cues do indeed reflect the high processing costs associated with this form of morphology; in a situation in which agreement cues can be used locally (without waiting for additional information), children below the age of 3 can and do make extensive use of subjectverb agreement to assign semantic roles.

2. OVERVIEW OF ITALIAN AND SERBO-CROATIAN GRAMMAR 2.1. Serbo-Croatian The Serbo-Croatian case system is so complex and so irregular that it poses a serious challenge to Serbo-Croatian children (Slobin and Bever, 1982). First, in contrast with so-called agglutinative languages like Turkish or Hungarian (where inflections line up like beads on a string), the inflectional system of SerboCroatian is fusional (a single morpheme may stand for several different morphological contrasts). According to Slobin (1982), children prefer agglutinative languages like Turkish, in which there is a one-to-one relationship between form and function. Second, there are many cases of irregularity and homonymy in SerboCroatian, a fact that makes it difficult for children to isolate stems. Third, case assignment interacts with several other dimensions, including gender and animacy. There are seven case inflections in Serbo-Croatian, in three genders, yielding a paradigm of 21 case-by-gender combinations. For feminine and animate masculine nouns, there is a clear morphological contrast between nominative and accusative forms; by contrast, neuter and inanimate masculine nouns take the same form in the nominative and accusative case. In Slobin's opinion, it should be more difficult for children to learn a case system of this kind, compared with the acquisition of a clear and transparent language like Turkish. Despite these problems, case morphology is still the most valid source of information about sentence meaning in this language. Indeed, Smith and Mimica (1984) and Smith and Bates (1987) have shown that adult speakers of SerboCroatian use case contrasts to assign agent-object relations whenever such a contrast is available, even in the face of competing information from semantics and

350

A. Devescovi et al.

word order. Nevertheless, research by Slobin and Bever (1982) bears out Slobin's prediction that Serbo-Croatian children will master the use of their difficult case system in sentence comprehension several years beyond the point at which Turkish children have begun to perform at adult levels. Serbo-Croatian listeners can also make use of other morphological cues to agent-object relations, to supplement their ambiguous case system. In the SerboCroatian conversational past tense, main verbs carry inflections that agree in gender with the subject of the verb. When there is a gender contrast between the subject and direct object of a sentence, gender agreement is a maximally reliable cue to agent-object relations. However, gender agreement is only moderately available as a cue to agent-object relations for two reasons: first (as noted), gender agreement does not occur except in the conversational past tense; second, it is useful as a cue only when there is a gender contrast between agent and object. Smith and Bates have shown that adult listeners only make use of gender agreement when case marking is ambiguous; in those instances, however, gender agreement is a powerful cue that determines sentence meaning at close to ceiling levels, overcoming competing semantic and/or word-order cues. 2.2. Italian Italian has lost the Latin system of case morphology; the only vestiges of case marking that remain are the nominative/accusative/dative distinctions in the system of personal pronouns (similar to the "I" versus "me" contrast in English). This fact has led some psycholinguists to classify English and Italian together in studies of sentence comprehension and production (Slobin, 1982, 1985; Slobin and Bever, 1982). SVO is also the canonical sentence order in both these languages. However, like many other inflectional languages (but unlike English), Italian allows a great deal of word order variability, especially (but not only) in informal conversation. In order to understand the meaning of many sentences in informal speech, Italian listeners have to rely on other sources of information. In particular, MacWhinney, Bates, and Kliegl (1984) have shown that Italians rely heavily on a rich system of morphological agreement to interpret simple transitive sentences. First of all, in Italian the subject agrees in person and number with the verb. This is another feature Italian shares with English, but the Italian system of agreement is considerably richer and more informative than the English (see Table 1). Within the simple present tense (indicative mood), all six combinations of person (first, second, third) and number (singular, plural) are marked contrastively on the verb. For most of the other tense, aspect, and mood conditions in the Italian verb conjugation paradigm, the same six contrasts are preserved. Although this means that Italian children have a heavy load of verb morphology to learn (Bates, 1976; Pizzuto and Caselli, 1989), it also means that listeners can usually identify the

Sentence Comprehension in Italian and Serbo-Croatian

351

TABLE 1 DESCRIPTION OF WITHIN-SUBJECT VARIABLES Word order NVN = noun-verb-noun NNV = noun-noun-verb VNN = verb-noun-noun Agreement AGO = both nouns agree AG1 = first noun agrees AG2 = second noun agrees Animacy AA = both nouns animate AI = animate first noun LA = animate second noun

// cane spinge il gatto. 'The dog pushes the cat.' // cane il gatto spinge. 'The dog the cat pushes.' Spinge il cane il gatto. 'Pushes the dog the cat.' I cani spingono i gatti. 'The dogs push the cats.' I cani spingono il gatto. 'The dogs push the cat.' I cani spinge il gatto. 'The dogs pushes the cat.' Il cane spinge il gatto. 'The dog pushes the cat.' Il cane spinge la matita. 'The dog pushes the pencil.' La matita spinge il gatto. 'The pencil pushes the cat.'

subject of a sentence entirely on the basis of verb morphology. This may be one reason why Italian permits so much word order variation; it may also help to explain why overt subjects are (legally) omitted from sentences in informal Italian conversation approximately 70% of the time (Bates, 1976). MacWhinney, Bates, and Kliegl (1984) have shown that these contrasts between English and Italian have a dramatic impact on the interpretation of simple sentences by normal adults. Given a sentence like (1), a large majority of English listeners choose the cat as the subject of the sentence, whereas 95% of Italians choose the dogs as the subject. (1) Il gatto inseguono i cani. 'The cat are chasing the dogs.' Second, the article and all noun modifiers agree in gender and number with the noun itself, illustrated in sentence (2): (2)

Una macchina vecchia puo' essere pericolosa 'A (sing, fern.) car (sing, fern.) old (sing, fern.) can (3rd. sing.) be dangerous (sing, fem.)'

This means that Italian listeners have a number of morphological cues available to ensure the identity of coreferential noun phrases. In addition, there are two sets

352

A. Devescovi et al.

of personal pronouns (a strong form like io, tu, lui/lei, and a weak form like mi, ti, lo/le), which agree in number and gender with their modifiers. These are illustrated in the following examples: (3) Io sono stanca. 'I am tired (fern.).' Tu sei stanco. 'You are tired (masc.).' Noi siamo stanchi 'We are tired (pl. masc. and/or pl. masc. and fem.)' The weak clitic forms (direct and indirect objects) also agree in person, number, and gender with the strong nominal forms to which they refer, as in (4) in which John is the actor and Maria is the receiver of his action. (4) La bacia Maria Giovanni. '(her) is kissing Maria Giovanni.' Given the array of morphological cues that are available in Italian, it is not surprising that Italian adults rely so much on inflectional morphology to interpret sentences in their languages. This reliance on morphology may have led to a correspondingly low reliance on word order cues. The results of two studies comparing sentence processing in English and Italian (Bates, McNew, MacWhinney, Devescovi, and Smith, 1982; MacWhinney et al., 1984) suggest that Italians use semantic cues when morphological markers are not informative enough. For example, in sentence (5) there are no morphological contrasts to indicate agent/ object relations; to interpret this sentence, listeners must choose between the canonical word order in their language (SVO) and semantics (i.e. the fact that apples are inanimate). (5) La mela spinge la bambina. "The apple is pushing the girl.' Under these circumstances, English listeners base their interpretation on word order cues (choosing the apple as the actor), whereas Italians are much more likely to base their interpretations on semantic plausibility. If semantic information is not available, Italians can use word order to make their interpretations; however, their use of word order information interacts with aspects of topicalization and contrastive stress, to an extent that is never seen in English listeners. Within the Competition Model, we should predict an order of acquisition in Italian children that mirrors the dominance of cues in adult sentence comprehension: morphological cues will be acquired first, followed by semantic contrasts, with word order regularities entering the system last (in combination with aspects of topicalization and stress). However, the Competition Model also allows for an

Sentence Comprehension in Italian and Serbo-Croatian

353

interaction between cue validity and cue cost. Several predictions follow from the cue cost factors that we have outlined so far, for Serbo-Croatian and Italian, respectively. First, although Serbo-Croatian inflectional system is relatively complex and ambiguous, it does permit "local" processing and therefore makes relatively low demands on short-term memory. This fact leads to the prediction that SerboCroatian children will master their case system (i.e., rely primarily on case cues in sentence interpretation) well before the point at which they reach adult levels in the use of gender agreement as a cue to sentence meaning (i.e., "distributed" cues). This possibility is addressed in Experiment 1. Second, in line with previous findings by Kail and Charvillat for French, we may expect relatively late acquisition of Italian subject-verb agreement as a cue to sentence meaning in a comprehension task, compared with the acquisition of case morphology in Serbo-Croatian. This possibility is addressed in Experiment 2. Finally, to disentangle the contributions of cue validity and cue cost in the acquisition of Italian subject-verb morphology, Experiment 3 will examine the use of person marking on the verb as a cue to agency in a situation in which the agreement cue can be interpreted and assigned locally.

3. EXPERIMENT 1 3.1. Design and Procedure Subjects in this study were 28 Serbo-Croatian children divided in 2 age groups (3-4, 5-6) and a group of 14 adults. All subjects were presented with a series of sentences in the Serbo-Croatian conversational past tense. Each sentence consisted of two nouns and a transitive action verb in NVN word order. With respect to the case-inflection factor, there were three types of sentence: nominativeaccusative, accusative-nominative, and ambiguous. For case-ambiguous sentences nouns were drawn from a pool of nouns that are not inflected for agent-object relation in Serbo-Croatian. With respect to gender agreement, there were also three types of sentences: items in which the verb agreed with the first noun (AG1), items in which the verb agreed with the second noun (AG2), and items with no gender contrast. For each sentence type, nouns were randomly drawn from a pool of nouns that fit the specifications of that sentence type (e.g., animate-inanimate, case ambiguous, no gender contrast). Verbs were randomly assigned from a list of eight transitive action verbs. This procedure yielded five unique sets of 27 sentences; each subject was randomly assigned one of these five sentence sets. Subjects were tested individually. Toy models of animate and inanimate nouns were placed on a table in front of the subject before presentation of the sentence

354

A. Devescovi et al.

(with position of the objects randomized over trials). On each item, the subject indicated their interpretation by choosing one of the models and moving it against the other. 3.2. Results and Discussion The experimental design allowed a 3 X 3 X 3 X 3 (age X case X gender agreement X animacy) mixed analysis of variance (ANOVA), with age serving as a between-subjects variable and the linguistic variables treated as repeated measures. In addition, three further analyses involving the three within-group factors were carried out, one for each age group. These subanalyses permitted us to determine (a) the age level at which each main effect and interaction became significant, and (b) the proportion of variance accounted for by each factor and by the interactions among factors. We will present the results within each age group first and follow with a report of the full developmental analysis. 3.2.1. ADULTS In the analysis of Serbo-Croatian adults, results largely replicated those reported in other studies (Smith et al., 1984; Smith and Bates, 1987). The ANOVA yielded main effects for case [F(2,22) = 900.94, p < .001], gender [F(2,22) = 112.95, p < 0.001] and animacy [F(2,22 = 5.218, p < 0.014], ordered in accordance with the amount of variance accounted for by each factor. There were also three significant interactions: case by gender [F(4,44) = 72.19, p < .0001]; gender by animacy [F(4,44) = 4.04, p < .01]; and case by gender by animacy [F(8,88) = 5.20, p < .0001]. The case by gender interaction (the largest of these three) is illustrated in Figure 1. The main effects and interactions are all compatible with the competition, coalition, and conspiracy effects predicted by the Competition Model. Main effects follow the order of importance predicted by cue validity: case > gender agreement > animacy. In the absence of competing gender agreement information, adult subjects made their agent-object assignments in accordance with case inflections 100% of time. Even when case is set in competition with gender agreement, subjects responded predominantly in accordance with case inflections (see Fig. 1). For example, for nominative-accusative sentences in which the second noun agreed with the verb (AG2), subjects chose the first noun as agent 91.5% of the time. For accusative-nominative sentences, in which the first noun agreed with the verb (AG1), subjects chose the second noun as agent 94.5% of the time. The significant case X gender interaction reflects the fact that adult subjects based their decisions on gender agreement primarily in that subset of sentences that are ambiguous with respect to case. On such sentences, Serbo-Croatian adults chose the

Sentence Comprehension in Italian and Serbo-Croatian

355

Figure 1. Serbo-Croatian adults Agreement X Case interaction.

first noun as agent 95.5% of the time when it agreed with the verb. When the second noun agreed with the verb, they chose that noun as agent over 83.5% of the time. The slight disparity between these two gender-determined sets reflects a "coalition" between agreement and word order in the first set (corresponding to a small SVO bias for these NVN items) and a "competition" between agreement and word order in the second. The magnitude of the SVO bias itself is best seen in semantically and morphologically ambiguous sentences (CO/AGO), where adult choice of the first noun as agent averages 96%. Finally, animacy had a significant but very small effect on adult subjects' agent—object assignments. Collapsing across all other experimental conditions, they chose the first noun as agent about 60% of the time for animateinanimate sentences (representing a coalition with SVO order), and 46.5% of the time for inanimate-animate sentences (representing a competition with word order). A significant animacy X agreement interaction [F(4.44) = 4.038, p < 0.007] reflects the fact that animacy had its greatest effect in sentences that are genderambiguous. Finally, a significant interaction between case, gender, and agreement was obtained [F(8,88) = 5.202, p < 0.0001], reflecting the fact that animacy is used primarily in sentences that are ambiguous with respect to both case and gender agreement. On these items, adult Serbo-Croatians choose the first noun 100% of the time in morphologically ambiguous AI sentences (reflecting a slight boost

356

A. Devescovi et al.

over the 96% first noun choice observed when semantic cues are ambiguous as well); by contrast, on morphologically ambiguous IA items these subjects choose the second (animate) noun only 25% of the time. This means that SVO word order is more important than semantics as a "back-up" system for Serbo-Croatian adults—in contrast with previous findings for Italian, where semantic cues clearly override word order in morphologically ambiguous sentences. 3.2.2. 3-4-YEAR-OLDS At this age level, there were significant main effects for case [F(2,22) = 28.356, p < 0.0001] and animacy [F(2,22) = 31.918,/? < 0.0001], accounting for 34.9% and 56.5% of experimental variance, respectively. In other words, semantic cues are more important than case information for young Serbo-Croatian children. The main effect of agreement did reach significance [F(2,22) = 5.072, p < 0.015], but it was clearly much smaller than the other two factors in this age range (see Figure 2 for a plot of variance accounted for by each main effect, at each age level). There were no significant interactions among the linguistic factors at this point in development, suggesting that the adult patterns of competition, coalition, and conspiracy have not yet appeared. Although word order was not a factor in this design, a first noun bias among young Serbo-Croatian children can be seen in

Figure 2. Variance resulting from main effects in Serbo-Croatian language.

Sentence Comprehension in Italian and Serbo-Croatian

357

the fact that the first noun was chosen 71% of the time in morphologically and semantically ambiguous sentences. 3.2.3. 5-6-YEAR-OLDS In most respects, performance by 5-6-year-old children was similar to the performance by Serbo-Croatian adults. Among the main effects, case was by far the most significant variable [F(2,22) = 75.316, p < 0.0001], accounting for 82.1% of the experimental variance. Agreement also reached significance [F(2,22) = 10.693, p < 0.001], although it accounted for only 7.8% of the experimental variance. Finally, the main effect of animacy was still significant [F(2,22) = 3.527, p < 0.047], but it accounted for only 5.1% of the experimental variance (compared with 56.5% for 3-4-year-olds). Thus, as Figure 2 shows, an important developmental change has taken place between 3.5 and 5.5 years of age: a switch from primary reliance on semantic cues to the adult pattern, in which case cues are the most important source of information about agent-object relations. There was also a significant interaction between case and gender agreement (F(4,44) = 2.66, p < .05) that did not appear at the earlier age level. However, 5-6-year-old Serbo-Croatian children have not yet reached an adult pattern in their use of agreement morphology. As we noted earlier, adults use gender agreement as their primary source of information when case is ambiguous; they fall back on word-order biases only when case and agreement are both ambiguous. The case-by-gender interaction in 5-6-year-old children takes a somewhat different form (see Figure 3). In particular, these children seem to rely more heavily on word order to mediate interactions between case and agreement. When case and agreement both point to the first noun, they choose the first noun 93% of the time; when case and agreement both point to the second noun, second noun choice averages 76.4%. Thus, in contrast with the adult listeners, these children are attracted to a word order solution roughly 23% of the time even when case and agreement conspire to produce an opposite result. The effect of word order is clearer still on items in which case and agreement compete: on C1/AG2 items, where case aligns with the SVO bias against agreement, first noun choice is roughly 83% (compared with 92% for adults); but on C2/AG1 items, where agreement aligns with the SVO bias against case, first noun choice is 54.2% for children (compared with only 5.5% for adults). The relative weakness of gender agreement (and corresponding preference for word order) is also illustrated in caseambiguous items, where we find a direct competition between gender agreement and SVO on AG2 stimuli: here 5-6-year-old children choose the first noun 79.2% of the time, compared with 16.7% first noun choice by adults. In other words, in a face-off between word order and agreement, agreement "wins" for adults but word order biases "win" for Serbo-Croatian children under 7 years of age. And

358

A. Devescovi et al.

Figure 3. Serbo-Croatian 5-6-year-olds' Agreement X Case interaction.

yet we cannot conclude that word order is stronger overall in these children: on morphologically and semantically ambiguous items, adults show a 96% bias toward SVO, compared with only 87.5% in 5-6-year-old children. The best interpretation seems to be that children in this age group do not "trust" agreement cues; they prefer to use other sources of information, particularly case and word order. From this point of view, the order of acquisition of grammatical cues in Serbo-Croatian deviates from the cue validity ordering obtained with SerboCroatian adults. 3.2.4. CROSS-AGE ANALYSIS These interpretations are confirmed by the overall ANOVA, which yields several robust interactions between grammatical cues and age. In addition to the expected main effects (all significant at p < .0001, in the directions already described), there was a significant interaction between age and case [F(4,66) = 19.19, p < .0001], reflecting an increase in the importance of case information from 3-4 years of age. There was also a significant interaction between age and gender agreement [F(4,66) = 8.29, p < .0001], with agreement increasing steadily in importance from 3 years of age to adulthood. The age-by-animacy interaction [F(4,66) = 13.03, p < .0001] reflects a precipitous drop in use of animacy cues between 3 and 6 years of age (see Figure 2). Finally, there was a significant and very robust three-way interaction of age, case, and agreement [F(8,132 = 14.756, p < 0.0001]. When case is ambiguous, adults rely primarily

Sentence Comprehension in Italian and Serbo-Croatian

359

on agreement, choosing the noun that agrees with the verb as the object regardless of other cues (i.e., sentence position and/or animacy). In the same situation, younger children behave at close to random levels, whereas 5-6-year-olds tend to choose the first noun as the agent (regardless of semantics and/or agreement marking). In other words, a word order strategy of some sort seems to have emerged among these Serbo-Croatian children, corresponding to a preference for the standard SVO order used throughout this experiment. The old semantically based strategy has been supplanted, use of case morphology is now well established, and yet these children still appear to be unable or unwilling to use "long-distance" agreement cues to resolve case-ambiguous sentences.

4. EXPERIMENT! 4.1. Design and Procedure The subjects in this study were 50 middle-class Italian children (25 boys, 25 girls), divided into 5 age groups (with mean ages of 3.6,4.6, 5.6, 7.6, and 9.6). There was also a control group of 10 adults, mainly university students. All subjects were presented with 54 sentences. Each sentence consisted of two nouns preceded by a definite article, and a transitive action verb conjugated in the third-person present indicative (as illustrated in Table 1). The sentences were made up according to an orthogonal combination of three variables: word order (NVN;VNN;NNV), animacy (both nouns animate: AA; first noun only animate: AI; second noun only animate: IA), and agreement (verb in agreement with the first noun only: Agl; verb in agreement with the second noun only: Ag2; ambiguous agreement: AgO). The verb was singular in 50% of the items, plural in the remaining 50%. The orthogonal combinations in this experimental design (3 X 3 X 3) yielded 27 different types of sentences, with two sentences within each cell. Three different sets of sentences were prepared. All nouns and verbs were randomly assigned to the aforementioned conditions. Each subject was randomly presented with one of these three randomized sets. The experimental procedure was the same that has been used for Serbo-Croatian. 4.2. Results and Discussion The experimental design allowed a 6 X 3 X 3 X 3 (age X word order X animacy X agreement) ANOVA: age was treated as a between-group variable, and all other variables were treated as repeated measures. In addition, separate 3 X 3 X 3 analyses were carried out within each age level. We will present the

360

A. Devescovi et al.

within-age analyses first, followed by a discussion of interactions with age obtained in the full design. 4.2.1. ADULTS Results for adult subjects largely replicate those reported in MacWhinney et al., (1984). All three main effects reached significance, including a large and significant effect of the subject-verb agreement [F(2,18) = 227.72, p < 0.0001], and smaller effects for word order [F(2,18 = 8.27, p < 0.0003], and animacy [F(2,18) = 19.54, p < 0.0001]. All of these main effects are in the expected direction. Italian adults systematically chose the noun that agrees with the verb as the agent, regardless of its position in the sentence. The significant effect of word order reflects an SVO interpretation of sentences with NVN structure, with random performance on the other two word orders. Finally, the animacy effect reflects a preference for the "animate" noun as the agent. Our conclusion that agreement is by far the most informative cue for Italian adults is confirmed by the fact that this main effect accounts for 83.46% of the experimental variance (compared with 1.67% accounted for by word order and 6.48% accounted for by animacy; Fig. 4). Although animacy is less important than agreement, it seems to be the most important source of information when agreement is ambiguous, as demonstrated by the significant interaction between agree-

Figure 4. Variance resulting from main effects in Italian language.

Sentence Comprehension in Italian and Serbo-Croatian

Figure 5.

361

Italian adults' Agreement X Animacy interaction.

ment and animacy illustrated in Figure 5 [F(4,36) = 11.25, p < 0.0001]. None of the other interactions reached significance in this adult sample. 4.2.2. 3-YEAR-OLDS

There were two significant main effects for children at this age: agreement [F(2,18) = 4.23, p < 0.03] and animacy [F(2,18) = 39.75, p< 0.0001] (Fig. 3). Three-year-olds systematically chose the animate noun as the agent, regardless of its position in the sentence (Fig. 4). The small agreement effect that we see at this age comes primarily from AG1 items (which represent a small "conspiracy" between agreement and canonical word order). On those items, 3-year-olds chose the first noun 60.5% of the time, compared with random performance on morphologically ambiguous items (AgO) and on items in which the second noun agrees with the verb (AG2). In other words, the relative importance of agreement and animacy are completely reversed between 3 years of age and adulthood. Animacy accounts for 87% of experimental variance at age 3, whereas agreement accounts for only 3.78% (Figure 4). In some respects, this pattern resembles the one we obtained for Serbo-Croatian children at 3-4 years of age: semantic information is most important, followed by case morphology. There is, however, a significant quantitative difference that

362

A. Devescovi et al.

highlights the different importance that each linguistic group gives to their "best" morphological cue: for Serbo-Croatian children, animacy accounted for only 56.5% of experimental variance (compared with 87% in Italian); moreover, the magnitude of the case effect in Serbo-Croatian (accounting for 34.7% of the variance) is considerably stronger than the subject-verb agreement effect in Italian (3.78%). 4.2.3. 4-YEAR-OLDS At this age all main effects reached significance: animacy [F(2,18) = 54.68, p < 0.0001], agreement [F(2,18) = 6.50, p < 0.007], and word order [F(2,18) = 34.57, p < 0.0001]. Animacy still accounts for most of the experimental variance (58.94%), compared with 23.88% for word order and only 3.23% for agreement (Figure 4). There was also a significant interaction between animacy and word order [F(4.36) = 5.80, p < 0.001]. Note that there has been no increase between 3 and 4 years of age in the amount of variance accounted for by agreement. Animacy is still the strongest determinant of sentence interpretation. Replicating the trend observed in 3-year-olds, the small agreement effect comes primarily from AG1 items in which the first noun agrees with the verb (i.e., agreement buttressed by word order). The major change between 3 and 4 years of age revolves around the emergence of a strategy for choosing the first noun as agent, particularly for NVN items, confirming the results of previous studies on the development of sentence comprehension in Italian (Bates et al., 1984; Caselli and Devescovi, 1982). Interestingly, this is also close to the point at which a word-order strategy emerged in Serbo-Croatian (Experiment 1). Overall, the Italian word-order strategy is secondary compared to use of animacy as a cue to agent-object relations. This is clear from a comparison of variance accounted for, and from the animacy-by-word order interaction illustrated in Figure 6. However, the competition observed between NVN and animacy at this age is very strong. In the two noncanonical orders, the animate noun is systematically chosen as the agent, whereas in the NVN standard order the SVO interpretation prevails. Within NVN sentences, the animate noun is chosen as the agent more than 90% of the time when it precedes the verb (AI); the animate noun is chosen only 35% of the time when it follows the verb (IA). Hence, even though animacy is a stronger cue to meaning overall, SVO wins over animacy in a direct competition. 4.2.4. 5-YEAR-OLDS

At this age only the main effects are significant, animacy [F(2,18) = 29.60, p < 0.0001], word order [F(2,18) = 13.59, p < 0.0001], agreement [F(2,18) =

Sentence Comprehension in Italian and Serbo-Croatian

Figure 6.

363

Age X Word Order interaction in Italian language.

13.50, p < 0.0001]. The interaction between semantics and word order just misses significance [F(4,36) = 2.53, p < 0.057]. Animacy is still the strongest cue overall, accounting for 49.52% of the experimental variance. The agreement effect has grown in magnitude, now accounting for 20.80% of the experimental variance, higher than the 14.57% associated with word order (Fig. 4). Although none of the interactions reach significance at this age, there are still some discernible changes in the role played by word order. Differences among the three word orders have flattened at this age, because—in contrast with the two younger groups 5-year-olds have generalized the first noun strategy to VNN and NNV orders, which are interpreted as VSO and SOV, respectively (Fig. 6). The animacy-by-word order interaction just misses significance [F(94,36) = 2.54, p < .06]. When there is a direct competition between SVO order and animacy (on NVN/IA items), 5-year-olds choose the first noun 70% of the time; on VNN and NNV items, the conflict between animacy and the first noun strategy results in random behavior. We will discuss this resurgence of word order among 4-5-yearold children in more detail later on. 4.2.5. 7-YEAR-OLDS

Once again, all three main effects reached significance: agreement [F(2,18) = 27.49, p < 0.0001], word order [F(2,18) = 8.614, p < 0.002], and animacy

364

A. Devescovi et al.

[F(2,18) = 39.78, p < 0.037]. In addition, there was a significant interaction between animacy and agreement [F(4,36) = 2.98, p < .04]; the interaction between order and animacy just missed significance [F(4,36) = 2.18, p < .10]. The most important change at this age is reflected in the relative size of the agreement effect, which now accounts for 74.6% of the experimental variance (compared with 9.3% for word order and 8% for animacy). At last children have begun to display the pattern of morphological dominance that characterizes adult sentence interpretation in this language. The tendency to overgeneralize the first noun strategy that we saw among 4-5-year-olds has also begun to abate. However, the interactions among cues still look qualitatively different in 7-year-olds compared with adults. On AVI/AG2 items, where agreement must compete with a combination of animacy and SVO biases, performance by 7-year-old children is essentially random (see Figure 7). By contrast, agreement wins 70% of the time for adults in the same situation. Hence agreement is still not as powerful for children as it is for adults. 4.2.6. 9-YEAR-OLDS At this age, all main effects and interactions reached significance at p < .01, with one exception: the interaction between animacy and word order just missed significance [F(4,36) = 2.46, p < .07]. Agreement is by far the strongest vari-

Figure 7. Italian 7-year-olds' Animacy X Agreement interaction.

Sentence Comprehension in Italian and Serbo-Croatian

365

able, with a main effect accounting for 87.6% of the variance [F(2,18) = 293.4, p < .0001]. Although the animacy and word order effects do reach significance (for animacy, F(2,18) = 22.62, p < .0001; for word order, F(2,18) = 8.94, p < .002), each accounts for less than 3% of the experimental variance. In general, then, it seems that 9-year-olds are close to the mature pattern of sentence interpretation in their language. The various interactions all reflect the fact that animacy and word-order effects are strongest in morphologically ambiguous items. This is particularly clear in the significance three-way interaction of agreement, animacy, and word order [F(8,72) = 3.13, p < .004]. When SVO word order and animacy "gang up" against agreement (in NVN/AI/AG2 items), morphology wins the competition 80% of the time. The only detectable difference between 9-year-olds and adults revolves around the relative strength of word order versus animacy on sentences without a subject-verb agreement contrast. On these items, animacy clearly "wins" on NNV and VNN orders, but animacy "loses" on NVN/IA sentences 85% of the time (compared with a 50/50 standoff for the same items in our adult controls). 4.2.7. CROSS-AGE ANALYSIS The developmental trends that are implicit in our analysis so far emerge clearly in an ANOVA including age as a factor (see Figure 4). First, there is a sharp drop in the importance of animacy between 5 years and adulthood, reflected in a significant age-by-animacy interaction [F(10,108) = 7.71, p < .0001]. Second, a sharp rise in the impact of agreement cues creates a significant age-by-agreement interaction [F(10,108) = 31.96, p < .0001]. There is also a significant interaction between age and word order [F(10,108) = 3.76, p < .0001], corresponding to the rise and subsequent fall in the importance of word order-cues illustrated in Figure 4. Higher order interactions include a three-way relationship between age, agreement, and animacy [F(20,216) = 2.78, p < .0001], a three-way relationship of age, agreement, and word order [F(20,216) = 1.74, p < .03], and a complete four-way interaction of age by animacy by agreement by word order [F(40,4320 = 1.58, p < .02]. These complex interactions can be summarized briefly as follows: 1. Animacy is the dominant cue to meaning for Italian children up until age 7. 2. The development of word order follows an inverted U-shaped patter, with a surge in the importance of word-order information between 4 and 9 years of age, reaching its peak at 5 years of age when children tend to overgeneralize the first noun strategy to all three word order types. 3. Agreement does not become the most important cue to meaning until age 7, and it does not reach the levels displayed by adults until age 9.

366

A. Devescovi et al.

As we predicted, Italian children succeed in mastering the "best" morphological cue in their language several years beyond the point at which use of case morphology reaches adult levels in Serbo-Croatian children. Furthermore, a comparison of Experiments 1 and 2 suggests that agreement cues are not fully mastered in either language until some point between 6 years and adulthood. This finding accords with Kail's report on the late acquisition of agreement morphology in French sentence comprehension (Kail, 1989) and may therefore reflect a very general fact about the use of agreement information in receptive language processing. In particular, these findings appear to support the view that the costs involved in processing distributed morphological cues mitigates against their use until some (relatively constant) point in development—perhaps a point at which changes in working memory make it possible for children to use such information efficiently. There are some alternative explanations for these delays in the use of agreement morphology. For example, one might argue that the Italian system of verb morphology is so complex and so difficult that it is difficult to acquire in the absolute—for use in production, comprehension, or any other performance domain. Evidence against this interpretation comes from developmental studies of spontaneous speech (Bates, 1986; Pizzuto and Caselli, 1989), which clearly show that Italian children have attained productive control over most aspects of noun and verb morphology by 3 years of age. To the extent that this is true, it suggests an excessive gap between development of comprehension and development of production. To explore this issue further, we have carried out a third study focused on Italian children's ability to decode a different aspect of agreement morphology, a system of person marking which (in many cases) permits the listener to assign semantic-syntactic roles immediately (i.e., as soon as the verb is processed). If the assignability hypothesis is correct, then the use of "local" agreement cues in sentence interpretation should pose no problem for Italian children.

5. EXPERIMENT 3 5.1. Design and Procedure Thirty upper middle-class children (15 boys, 15 girls) took part in the experiment, with 10 subjects in each of three age groups (2, 3, and 5 years of age). All subjects were presented with six verbs normally used in everyday conversations with very young children (e.g., drink, eat, sleep). The verbs were conjugated in the first three persons of the present tense, yielding a total of 18 singular verbal forms. In some verb paradigms (i.e., verbs that end in "-are" in the infinitive), the third person of the present tense is homophonous with the second person singu-

Sentence Comprehension in Italian and Serbo-Croatian

367

lar of the imperative; in other verb paradigms (i.e., verbs whose infinitive forms end in "-ere" or in "-ire"), the distinction between third-person indicative and second-person imperative is unambiguously marked. For example, the verb mangiare (to eat) takes the form mangia in the third-person singular indicative, mangi in the second-person indicative, and mangia in the second-person imperative. By contrast, the verb bere (to drink) takes the form beve in the third-person singular, bevi in second-person indicative, and bevi in the second-person imperative. This means that children could confuse the second- and third-person interpretations for -are verbs but not for -ire or -ere verbs. In this experiment, we used three verbs from each group (i.e., three potentially ambiguous -are verbs and three unambiguous -ire/-ere verbs), treating verb type as a factor in the design. Hence the full design of the experiment is a 3 (age) by 3 (verb person) by 2 (verb type) design, with age treated as a between-subjects factor and verb person and verb type treated as repeated measures. Each child was tested in a pretend play situation involving the child himself, the researcher, and a large teddy bear. Actions from everyday "scripts" were acted out: getting up, going to sleep, eating, taking a walk, an so on. On each test trial, the child was asked to act out a single verbal form ("Now show me that I am/you are/he is eating"), treating one of the three participants as the actor who must operate on the toy objects (cups, blankets, etc.). On all these trials, the subject of the verb form was omitted (yielding structures that are well formed, frequent, and perfectly natural in Italian). Hence a single item might be "Now show me MANGIO," with the verb form indicating the Experimenter (the first-person speaker) ought to carry out the act. All three forms of the same verb were always presented in a block (i.e., all three forms of the verb mangiare), but the order of presentation of the three verbal persons was randomly varied within each block. Hence the child could only rely on verbal morphology in order to understand which of the three participants was the subject of the action he was required to act out. Prior to administration of the test trials, there was a brief practice period to familiarize the child with the idea of choosing among the three potential actors (experimenter, child, or bear). A toy telephone was placed on the floor equidistant from all three participants, and the Experimenter modeled "telephoning," placing the receiver against her own ear (saying "Now I call"), the child's ear (saying "Now you call"), and the bear's ear (saying "Now he/she calls"). Children in all three age groups appeared to understand the game immediately. 5.2. Results and Discussion Four ANOVA were carried out: (a) a 3 X 3 X 2 analysis across age groups and (b) individual 3 (verb person) by 2 (verb type) analysis within each age level. We will begin by presenting results for individual age groups and then discuss the full analysis.

368

A. Devescovi et al.

At 2 years of age, both main effects reached significance: person [F(2,18) = 9.079, p < 0.002] and type of verb [F(l,9) = 6.06, p < 0.036]. There was also a significant interaction between person and verb type at this age [F(2,18) = 6.726, p < 0.007]. As Figure 8 shows, the highest rate of correct answers occurred on the first- and second-person singular, this last one being slightly higher. The greatest number of mistakes involved the third-person singular. From one point of view, this result is surprising, because Italian children tend to produce the third-person singular as the "unmarked" form of many verbs during the earliest stages of grammatical development (Bates, 1976; Pizzuto and Caselli, 1989). Superior performance on the first and second person in this comprehension task could reflect the fact that first and second person use are typical in parent-child dialogue (Bruner 1975); this is particularly true at this stage of development, when narrative descriptions of events involving third parties are relatively rare (Karmiloff Smith, 1985). However, an examination of the person-by-verb type interaction suggests a different interpretation. As we mentioned earlier, -are verbs present the possibility of a confusion between the third-person indicative and the second-person imperative. In fact, as illustrated in Figure 9, mistakes on the third person occurred primarily on these potentially ambiguous -are items (i.e., a 65% error rate). Compare this with a 15% error rate on the unambiguous -ire/-ere verbs, where the second-person form is the same for indicatives and imperatives; on these items, third-person performance is almost as good as performance on the other two conjugations.

Figure 8. Italian 2-year-olds' percentage correct performance.

Sentence Comprehension in Italian and Serbo-Croatian

369

Figure 9. Italian 2-year-olds' Person X Type of verb interaction.

In short, these results suggest that 2-year-olds are able to use person marking to assign the agent role. The errors that we see on the third person reflect a potential ambiguity that is restricted exclusively to -are verbs. On these items, 2-year-olds seem to prefer a second-person imperative reading (e.g., carrying out the action themselves when the Experimenter says "Show me MANGIA"). Although such interpretations do violate the descriptive-indicative nature of the "game," they are not (strictly speaking) errors of person assignment. At 3 and 5 years of age, none of the within-subject effects reached significance. From 3 years of age onward, children respond correctly more than 90% of the time, on all three persons and on both verb types. Hence the imperative-indicative confusion we saw at 2 years of age for potentially ambiguous -are verbs has disappeared entirely by age 3. Children now understand this descriptive-indicative game very well, and they use person marking correctly and efficiently as a cue to sentence meaning. In the overall analysis, there was a significant effect of age on number correct, reflecting the fact that the person-marking game is fully mastered between 2 and 3 years of age [F(2.27) = 36.979, p < 0.0001]. It follows that younger children in the previous experiment were already able to decode verbal person morphology. One might argue that the singular-plural distinction manipulated in Experiment 2 develops much later. However, longitudinal studies of Italian free speech suggest that number marking is also acquired before 3 years of age in most cases (Bates,

370

A. Devescovi et al.

1976; Pizzuto and Caselli, 1989). We therefore feel justified in concluding that the late adoption of agreement cues in sentence interpretation in Experiment 2 (and perhaps in Experiment 1 as well) reflect the fact that children do not want to use their knowledge of agreement marking because of the processing costs involved in holding several items in memory and comparing them across the sentence before an interpretation can be assigned.

6. GENERAL DISCUSSION To summarize, we have shown that Serbo-Croatian children make extensive use of case information as a cue to sentence meaning by 3 years of age, although they do not reach adult levels until age 5. Compared with clearly marked case languages like Hungarian and Turkish, the irregular Serbo-Croatian system apparently does take longer to acquire. In this respect, we have replicated previous findings by Slobin and Bever (1982). In addition, however, we have shown that 5-6-year-old Serbo-Croatian children still fail to show the adult pattern in secondary use of gender agreement; the gender agreement effect is significant, but very small and clearly subordinate to their use of animacy (at 3-4 years) and/or word order (at 5-6 years). The agreement effect in Italian also reached significance in our 3-year-old group, and remained significant at every subsequent age level. However, these agreement effects were very small compared with the near-deterministic use of agreement by Italian adults. Agreement did not become the dominant cue to meaning until 7 years of age, and the adult pattern is not established until age 9. For the youngest children in both linguistic groups, animacy is the strongest source of information. At 5 years of age, in both languages (albeit to a different degree), word-order information begins to play a prominent role in sentence processing—although it never reaches the dominant levels that characterize performance by English-speaking children and adults (Bates et al., 1984). Attention to word order may be viewed as a holding action, recruitment of an additional source of information to supplement the relatively weak use of agreement cues in both these language groups. This may be less surprising in Serbo-Croatian (where agreement is always a "backup" to case), but it is quite surprising in Italian (where agreement is clearly the "best" source of information for adult listeners). From one point of view, these results support the cue-validity predictions of the Bates and MacWhinney Competition Model: specifically, children in both linguistic groups showed very early sensitivity to the dominant cues of their mother tongue. However, the relative strength of cues to meaning from 3 years to adulthood deviates from predictions based on cue validity. In the Competition Model, failures of cue validity must be explained by cue cost. The results of Experi-

Sentence Comprehension in Italian and Serbo-Croatian

371

ment 3 are particularly important in this regard. If the situation permits "local" use of agreement cues (marked on the verb in a one-word sentence), performance by 2-3-year-old Italian children is similar to the performance displayed by 2-3-year-old speakers of clear case-marked languages like Hungarian and Turkish. This finding is compatible with the notion of assignability: those morphological cues that can be interpreted through local processing are more easily learned than those that require all elements be processed and kept in memory. Because the Serbo-Croatian case system is relatively low in availability and reliability, we expected later establishment of a case-based strategy in this language (compared with previous work on Turkish and Hungarian). We would also expect gender agreement to come in later than case within this language. It is interesting, however, that there are commonalities across languages in the point at which agreement morphology reaches adult levels of use in sentence interpretation. In Serbo-Croatian, agreement was still subordinate to word order at 6 years of age. In Italian, agreement did not emerge as the strongest cue until 7 years of age. And in the earlier study by Kail and Charvillat (1988), French children do not rely on the information carried by agreement between object nouns and object clitics until some point after 6 years of age. Italian children's difficulty in relying on agreement when interpreting a simple sentence is also confirmed by another study in our laboratory (Devescovi and D'Amico, 1986), an Italian version of the experiment reported by Kail and Charvillat. The variables manipulated in this study included word order, clitic pronoun-object agreement, and subject-object agreement. All sentences were semantically reversible. In this experiment, where semantic contrasts are not available, 4 -5-year-old Italian children relied on word order as their primary cue to sentence meaning. Systematic use of subject-verb agreement and object-clitic agreement did not begin until 6-7 years of age. Again, there were small but significant effects of morphology at all the earlier ages. Hence the fact that younger children fail to use agreement information does not mean that they do not know the rules that govern both these forms of agreement. However (as we saw among the 3-5-year-old children in Experiments 1 and 2), agreement effects tend to appear primarily when there is a "coalition" between agreement and some other cue (especially word order). They can analyze single segments of the sentence and assign agreement correctly at a local level, but they apparently cannot or do not want to hold separate elements in memory and compare them across the course of the sentence. As a result, agreement is used only to strengthen a choice that has already been made on the basis of other cues (e.g., word order). The hypothesis that very young children prefer local processing strategies could also account for the early use of animacy to support or substitute for morphosyntactic strategies. In fact, younger children might use the animacy contrast as a local cue in this experimental situation; because the animals and objects to be used on each trial are laid out directly in front of the child, all he or she really has to do

372

A. Devescovi et al.

is to listen for the verb and assign roles immediately at that point in the sentence. There is no need to keep in memory those elements that have already been processed. This brings us to a rather different matter: the appearance of a word-order-based strategy in 4-5-year-old Italian children. A preference for interpreting NVN sentences as SVO has also been observed in other languages: Bever (1970), in a study on English children's comprehension of the passive, Sinclair and Bronkart (1972) in a study of sentence comprehension in French, Kail and Charvillat (1988) in a study of Spanish. Our study of Serbo-Croatian also supports these observations; although word order was not directly manipulated in Experiment 1, children showed a bias toward choice of the first noun as agent that reached its peak in the 5-6-year age range. Finally, it is worth pointing out that Turkish children in the Slobin and Bever (1982) study began to show a very small but significant preference for SOV around 4-5 years of age. One explanation for this phenomenon might revolve around the formation of a word-order "prototype" or canonical sentence frame, a development that might depend on the number of canonical strings to which children have been exposed (Slobin, 1981). However, this proposal fails on at least two counts. First, it does not explain why 5-year-old Italian children overgeneralize their strategy to NNV and VNN sentences (something that does not occur in English despite the strength and prevalence of SVO). Second, it is difficult to explain why this prototype should emerge at the same time in a wide variety of languages, when the absolute number of canonical sentence forms in the input to children varies widely from one language to another. There is another possibility, which is also related to the late emergence of agreement morphology in sentence comprehension. At one level of analysis, word order and agreement have something in common: both sources of information require the child to hold several elements in memory and compare them in some fashion. To perceive a "prototypical" SVO or SOV, a child must simultaneously perceive or remember three different constituents and the relationships among them. To use long-distance agreement cues, the child must also perceive and remember several different elements, and compute the agreement relationships among them. It is possible that phrase structure types like SVO or SOV do not behave like a complete unit until 4-5 years of age, because of certain poorly understood constraints on memory that preclude such long-distance comparisons. Before that point, children may use word-order information in a local, fragmentary way, for example, tending to pick the noun before the verb as subject (S-V) or the noun after the verb as object (V-O). This hypothesis rests on the assumption that there is some kind of exogenous change in short-term or working memory around 4-5 years of age, a change that is relatively independent of the child's current language level and/or the nature of his or her input language. Evidence for a shift of this kind can be found in a number of nonlinguistic tasks (e.g., a series of visual-manual coordination tasks

Sentence Comprehension in Italian and Serbo-Croatian

373

studied by Pierre Mounod and his colleagues; Mounoud, 1983; Mounoud, Vivani, Houert, and Guyon, 1985). These studies show that around 6 years of age children change from a "local" control of experimental elements to a form of control that takes the whole array into account. We may speculate that the appearance of such an "array processor" also permits the human child to represent and operate on a multielement array of linguistic forms, scanning and comparing those elements to extract relationships that are not evident when the same elements are considered one at a time. In linguistic development, the use of a word order-based strategy could mark this change from a local processing of linguistic input to an overall processing that seems to be well established by 7 years of age. The very same factors that permit the child to contemplate an array of separate content words might also lead to flexible and efficient comparison of the agreement markers located on those content words. Hence we could explain the inverted-U-shaped surge of word-order strategies in inflected languages: the new ability has its first impact on the relationship between content words, followed closely by mastery of distributed agreement cues; in turn, mastery of agreement cues in real-time sentence comprehension feeds back onto these temporary word-order strategies, reducing them to levels that are appropriate for that language. The new processing abilities that are available after age 5 might have implications for other areas of language processing as well. For example, many studies have indicated that something new happens after 7 years of age, when children are at last able to coordinate linguistic features that go beyond the single sentence. This is, for example, the age at which children consistently use text cohesion markers in their "narrative" productions (Karmiloff-Smith, 1985; Orsolini, 1990; Orsolini and Devescovi, 1987,1988). Of course, this last shift could be (at least in part) a result of school experience, which brings with it a strong emphasis on narrative skills. But we still need to account for the fact that middle-class children who are exposed to stories and other narrative forms from their preschool years (if not before) nevertheless show a marked increase in the ability to process and produce narrative discourse between 5 and 7 years of age. If these coincidences do not happen by chance, we may hypothesize that a new working memory configuration facilitates those aspects of language comprehension and production processes that emphasize long-distance relations. It also makes it possible to solve other cognitive problems that make similar demands on the child's ability to recall and compare several disparate elements in one complex operation. To be sure, these new abilities interact with structural and statistical differences between natural languages. And many aspects (perhaps most aspects) of grammatical learning take place long before this purported change in processing capacity. It is clear that children below the age of 3 are able to pick up regularities in the relationship between form and meaning, and regularities in the relationship between two or more linguistic forms. However, we suggest that they

374

A. Devescovi et al.

do so in a gradual and perhaps piecemeal fashion, supported by their knowledge of the context and by global cues (e.g., prosody—Morgan, 1986) that hold large units together for them long enough for a certain amount of learning to take place. When should we conclude that children "know" the rules that govern their language? Many investigators assume that aspects of grammar are acquired as soon as they are systematically applied in at least one well-supported situation of language use (e.g., Fodor and Crain, in press). One such piece of evidence is sufficient to establish the existence of "competence"; further delays in the child's application of that knowledge are explained (or left unexplained) by invoking some aspect of "performance". Delays in the use of agreement markers in sentence comprehension would constitute a particularly clear example of such a competence-performance lag. But there is an alternative view: language acquisition may be a protracted process, a constant interplay between knowledge (which may be partial) and use (which may be limited and inconsistent for many years). Within this framework, the first systematic use of a linguistic form has no privileged status; each step toward mature use of a natural language is equally important, and the process is not over until it is over (see Marchman, Bates, Burkhardt, and Good, 1991, for a discussion of how the passive develops from 2 through 20 years of age). In the words of the Italian playwright Edoardo de Filippo: Gli esami non finiscono mai. 'The testing never ends.'

REFERENCES Ammon, M. S., and Slobin, D. I., (1979). A cross linguistic study of the processing of causative sentences. Cognition, 7, 341-354. Bates, E. (1976). Language and context: the acquisition of pragmatics. New York: Academic Press. Bates, E., and MacWhinney, B. (1982). Functional approaches to grammar. In E. Wanner and L. Gleitman (Eds.), Language Acquisition: The state of the art. New York: Cambridge University Press. Bates, E., and MacWhinney, B. (1987). Competition, variation and language learning. B. MacWhinney (Ed.), Mechanisms of language acquisition (pp. 157-193). Hillsdale, NJ: Lawrence Erlbaum. Bates, E., and MacWhinney, B. (1989). Functionalism and the Competition Model. In B. MacWhinney and E. Bates (Eds.), The crosslinguistic study of sentence processing. New York: Cambridge University Press. Bates, E., MacWhinney, B., Caselli, C, Devescovi, A., Natale, F, and Venza, V. (1984). A cross-linguistic study of the development of sentence interpretation strategies. Child Development, 55, 341-354. Bates, E., McNew, S., MacWhinney, B., Devescovi, A., and Smith, S. (1982). Functional constraints on sentence processing: A cross-linguistic study. Cognition, 11,245-299.

Sentence Comprehension in Italian and Serbo-Croatian

375

Bever, T. G. (1970). The cognitive basis for linguistic structures. In J. R. Hayes (Ed.) Cognition and the development of language, New York: Wiley. Bickerton, D. (1984). The language bioprogram hypothesis. The Behavioral and Brain Sciences, 7 :2, 173-187. Bowerman, M. (1973). Early syntactic development: a cross-linguistic study with special reference to Finnish. Cambridge: Cambridge University Press. Braine, M. D. S. (1976). Children's first word combinations. Monographs of the Society for Research in Child Development, 41:1, Serial Number 164. Brown, R. (1973). A first language, Vol. 1: The early stages. Cambridge MA: Harvard University Press. Bruner, J. (1975). From communication to language. A psychological perspective. Cognition, 3, 255-287. Caselli, C., and Devescovi, A. (1982). Sentence comprehension in Italian children from 2 to 5 years of age. Italian Linguistics, 2, VII, 1-7. Devescovi, A., and D'Amico, S. (1986). Clitic pronoun comprehension in Italian children from 4 to 6 years of age. Paper presented to 2nd European Conference on Developmental Psychology—Development: European Perspectives. Rome, Italy, September 10-13. Elman, J. (1989). Representation and structure in connectionist models. Center for Research in Language Technical Report #8903, University of California, San Diego. Fodor, J., and Crain, S. (in press). On the form of innate linguistic knowledge. Cambridge, MA: MIT/Bradford Books. Frankel, D. G., and Arbel, T. (1981). Developmental changes in assigning agent relations in Hebrew: The interaction between word-order and structural cues. Journal of Experimental Child Psychology, 32, 102-114. Hakuta, K. (1982). Interaction between particles and word-order in the comprehension of simple sentence in Japanese children. Developmental Psychology, 18, 62-76. Jacobs, W. J., and Nadel, L. (1985). Stress-induced recovery of fears and phobias. Psychological Review, 92, 512-531. Johnston, J., and Slobin, D. I. (1979). The development of locative expression in English, Italian, Serbo-Croatian, and Turkish. Journal of Child Language, 6, 529-545. Just, M., and Carpenter, P. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 187, 329-354. Kail, M. (1989). Aspects of sentence comprehension in French and Spanish. In B. MacWhinney & E. Bates (Eds.), The cross-linguistic study of sentence processing. New York: Cambridge University Press. Kail, M., and Charvillat, A. (1988). Local and topological processing in sentence comprehension by French and Spanish children. Journal of Child Language, 15, 637-662. Karmiloff-Smith, A. (1985). Language and cognitive processes from a developmental perspective. Language and Cognitive Processes, 1, 60-85. MacWhinney, B., and Bates, E., (Eds). (1989). The crosslinguistic study of sentence processing. New York: Cambridge University Press. MacWhinney, B., Bates, E., and Kliegl, R. (1984). Cue validity and sentence interpretation in English, German, and Italian. Journal of Verbal Learning and Verbal Behavior, 23, 127-150. MacWhinney, B., Pleh, C., and Bates, E. (1985). The development of sentence comprehension in Hungarian. Cognitive Psychology, 17, 178-209.

376

A. Devescovi et al.

Marchman, V., Bates, E., Burkhardt, A., and Good, A. (1991). Functional constraints on the acquisition of the passive: towards a model of competence to perform. First Language, 11, 65-84. McClelland, J. L., and Rumelhart, D. E. (1981). An interactive activation model of context effects in letter perception: Part I, An account of basic findings. Psychological Review, 88, 378-408. McNeill, D. (1975). Semiotic extension. In R. Solso (Ed.), Information processing and cognition: The Loyola Symposium. Hillsdale, NJ: Erlbaum. McNeill, D. (1979). The conceptual basis of language. Hillsdale, NJ: Laurence Erlbaum. Morgan, J. (1986). From simple input to complex grammar. Cambridge, MA: MIT Press. Mounoud, P. (1983). Lo sviluppo come costruzione di rappresentazioni. G. Bocchi, M. Ceruti, D. Fabbri-Monteseno, A. Munari (Eds.), L'altro Piaget (pp. 55-69). Milano: Emnie Edizioni. Mounoud, P., Viviani, P., Hauert, C. A., and Guyon, J. (1985). Development of visuomanual tracking in 5 to 9 year old boys. Journal of Experimental Child Psychology, 115-132. Orsolini, M. (1990). Episodic structure in children's fantasy narratives: "Breakthrough" to decontextualized discourse. Language and Cognitive Processes, 5 (1) 53-79. Orsolini, M., and Devescovi, D. (1987). Structure, coherence and cohesion in children's narratives. Paper presented to the 4th. International Congress for the Study in Child Language/Lund, Sweden, 19th-24th July. Orsolini, M., and Devescovi, D. (1988). Toward a planned discourse—Episodic structure, thematic continuity, pauses and intonation in orally produced stories. Paper presented to the Symposium on Early Text Construction in Children, Rome (Italy) October 26th-27th. Osgood, C. E., and Tanz, C. (1977). Will the real object in bitransitive sentence please stand up? A. Julliard (Ed.) Linguistic studies offered to Joseph Greenberg on the occasion of his sixtieth birthday. Saratoga, CA, Anme Libri. Pinker, S. (1982). A theory of the acquisition of lexical-interpretive grammars. In J. Bresnan (Ed.), The mental representation of grammatical relations. Cambridge, MA: MIT Press. Pizzuto, E., and Caselli, M. C. (1989). Deixis and anaphora in the acquisition of Italian: The interplay between discourse and morphology. Paper presented to Eleventh Annual Boston University Conference on Language Development, Boston, MA, October. Pizzuto, E., and Caselli, M. (in press). L'acquisizione della morfologia nello sviluppo del linguaggio spontaneo. Acta Phoniatrica Latina. Rumelhart, D. E., and McClelland, J. L. (1986). Parallel distributed processing: Explorations in the microstructure of cognition, Vol. 1. Cambridge, MA: Bradford/MIT Press. Rumelhart, D. E., and McClelland, J. L. (1987). Learning the past tenses of English verbs: Implicit rules of parallel distributed processing? In B. MacWhinney (Ed.), Mechanisms of language acquisition. Hillsdale, NJ: Laurence Erlbaum. Sinclair, H., and Bronckart, J-P. (1972). SVO—A linguistic universal? A study in developmental psycholinguistics. Journal of Experimental Child Psychology, 14, 329-348. Slobin, D. (1981). The origins of grammatical encoding events. In W. Deutsch (Ed.), The child's construction of language. New York: Academic Press.

Sentence Comprehension in Italian and Serbo-Croatian

377

Slobin, D. (1982). Universal and particular in the acquisition of language. In E. Wanner and L. Gleitman (Eds.), Language acquisition: The state of the art. New York: Cambridge University Press. Slobin, D. (Ed.). (1985). The cross-linguistic study of language acquisition. Hillsdale, NJ: Erlbaum. Slobin, D., and Bever, T. (1982). Children use canonical sentence schemas: a cross-linguistic study of word order and inflections. Cognition, 12, 229-65. Smith, S., and Bates, E. (1987). Accessibility of case and gender contrasts for agent-object assignment in Broca's aphasics and fluent anomics. Brain and Language, 30, 8-32. Smith, S., and Mimica, I. (1984). Agrammatism in case-inflected language: comprehension of agent-object relations. Brain and Language, 21, 274-290. Sokolov, S. (1988). Cue validity in Hebrew sentence comprehension. Journal of Child Language, 15,129-55. St. John, M., and McClelland, J. (in press). Learning and applying contextual constraints in sentence comprehension. Technical Report, Department of Psychology, CarnegieMellon University. Weist, R. M. (1983). The word order myth. Journal of Child Language, 15, 129-55.

This page intentionally left blank

ROLE OF CONTEXT IN THE COMPREHENSION OF AMBIGUOUS ITALIAN IDIOMS LUCIA COLOMBO Department of General Psychology University of Padua Padua, Italy

1. INTRODUCTION Ambiguous idioms are those expressions that have both a figurative meaning and a literal interpretation. For instance in "the man kicked the bucket" two interpretations are possible, one of which is that the man suddenly died. This meaning of "dying" assigned to the phrase "kick the bucket" is associated with it and stored in memory. The other interpretation is formed by integrating the literal meanings of the single words that form the idiom. Because of the characteristics of the meaning retrieval process, it has been assumed in one model of idiom comprehension, the lexical representation model, that idiomatic phrases are stored in the lexicon as "long words," and the meaning associated is automatically activated in the same vein as for "normal" words (Swinney and Cutler, 1979). The time course of the activation of the figurative meaning and of the literal interpretation's construction may obviously differ, but it is generally assumed that because meaning activation is only dependent on the strength of association between the words that form the idiom, it will be faster than the construction of the literal interpretation, a process that involves not only the level of lexical retrieval, but also sentence-level processes. Thus, the comprehension of ambiguous idioms presents different characteristics with respect to those offered by lexical ambiguity, in which a single word may take more than one independent meaning (i.e., meanings that are not related to Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

379

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

380

Lucia Colombo

each other, as is the case for homonyms). However, the parallelism between the two phenomenons, lexical and idiomatic ambiguity, is such as to make an interesting comparison. One of the most important issues involved in understanding lexical ambiguity, for instance, is the extent to which the meanings associated to a given words are retrieved automatically and independent of context. One very influential theory claims that as the mearing associated to a given word form is automatically activated, the two meanings of a homonym are activated independent of the bias provided by the sentence context (Seidenberg, Tanenhaus, Leiman, and Bienkowski, 1982; Swinney, 1979). The empirical verification of this claim has very important consequences for the construction of a model of sentence comprehension, and for the way the architecture of the human language processor is conceived. Evidence showing that only the meaning biased by the context is activated would imply that lexical and sentence-level processing are not computed by separate and independent mechanisms, carried out by independent "modules" (according to the fodorian terminology; Fodor, 1983; Forster, 1979) but that they do interact. Similar arguments have been provided when considering ambiguity in syntactic parsing (Frazier and Fodor, 1978; Frazier and Rayner, 1987). Also in the case of ambiguous idioms, one can conceive that, as the idiomatic meaning is associated to the idiomatic phrase, it should be automatically activated, independent of whether the sentence context biases the figurative or the literal meaning of the idiom. As was stated, one influential theory of lexical ambiguity has claimed that lexical meanings are automatically activated, and the influence of context only occurs at a later stage, when the meaning appropriate to the context must be selected and integrated into the sentence representation. Support for this model has come mainly from experiments conducted with the cross-modal paradigm. A sentence, biasing one of the meanings of the word, is auditorily presented, and at the offset of the critical ambiguous word a target word related to one of its meanings is visually displayed, and a lexically decision must be made on it (Seidenberg et al., 1982; Swinney, 1979). Although the first studies found rather uniform support for the automatic access model, later studies showed evidence of selective effects of context when different aspects of the context are taken into account (Kellas, Paul, Martin, and Simpson, 1991; Simpson, 1981; Tabossi, 1988; Tabossi, Colombo, and Job, 1987). One such study was conducted in Italian (Tabossi et al., 1987). The experiments reported in this paper showed that when the context biased salient aspects of the meaning of a lexically ambiguous word, only the target word related to the context-biased meaning was activated. This view was further supported by data from Paul, Kellas, Martin, and Clark (1992), who suggested that the extent to which context influences the initial activation of meaning is determined by the overlap between contextually activated features and the sentence words' features. Other studies, conducted in English, pointed out the importance of the relative dominance (frequency) of the ambiguous word's meanings

Context and Ambiguous Italian Idioms

381

(Duffy, Morris, and Rayner, 1988; Rayner and Frazier, 1989; Simpson, 1981; Simpson and Burgess, 1985; but see Onifer and Swinney, 1981). Other more intermediate views have also been proposed between the two extremes—the context-independent and the context-dependent—such as the integration model (Rayner and Frazier, 1989). This very brief review of the literature on lexical ambiguity does not provide definitive suppport to any of the theories proposed, but shows that different aspects of the context and of the ambiguity must be considered in order to decide whether the access of meaning is completely autonomous and contextindependent. It may also give useful indications relative to the parallelism I have suggested between ambiguous words and ambiguous idioms.

2. THE COMPREHENSION OF FIGURATIVE LANGUAGE The models that have been proposed on the representation and the comprehension of idiomatic expressions predominantly suggest that the figurative meaning is very fast to become active, in particular because idioms are sequences of words that have acquired a meaning that is often not related to, nor derivable from, the literal interpretation of the expression. Moreover, very often idiomatic expressions do not even have a possible literal interpretation (i.e., "to burn the midnight oil"). Thus, the dominant view is that the meaning associated with an idiom is retrieved automatically (Gibbs, 1980; Swinney and Cutler, 1979) as opposed to a view according to which it is retrieved in a special "idiom list" only after the literal meaning has been found to be implausible (Bobrow and Bell, 1973). Some peculiar aspects of the idioms may suggest that they are indeed represented somewhat differently with respect to normal words. One instance is relative to the well formedness of the expressions. Some idioms are characterized by violations of either selectional restrictions ("rack your brains") or of syntactic wellformedness, or predictability ("to lose face"). Another peculiar characteristic of idioms is syntactic frozenness: the figurative meaning of some idioms, in fact, can only be maintained when they are in their base form. This aspect has been shown to be important from the processing point of view, as frozen idioms are rated as acceptable in sentence acceptability judgments faster than control expressions (Gibbs and Gonzales, 1985; Swinney and Cutler, 1979). Thus, idioms may be seen as peculiar expressions that require specialized processing mechanisms and special representations. However, this view is by no means the most dominant one, in particular since it has been shown that some idiomatic expressions may be decomposable and analyzable. That is, for some idioms the meaning of the single words contribute to the figurative meaning of the whole phrase (Gibbs and Nayak, 1989; Gibbs,

382

Lucia Colombo

Nayak, and Cutting, 1989; Nunberg, 1978; Wasow, Nunberg, and Sag, 1983), and in a sentence acceptability task, decomposable idioms are evaluated faster than nondecomposable idioms, and faster than literal phrases (Gibbs et al., 1989). This advantage of decomposable idioms is only found, however, when they are in the base form and disappears when they are in a syntactically more complex form. In fact, McGlone, Glucksberg, and Cacciari (1994) found that reading times for "Sam spilled the beans" were faster than for the control condition ("Same told him all"), but they were no faster for "Sam didn't spill a single bean" than for "Sam didn't say a single word". Although many of the data reported on idiom comprehension have used offline paradigms, it is critical, in evaluating whether the processing of idioms is part of normal language processing, to consider what happens to the comprehension system while the idiom is presented, or immediately after, in particular when there are no specific cues to signal the presence of an idiomatic expression. To this extent an important characteristics of idioms has been shown to be predictability (Cacciari and Tabossi, 1988; Tabossi and Zardon, 1993; Titone and Connine, 1994). Predictability has been operationally defined as "the point in the string after which the probability of the fragment to continue idiomatically is very high, even though a different, literal completion may still be possible." (Tabossi and Zardon, 1993). So given the fragment "to be in seventh . . ." subjects tend to complete it with "heaven" rather than with a literal expression, and the point after which the completion is with most probability idiomatic has been named the "key" in the configuration model (Cacciari and Tabossi, 1988; Tabossi and Zardon, 1993, 1995). According to this model, only when the key is available to the listener is the figurative meaning of the idiom activated. The influence of predictability has been demonstrated also by Titone and Connine (1994). In their experiments, in which the cross-modal priming paradigm was used, they showed that for high-predictable idioms in a neutral context there was evidence of activation at a position before the last word of the idiom, both when the idioms were ambiguous (i.e., with a possible literal interpretation) and when they were unambiguous. This activation decreased at the offset of the idiom but only for ambiguous idioms. For low-predictable items there also was evidence of idiom activation at the penultimate position, but the amount of activation was smaller than for high-predictable idioms. Thus their experiments showed that predictability and ambiguity had different effects on the activation rate of the figurative meaning. They also showed that for ambiguous idioms, the literal meaning of the last word of the idiom was also active at offset for high-predictable idioms, although it was not so when the idioms only had a figurative meaning. Titone and Connine' s results, showing evidence of automatic activation of the figurative meaning and of the literal meaning of the words of the idiom, although the extent and the temporal course of activation are modulated by predictability

Context and Ambiguous Italian Idioms

383

and ambiguity, concerned the comprehension of idioms in a neutral context. Another aspect that is important to consider is, however, the role of context in the comprehension of sentences including idiomatic expressions. If idioms are represented and processed by special processing mechanisms, the context may function as a special cue to index these special representations and mechanisms. Moreover, if idioms are ambiguous, and may take a plausible literal interpretation, there might possibly be different effects on the extent to which there is activation of the idiomatic meaning, the way construction of a literal interpretation is carried out, and on their temporal courses. The major focus of the present study is on the role of context on the comprehension of ambiguous idioms. Three experiments will be presented. The experiments were carried out in Italian and will offer a comparison to the studies conducted mainly in English. The first experiment is an off-line experiment in which fragments of idioms are presented to be completed by subjects. In experiments 2 and 3 the relative time course of construction of a literal interpretation and activation of the figurative meaning are investigated with the self-paced reading paradigm. This is an on-line method, which allows inferences to be made on the representations and processes involved in the task while they occur. It makes an interesting comparison to the more frequently used cross-modal priming paradigm. The latter is generally used to measure the amount of facilitation of a target word related to the critical probe (ambiguous word or idiom). For instance, for ambiguous words the target is related to one of the meanings. The facilitation that occurs is therefore an index of the meaning's activation. For idioms, evidence of facilitation when a target word is literally related to the last word of the idiom can be interpreted as showing that the literal meaning of the last word of the idiom has been activated, but not necessarily that the construction of the literal interpretation of the idiomatic phrase has been initiated or completed. Thus, evidence of activation of the literal meaning of a word of an idiom cannot say anything about the sentence-level processes occurring while the idiom phrase is processed. Only evidence of the lack of the word's literal meaning's activation can be unambiguously interpreted, because if the meanings of the literal phrase's words are not activated, it is very unlikely that the construction of a literal interpretation has been carried out at all. In all experiments care was taken that subjects did not notice at all (or only in the least) (a) the presence of idiomatic expressions or (b) that the experiments were aimed at investigating the processing of figurative language. Any theory that strongly claims that the figurative meaning's activation is automatic, even admitting that the amount and the temporal course of this automatic activation may be sensitive to different dimensions of the idioms or to other aspects of the experiment, must also claim that it occurs independently of those aspects and dimensions.

384

Lucia Colombo

3. THE COMPREHENSION OF AMBIGUOUS IDIOMS IN AN OFF-LINE TASK: EXPERIMENT 1 The main aim of experiment 1 was to verify, with an off-line task, whether the interpretation assigned to ambiguous idiomatic strings is influenced by the context. Subjects were presented with a sentence containing the initial part of the idiom, that is, the verb, preceded by either a neutral context or a context biasing the literal interpretation or the figurative interpretation, and were required to complete the sentences. For instance, subjects might be asked to complete sentences such as: "The rich old man was very ill and his inheritor was hoping that he kicked..." If the idiom tends to be perceived as a unit, the presentation of its first part should tend to elicit its completion, independently of the context, because the connection strength between the lexical nodes of the words of the idiom should be strong, as compared to those for phrases that are not idiomatic. In the same vein, the presentation of a fragment of a word "capt" may cue the initial activation of all words consistent with the input ("captain, captive") as is assumed in a wellknown model of spoken-word recognition that has received support from experimental data (Marslen-Wilson, 1987; Zwitserlood, 1989). On the other hand, it might be that because the idioms may have a literal interpretation, the connection strength between the verb and the noun phrase (NP) of the idiom is not so large because it is somewhat attenuated, as the possible continuations of the phrase are very many. Indeed, if we consider a familiar Italian idiom like prendere un granchio (which means roughly 'to make a mistake'), the verb prendere ('to take') is a very frequent one, and one with several meanings, dependent on context, as in its English translation. For instance, one can prendere un libro dallo scaffale ('take a book from the shelf), or prendere un raffreddore (literally 'take a cold'). Thus, considering all the possible contexts in which the verb has been experienced, the probability that the NP of the idiom is given as a completion might not be so high.

3.1. Method 3.1.1. SUBJECTS Forty-eight students of the University of Padua were tested. 3.1.2. MATERIALS Eighteen ambiguous idioms and the contexts biasing the literal and figurative contexts used in experiment 3 of the Colombo study (1993; and there listed) were used in this experiment. To evaluate whether idioms were fairly familiar expres-

Context and Ambiguous Italian Idioms

385

sions in their figurative meaning, 23 subjects who did not take part in experiment 1, were asked to make a paraphrase of the 18 idioms. Of these, 15 were familiar to above 80% of the subjects; three were a bit less familiar (one was identified by 61% of the subjects, and 2 by more than 70% of the subjects). The idiom strings were also evaluated for familiarity by six new subjects, on a 5-point scale, ranging from very familiar (1) to very unfamiliar (5). The mean rating assigned to the 18 idiom strings was 2.38 (SD = .73). To these materials a list of 18 sentences forming the neutral contexts was added. The neutral contexts were constructed so as to be rather vague and to be congruent with both a literal and a figurative intepretation of the idiom, such as, "Vorrei sapere cosa hafatto Giulia quando ha preso ..." ('I'd like to know what Giulia did when she took . . .'). There were three lists; one list contained 18 neutral context sentences plus 25 filler sentences. Each of the remaining two lists contained 18 experimental sentences, 9 with a literal context and 9 with a figurative context, with different idioms, so that the group of subjects with the same list was not presented with the verbs relative to the same idioms, plus 25 filler sentences. The neutral context sentences were presented by themselves, so as not to induce an "idiomatic mode of processing" (Bobrow and Bell, 1973). 3.1.3. PROCEDURE Subjects were presented with one of the sentence lists, and were asked to read each sentence and make well formed and plausible completions. They were required to proceed until the end, and not to go back when a sentence was completed. 3.2. Results and Discussion From the completions data, 1.85% of the scores were excluded because they were missing and 1.04% because of an experimenter's error. The completions were scored according to three categories. One category included completions in which the NP of the idiom was exactly reported; the second category included completions formed by NPs referring to a literal interpretation of the idiom string (such as 'the pail' for 'kick'). The third category included completions in which the NP provided by subjects was formed by a string with a figurative meaning, different from the critical idiom. For instance, in this category completions such as le mani after the verb mangiarsi ('to eat') could be included, which is part of the idiom mangiarsi le mani ('to regret something very angrily') rather than il fegato, which is the NP of the critical idiom mangiarsi ilfegato ('to be enraged'). Also in this category were included completions that did not form an idiomatic string but nonetheless had a figurative meaning.

386

Lucia Colombo TABLE 1 PERCENTAGES OF DIFFERENT TYPES OF COMPLETIONS IN THE THREE CONTEXT TYPESa Context

I

L

F

Idiomatic Literal Neutral

54.6 26.2 9.6

9.5 66.8 67.6

35.9 7.0 22.8

a

I, idiom; L, literal; F, figurative.

The percentages of frequency scores for the three categories, relative to the three types of contexts, are shown in Table 1. They indicate that subjects' completions significantly differ within each context type: (X2(2) = 555.8, p < 001 for the idiomatic context; X2(2) = 1010.3, p < .001 for the literal context; x2(2) = 943.8, p < .001, for the neutral context) and tend to be strongly influenced by the context. When the context biases the idiomatic meaning, the completions reporting exactly the final part of the critical idioms are rather high (54.6%). If we add to the latter the idiomatic completions different from the critical ones, the proportion of figurative types of completions increases up to 90.5%. The reason for the larger proportion of figurative completions is given by the fact that for many of the verbs of the critical idioms there are several different possible completions in figurative domains, which sometimes are themselves idioms. When the context biases a literal interpretation of the critical phrase, the completions fall more frequently in the literal completion category. The idiomatic completions in this context type are also fairly high (26.2%), but this may be due to the fact that for some idiomatic phrases the NP of the idiom itself was a very plausible completion, although the interpretation was still literal. For instance, for the sentence "One evening Andrea went into the garden with Maria and looking at the sky has shown her . . ." the completion "the stars" is quite plausible and likely, but it is much more likely that the literal interpretation was meant, because the figurative interpretation of vedere le stelle ('to feel an acute pain') would be rather odd in this context. The percentage of these ambiguous types of completions was rather high due in particular to five idioms (above 40%); when these were removed the percentage of idiomatic completions in the literal context fell down to 9.3%. Completions in the neutral context are interesting because they reflect the extent to which a literal or idiomatic completion is assigned, independently of context. As is apparent, they tend to fall in the literal category (67.6%), while those in the figurative and idiomatic categories together are only 32.4%. The comparison of this figure with that corresponding to the same categories pooled in the idiomatic context condition (90.5%) suggests that when both a literal and a figurative inter-

Context and Ambiguous Italian Idioms

387

pretation is possible, subjects tend to assign the former rather than the latter, despite the fact that many completions of the figurative type are possible. We can conceive of this completion task as a measure of the extent to which the idiom has been lexicalized (Titone and Connine, 1994). If this is the case, the proportion of idiomatic completions should correlate with the familiarity of the idioms. The higher the familiarity, the more likely it should be that a completion corresponding exactly to the final part of the idiom is given. The correlation between the familiarity ratings (see the method section), and the proportion of idiomatic completions in the three types of contexts is instead not very high, is nonsignificant, and is relatively higher only in the idiomatic and neutral contexts (r = —.37, for the completions in the idiomatic context; p > .05; r = —.31, for those in the neutral context; p > .1; r = —.04, p > .1 for those in the literal context; recall that low scores corresponded to high familiarity ratings). Thus the lack of a significant correlation with familiarity, coupled with the predominance of literal completions in the neutral context, suggest that it is mainly the context that dictates the probability of idiomatic completions and therefore also the probability that a figurative interpretation is assigned to the idiom. These data would suggest that when idiomatic expressions also have a plausible literal interpretation, some sort of cue, either the context or some other sort, is necessary for the figurative interpretation to become active. However, the data of experiment 1 are relative to an off-line task, which reflects the final state of the comprehension process. So it could be that the idiomatic interpretation is active for a short time and then inhibited, even in neutral contexts. To know whether there is an automatic activation of the idiomatic meaning it is important to have some on-line measure during, or immediately after, the idiom presentation. Thus, any conclusion will be postponed until the general discussion.

4. THE COMPREHENSION OF AMBIGUOUS IDIOMS IN ON-LINE TASKS: EXPERIMENTS 2 AND 3 In a former study Colombo (1993) has shown in an on-line task that when ambiguous idioms are embedded in sentence contexts, the idiomatic meaning is activated only when the context biases the figurative meaning of the idiom. In experiments 2 and 3 of the Colombo (1993) study subjects were visually presented with sentences such as, "The man was painting the room when he stumbled and kicked the bucket that was on the kitchen floor" in which the idiom "kick the bucket" is used in the literal sense in a perfectly plausible way. This condition was contrasted with the following: "The rich old man was very ill and his inheritor was hoping that he kicked the bucket that was on the kitchen floor." Obviously,

388

Lucia Colombo

the latter sentence becomes nonsensical in its last portion, where the relative clause starts. This is because the relative pronoun and the subsequent relative clause naturally refer to the noun of the preceding (idiomatic) NP. To make reference to just the NP (rather than to the whole idiom phrase) is only possible when assigning a literal interpretation. However, because the idiom is embedded in a figurative context, it acquires a figurative meaning, and it does so only when taken as a unit. Thus, while an idiomatic interpretation is still current, assignment of a literal interpretation to the NP before the relative clause leads to the simultaneous presence of two incompatible interpretations. This type of structure therefore was supposed to create a garden path, because when the figurative meaning is active and has been, or is going to be, integrated into the preceding context, the presentation of the relative clause disrupts the processing flow, presumably inducing the processor to reexamine the ongoing interpretation. In Colombo's (1993) experiment 3 there was also a control condition, in which the noun of the idiom, embedded in a figurative context, was replaced by another noun, so that the sentence no longer contained an idiomatic expression, as in the following instance: "The rich old man was very ill and his inheritor was hoping that he kicked the pail that was on the kitchen floor." This type of sentence was also implausible, and was used as a control of the idiomatic context condition. As mentioned above, in the figurative condition it was expected that the disruption produced by the relative clause would cause the processor to reexamine the current interpretation, and this presumably would cause some delay with respect to the literal condition, in which, because of the lack of ambiguity, processing should occur more fluently. But because the idiomatic sentence was implausible, it was necessary to compare it also with another implausible condition, with no ambiguous interpretation. Indeed, it was found that the presentation of the relative clause produced a delay in latencies for the figurative context condition, as measured with respect to the literal condition and also with respect to the implausible/nonfigurative condition. Apparently, subjects immediately decided that the control condition was implausible and therefore did not spend much time on it. In the idiomatic context condition, instead, it took time for the processor to resolve the inconsistency between the interpretation created before the relative clause, and the interpretation structurally possible after the relative clause. In experiment 2 of Colombo's study, the same type of idioms were embedded in neutral and literal sentence contexts. These type of contexts created really ambiguous conditions, in particular the neutral one, for which both the figurative meaning and the literal meaning could be simultaneously possible, and were contrasted with a control condition in which the verb of the idiom was replaced by another verb. In all these conditions the sentences were plausible. If, as claimed by some, the figurative meaning is automatically activated, even in nonbiasing contexts, two interpretations should at some point be active. This might produce

Context and Ambiguous Italian Idioms

389

a delay, because the processor should spend some time to resolve the ambiguity. Evidence from eye fixations studies (Rayner and Frazier, 1989), for instance, showed that gaze durations on ambiguous words, presented in a neutral context, were longer when the two meanings of the ambiguity were equally likely, than when one meaning was dominant over the other. As Rayner and Frazier (1989; see also Duffy, Morris, and Rayner, 1988; Rayner and Duffy, 1986) argue, "gaze durations ... were lengthened presumably because of the initial access and selection processes associated with the two meanings." When, instead, the two meanings were not equiprobable, fixation latencies were longer in the disambiguating context when this was consistent with the secondary meaning. Once again, accessing the subordinate meaning, which was available slowly, and integrating it into the context was shown to be a time-consuming process. On the basis of this type of data, in the Colombo's study it was assumed that the simultaneous activation of two meanings and the selection and integration processes do take time, and so at least in the neutral context condition latencies were expected to be longer with respect to the control condition. However, no delay was apparent in either experimental condition as measured with respect to its control. The results of Colombo's study clearly contrast with the rather dominant claim in the literature that the figurative meaning of idiomatic expressions is automatically activated independently of the context (Gibbs, 1980). At least when the idioms are ambiguous, this claim is not supported. In the present study, other experiments are presented that replicate and give further support to the former findings. In experiment 3 of Colombo's (1993) study, it was found that the idiomatic context condition yielded longer latencies in the region immediately following the idiom with respect to the literal condition, and with respect to the idiomatic control condition. This was interpreted as an effect of the incompatible interpretations present only in the idiomatic context condition. However, it might be that the figurative meaning is to some extent activated also in the literal context condition, and perhaps quickly rejected before the time of the relative clause's presentation. Thus, some delay may be detectable at some point during the processing of the sentence, assuming once again on the basis of several authors' data and models (Duffy et al. 1988; Frazier and Rayner, 1987, 1990; Rayner and Duffy, 1986; Rayner and Frazier, 1989) that the processing system delays when a selection is to be made. This delay can only be measured with respect to a proper control, however. Therefore, in experiment 1 of the present study three conditions were included. The first two were an idiomatic context condition and a literal context condition similar to those used in experiment 3 of the Colombo's former study. The third was a literal control condition, in which the context preceding the idiom was biasing the literal interpretation, but the noun of the idiom was substituted by another noun, so that in this condition the sentence contained no idiom (see Table 2). The

390

Lucia Colombo

TABLE 2 SAMPLES OF MATERIALS (TRANSLATED FROM ITALIAN) IN EACH CONDITION OF EXPERIMENT 2 Idiomatic Literal Literal Control

The disappointed athlete understood that he could not make it and decided to throw the sponge that was dripping water. In the bathtub where he was taking a bath the child wanted to throw the sponge that was dripping water. In the bathtub where he was taking a bath the child wanted to throw the ball that was dripping water.

literal context condition could then be faster than the idiomatic context condition, but slower than the control, if some activation of the figurative meaning has occurred at some point. This is because the literal interpretation must necessarily be initiated (that is, the meanings of the words must be activated and integrated into the sentential context). If and when the idiomatic meaning becomes active, the processor must select which of the ongoing interpretations must be abandoned. This process, as said above, should require time. 4.1. Experiment 2 4.1.1. METHOD 4.1.1.1. Subjects of Padua.

The subjects were 21 volunteer students of the University

4.1.1.2. Materials From the set of ambiguous idioms used in the Colombo (1993) study 18 idioms were selected. These were all of the form "verb phrase noun phrase," like "kick the bucket" and could be interpreted either literally or figuratively. For each idiom, three types of sentences were created, one biasing a literal interpretation of the idiom, and containing the real idiom. The second was a control condition that had the same context as the former, but the noun of the NP of the idiom was replaced by another noun, possibly a synonym, which did not allow a figurative interpretation to be made. In the third sentence type the context biased the figurative intepretation of the idiom. Thus the list of experimental stimuli was made up of 18 idioms by three sentence types for a total of 54 experimental sentences. Three lists were constructed; in each list only one sentence type was included for each idiom. Test sentences were constructed so that the idiomatic context condition was always implausible, whereas the literal and control conditions were plausible. To the list of experimental sentences 58 filler sentences were added, none of which contained idioms, and 22 of which were implausible. In each of the three lists there were the 18 experimental sentences, only 12 of which contained idioms

Context and Ambiguous Italian Idioms

391

(the rest were control strings), plus the 58 filler sentences, for a total of 70 sentences, 28 of which were implausible. Thus the proportion of sentences containing idioms was kept relatively low (17%), with the aim of avoiding the creation of sets or expectations for idioms among subjects. Three groups of subjects were tested, and each group was randomly assigned to one of the list conditions. At the end of the experiment each subject was shown the list of idiomatic expressions and asked to pharaphrase the figurative meaning of the idiom. This was done in order to ascertain the degree of familiarity of the test idioms. In general they appeared from the data as quite familiar idioms. Of 18 idioms, 15 were known by above 85% of the subjects, 1 by 54% of the subjects, and the other 2 were in between. The data for unknown idioms were removed from further analyses. 4.1.1.3. Procedure The experiment was run with the MEL (Micro Experimental Laboratory) software. Subjects were seated in front of a computer screen. The sentences were presented on the screen chunks by chunks in a self-paced reading task. The chunks were formed by groups of words that did not necessarily correspond to phrases or clauses, in order to avoid expectations by subjects on when to press the advance button. The critical chunks, which were afterwards subject to analyses, were four. The first was formed by the verb of the idiom (S1), the second by the NP (S2), the third by the relative pronoun plus the immediately following verb (S3), and the fourth by the rest of the sentence (S4). The subjects were instructed to read and understand the sentences carefully, and then at the end of each sentence they had to decide quickly whether the sentence was plausible or not by pressing one of two buttons, labeled "yes" and "no." After the subjects read the instructions they had to make the experiment's presentation start by pressing the "plus" button of the numeric keyboard. This started the presentation of the sentence lists. When the subjects had read each chunk they pressed the "plus" bar so that the former chunk disappeared from the screen and the following appeared in the following region. After the presentation of the last segment of the sentence, when the subject had pressed the "plus" bar, the chunk disappeared from the screen and a question was presented: "Is the sentence plausible or implausible?" The subjects pressed the "plus" bar with their right hand in order to make the presentation advance. To make the plausibility decision they pressed the 1 button of the keyboard with the index finger and the 2 button with the middle finger of the left hand. 4.1.2. RESULTS The data were formed by the mean reading latencies of the experimental sentences' critical chunks, and by the mean latencies and errors for the plausibility

392

Lucia Colombo TABLE 3 MEAN READING LATENCIES FOR THE CRITICAL CHUNKS, MEAN RESPONSE LATENCIES, AND PROPORTION ACCURACY (IN PARENTHESES) IN THE PLAUSIBILITY DECISION IN EXPERIMENT 2 Idiomatic

Literal

Literal Control

S1 S2 S3 S4

727 805 824 829

745 819 764 785

765 811 768 794

Reaction time

689 (74)

691 (70)

724 (72)

decision. Only the data for which the plausibility decision was correct were kept for further analyses. The data relative to two sentences for each list were removed because they elicited too many errors in the decision task. From the reading latencies data latencies below 200 ms and above 1400 ms were removed. Thus 6.1% of the data were discarded. The mean reading latencies on each critical segment, the same as those used for presentation, are displayed in Table 3. The ANOVA conducted on the first and second segments of the idiom did not show any significant effect. The analysis on S3 showed a significant effect of type of condition only in the analysis by subjects [F(2,20) = 3.77, MSe = 6226.7, p < .05, by subjects; F < 1, by items]. T-tests among the means of the different conditions showed that the latencies in the idiomatic context condition were significantly longer than in the other two conditions (p < .05), but the latter were not significantly different from each other. Although the trend was in the same direction as for S3, the three conditions of S4 did not significantly differ. No significant effect was found in the analysis of the decision task for either latencies or errors. 4.1.3. DISCUSSION The results of experiment 2 show that when the context biases a figurative meaning of an ambiguous idiom, reading times in the postambiguity region are increased, in agreement with the predictions and with the data of Colombo (1993). Thus, when the continuation of the sentence is only consistent with a literal interpretation of the idiom, in contrast with the figurative interpretation biased by the context and presumably pursued by the processor, a garden-path situation is formed, and a reanalysis of the potential interpretations is required. This process takes times, and therefore subjects tend to pause for a longer time on the critical segment. During this time delay presumably the subjects decide that the sentence

Context and Ambiguous Italian Idioms

393

is implausible. At the end of the sentence the plausibility decision then is executed relatively quickly for these sentence conditions. In this experiment no difference was found between the literal context condition and its control in any of the critical segments. It was argued above that if the idiomatic meaning is activated independent of context, there should be some evidence of this fact. At some point the processor should entertain two hypotheses, and the decision on which one to discard should take some time. This might have resulted in a lengthening of reading latencies, presumably in the second segment containing the NP of the idiom. But this hypothesis has not found any support in the present data. The only evidence that might be taken as indicative of the fact that some activation of the idiomatic meaning has occurred when the idiom string is embedded in a literal context comes from the pattern of errors in the plausibility decision of experiment 3 in Colombo (1993). In that experiment it was found that there were more errors in the decision task at the end of each sentence in the literal context condition, one in which the idiom was presented, but the context biased the literal interpretation. This higher percentage of errors might be attributed to a late activation of the idiomatic meaning, which would interfere with the decision. Alternatively, it might be due to some characteristics of the sentences of the literal context conditions, such as, for instance, the degree of plausibility of the sentences. Experiment 3 was aimed at replicating the results of the former experiments and at controlling which of these possible interpretations of the results are correct. In experiment 3 there were thus two conditions determined by the context, biasing the idiomatic or the literal interpretation of the idiom, and two conditions determined by whether the sentences contained the idiom, or did not (control sentences). In this experiment, therefore, both the idiomatic context sentences and the literal context sentences had their own control. Moreover, the experiment was preceded by a pretest, in which the plausibility of the sentence contexts was evaluated by a different group of subjects. TABLE 4 SAMPLES OF MATERIALS (TRANSLATED FROM ITALIAN) IN EACH CONDITION OF EXPERIMENT 3 Idiomatic Id. Control Literal Lit. Control

The disappointed athlete understood that he could not make it and decided to throw the sponge that was dripping water. The disappointed athlete understood that he could not make it and decided to throw the cloth that was dripping water. It was in the bucket that the cleaning lady had to throw the sponge that was dripping water. It was in the bucket that the cleaning lady had to throw the cloth that was dripping water.

394

Lucia Colombo

4.2. Experiments 4.2.1. METHOD 4.2.1.1. Subjects Forty students of the University of Padua participated in this experiment. None of them had been a participant in the former experiments. 4.2.1.2. Materials for the Pretest The ambiguous idioms used in this experiment were selected from the list of 18 idioms of experiment 1 on the basis of the data from a pretest experiment, in which the sentence contexts were evaluated for plausibility. In the pretest all 18 idioms were used. Three types of context sentences were evaluated: the idiomatic and the literal contexts, both containing the idiom, and the literal control in which the NP of the idiom was replaced by another noun. As was described for experiment 2, the sentences of the idiomatic context condition became implausible after the idiom itself, when the relative clause was presented. However, it was deemed important to determine their overall plausibility with respect to the literal condition, until the point at which it became implausible. The sentences for the control conditions were exactly the same as the sentences for the related experimental conditions, except that the noun of the NP was replaced with either a synonym or an associate. For instance, for the literal context condition a sentence was, "It was in the bucket that the cleaning lady had to throw the sponge (meaning 'to give up')," and for the control the noun "sponge" was replaced by "cloth." The idiomatic context for the same idiom was "The disappointed athlete understood that he could not make it and decided to throw the sponge," and in the control condition "sponge" was also replaced by "cloth." In this way the sentence became implausible even before the relative clause. Thus its plausibility was not evaluated. Each of the three lists contained 18 test sentences to be evaluated and 25 filler sentences. The test sentences were formed by six sentences in each of the three conditions described above, with different idioms for each condition. The order of presentation of the sentences was random. Subjects were instructed to read carefully each sentence and then rate it for plausibility on a 5-point scale, ranging from very plausible to very implausible. Fifty-one University of Padua students rated the sentences. The mean ratings and the standard deviations relative to the three conditions for each idiom were calculated. The average rating for the literal context condition was 2.57 (SD = .81), for the literal control 2.89 (SD = .98), and for the idiomatic context 2.68 (SD = .49). T-tests were conducted on the three means, and none was found to be significantly different from the other. Thus, on the average, the sentences had equivalent plausibility ratings. However, the ratings were also analyzed separately for each idiom to see whether the three conditions for each idiom were rated with different levels of plausibility. The results of this analysis showed

Context and Ambiguous Italian Idioms

395

that for 8 of the 18 idioms the context sentences were not rated of equivalent plausibility. Thus the sentences relative to the three conditions for each of the eight idioms that had not yielded homogeneous conditions were changed, and these changed sentences were again rated by a new group of 30 University of Padua students. As before, for each idiom there were three sentence types, and three lists were made up, in one of which there were 8 test sentences plus 10 filler sentences. On the basis of the analysis of the latter data, only five idioms were found to have sentence context rated as equivalent for plausibility. Among these, and those from the former analysis, 12 idioms were selected with the relative contexts, for a total of 36 experimental sentences. 4.2.1.3. Materials for the Experiment The 36 experimental sentences selected on the basis of the pre-test were formed by the 12 context sentences with a figurative bias, the 12 with a literal bias, and the 12 controls of the literal bias sentences. To these were added 12 experimental sentences that were the same as the idiomatic context condition, except that the noun of the idiom was replaced by another noun, not forming an idiomatic expression. These 12 sentences formed the control of the idiomatic context condition. To the 48 experimental sentences thus obtained were added 48 filler sentences, not containing idiomatic expressions. Thus each subject was presented with the 48 filler sentences, plus the 12 experimental sentences, for a total of 60 sentences, 32 of which (6 test sentences plus 26 filers) were implausible. 4.2.1.4. Procedure The procedure was the same as for experiment 2, except for the critical segments. The first (S1) was formed by the verb of the idiom, the second (S2) by the NP of the idiom, the third (S3) by the relative pronoun, and the fourth (S4) by the rest of the sentence. In contrast to experiment 2, only the relative pronoun was included in S3, because this distinction was thought to show more clearly the garden-path effects. 4.2.2. RESULTS The dependent variables in this experiment were the latencies to read in the self-paced task, and the latencies and errors to respond in the plausibility decision at the end of each sentence. In the analyses were included the data for subjects whose latencies in the reading task were above 100 ms and below 1500 ms, and whose latencies in the decision task were above 150 ms and below 1500 ms. The data of six subjects were discarded, because five of them did not perform the task appropriately, as indicated by the decision task, and one subject's reading latencies were too long. The latencies for the remaining 34 subjects were analyzed by an ANOVA with two factors, context type (figurative vs. literal) and sentence type (experimental

396

Lucia Colombo

vs. control). Only the latencies for the critical segments, the same as used for the presentation, were analyzed in the reading task. Segment 1 (S1) was formed by the verb of the idiom, segment 2 (S2) by the NP of the idiom, segment 3 (S3) by the relative pronoun, and segment 4 (S4) by the rest of the sentence (usually a verb plus an NP or a prepositional phrase). The mean reading latencies and response times are displayed in Table 5. The ANOVA on the reading latencies on S1 showed a significant effect of context [F(l,33) = 4.75, MSe = 7322.7, p < .05, by subjects; no significant effect by items] and no other significant effect. Reading times for the sentences in the literal-bias context were faster than those in the idiomatic-bias context. Also the analysis of S2 showed a significant effect of context, in the same direction as for S1, [F(l,33) = 8.9, MSe = 13310.8, p < .01, by subjects; F(l,28) = 14.5, MSe = 5279.07, p < .01 by items]. The factor context was significant in S3 too, [F(l,33) = 4.16, MSe = 6801.01, p < .05 by subjects, not significant by items], while sentence type was only marginal [F(l,33) = 3.05, p < .1, only by subjects]. The critical interaction between the two factors was also significant [F(l,33) = 10.04, MSe = 5776.3,p < .01 by subjects; F(l,32) = 8.26, MSe = 6279.2, p < .01 by items]. Planned comparisons showed that reading times for the sentences biasing the idiomatic context were significantly longer with respect to their control, [t(33) = 3.06, p < .05], whereas those for the literal context condition did not significantly differ from their control. Last, in the analysis of S4, the factor context was again significant [F(l,33) = 6.8, MSe = 5377.7,p < .05 by subjects, not significant by items], while the interaction was marginally significant [F(l,33) = 4.00, MSe = 9487.7, p < .1 by subjects only]. Planned comparisons showed that again the difference between experimental and control sentences was significant for the idiomatic context condition [f(33) = 2.1, p < .05], not for the literal context condition. The ANOVA conducted on the decision latencies showed a significant effect of context [F(l,33) = 45.6, MSe = 27966.3, p < .001 by subjects, F(l,32) = 7.93, MSe = 35320.9, p < .01 by items], a marginal effect of sentence type [F(l,33) = 3.6, MSe = 23010.7, p > .05, by subjects only] and no TABLE 5 MEAN READING LATENCIES FOR THE CRITICAL CHUNKS, MEAN RESPONSE LATENCIES, AND PROPORTION ACCURACY (IN PARENTHESES) IN THE PLAUSIBILITY DECISION IN EXPERIMENT 3 Idiomatic S1 S2 S3 S4 Reaction time

566 642 634 662 449 (93)

Control 547 649 583 601 473 (81)

Literal 515 584 564 596 618 (69)

Control 534 589 570 601 692 (79)

Context and Ambiguous Italian Idioms

397

other significant effect. The main effect of context indicates that the decision latencies for the figurative context conditions were faster than those for the literal context conditions. The analysis of accuracy in the decision task revealed a significant effect of context [F(l,33) = 15.41, MSe = 386.4,p < .001 by subjects; F(l,11) = 6.09, MSe = 437.05, p < .05 by items] showing that accuracy was higher for the conditions with an idiomatic context. There was no effect of sentence type, but the interaction of context by type of sentence was significant [F(l,33) = 7.63, MSe = 568.4, p< .01, by subjects; F(l,11) = 5.08, MSe = 265.04,p < .05 by items]. The interaction occurred because accuracy was higher for the condition in which the sentences contained the idiom, with respect to its control, in the figurative context condition [t (33) = 2.25, p < .05], while there was no difference between experimental and control sentences in the literal context condition. 4.2.3. DISCUSSION The significant interaction that has been found in experiment 3, between context type and sentence type, in the portion of the sentence immediately after the idiom string, shows that reading latencies increase only when the idiom string is embedded in the figurative context condition. In this condition, it will be recalled, a subordinate phrase introduced by a relative pronoun was presented immediately after the idiom. This portion of the sentence, for which the referent of the relative pronoun must be the noun of the idiom phrase, was predicted to yield a condition of incongruency with the on-going interpretation of the sentence, which, because of the context bias, should refer to the idiomatic meaning of the ambiguous string. Thus, the results show that the idiomatic interpretation of the idiom has been activated by the figurative context. On the contrary, no difference between experimental and control conditions was found when the sentence context biases the literal meaning of the idiom. If the figurative meaning of the idiom had been automatically activated, the processor should at some point have entertained two interpretations simultaneously and should have taken a decision on which of the two was most likely. This should have produced evidence of some delay, at some point during sentence processing, as was argued above on the basis of experimental evidence on other types of ambiguities (Frazier and Rayner, 1987, 1990; Rayner and Frazier, 1989). But no such evidence was apparent. A comparison between the experimental and control conditions in the idiomatic context did not show any difference in any of the critical segments. This same lack of difference was found in experiment 3 of Colombo (1993). One might wonder whether at the point when the control condition becomes implausible, or immediately afterward, a delay shouldn't appear, as evidence that the processor has detected the implausibility. Apparently, the detection of implausibility did not

398

Lucia Colombo

require the processor to spend much time before making a decision. It will be recalled that half of the sentences subjects were presented with were implausible; therefore an implausible sentence was expected. On the other hand, in the experimental condition, where the garden path occurred, the processor may have had to decide whether a literal interpretation might still be possible, or whether the sentence had to be considered implausible. Apparently, deciding that a word is odd in a context where odd words are expected is not very effortful, whereas deciding among possible alternative interpretations is. Alternatively, the lack of difference between experimental and control conditions may be attributed to the fact that the self-paced reading technique is not very sensitive to semantic activity. According to this view, it would not be possible to detect differences in reading latencies due to the semantic activation of the figurative meaning of the idiom in the literal context, or to the detection of implausibility in the idiomatic context, unless this activation created problems to the processor in terms of selection and decision processes at the interpretive level. Perhaps these may be avoided when the context is sufficiently constraining. According to this interpretation, only when competing interpretations are active at the sentence level might differences be detected by the self-paced technique. This issue deserves further attention, and suggests the opportunity of a comparison among different paradigms. The results of the present experiment also show significant effects of context type. Reading latencies tended to be faster for the literal context conditions than for the figurative context conditions. These differences were independent of the presence of the idiom string, because they were present also in the control conditions. Moreover, subjects took longer and were less accurate to decide about the plausibility of the sentences in the literal context conditions. The advantage of the literal context conditions in reading times was apparent in the first portion of the idiom string, so that it must be attributed to differences in the type of context. This pattern of data suggests a tendency for subjects to find an interpretation of the figurative context conditions more difficult to construct than that of the literal context conditions. The two figurative context conditions, it will be recalled, became implausible at some point during sentence reading. Thus, when the plausibility decision task occurred, the decision was relatively easier. Because the implausibility of the experimental sentences in the figurative conditions required a more accurate analysis (the processor in this condition had to reject the interpretation that was most favored by the context and a possible literal interpretation and decide that the sentence was implausible), responses in the decision task were more accurate than those in the control condition. On the other hand, for the literal context condition more time was required in the decision task, probably because no implausibility had been apparent while reading the sentences. Further analyses were then presumably conducted, requiring longer latencies and a higher probability of error.

Context and Ambiguous Italian Idioms

399

In general, the present data confirm and further support the data found in experiment 2 and in former experiments with the same type of paradigm (Colombo, 1993), showing that the activation of the idiomatic meaning is strongly dependent on context. These results do not fit well, instead, with the idea that idiomatic strings are lexically represented as "long words" (Swinney and Cutler, 1989). If such hypotheses were correct, the presentation of the first part of the idiom should automatically activate the final part, and so relatively faster reading times in the segment corresponding to the second part of the idiom, that is the NP, should be expected. As shown in the off-line task of experiment 1, when the context biased the idiomatic meaning 54% of the completions reported the NP of the critical idioms. Thus, in the on-line task, one would expect faster reading times for the NP portion of the idiom, at least in the figurative context. But there was no such effect in the present data, as reading times were only about 6 ms faster in the experimental conditions than in the control conditions, and the difference is not significant in either case [f(33) < 1 for the difference between experimental and control conditions in both the figurative and the literal context]. One could argue that the idiom's NP was so hard to trigger because the idioms are not familiar idioms, but this is not the case, as shown by the familiarity test. On the other hand, it was remarked above that many continuations of the phrases are possible given the rather common verbs of the idioms, and therefore a strong effect in this direction was probably not to be expected. Evidence by Tabossi (1997, personal communication), can be relevant to this issue. In a recent experiment by Tabossi there were two types of idioms. One type was formed by idioms with an early key, that is, idioms for which the idiomatic completion was very likely given the first NP of the idiom, such as tirare l'acqua al suo mulino ('to pull water to his/her mill'). The second type was formed by late-key idioms, for which the idiomatic completion tended to occur only at the end of the idiom (avere le mani bucate, 'to have the hands pierced'). In this experiment, the target word, on which a lexical decision was made, was embedded in a neutral sentence context, and was a possible literal completion of the first part of the idiom, like vicino in tirare l'acqua al suo VICINO, rather than MULINO, the real idiomatic completion, which was not actually presented, and sporche in avere le mani SPORCHE, rather than BUCATE. So this experiment tested whether there was activation of the idiomatic string at the NP position before the end of the idiom by measuring the amount of interference created by its activation with respect to a control condition. The results showed evidence of interference for the literal target word for the early-key idioms, while there was no difference with respect to the control for late-key idioms. Thus, when predictability is low, the idiom as a unit is not automatically activated. In the present study the proportion of idiomatic completions was rather low in the neutral context (see experiment 1), indicating that these idioms would fall in the "late key" (or "no key"?) category. Thus, the results of Tabossi, showing no

400

Lucia Colombo

activation of the final part of the idiomatic string for late key idioms, are congruent with the on-line data of the present study, which show no effect on the NP position in self-paced reading. With respect to the hypothesis that idioms are represented as long words (that is, as units), one should expect the same pattern of data as for idioms to hold true for "normal" words in spoken-word recognition. For instance, the string 'capt' can give as possible continuations both 'captain' and 'captive,' unless lexical frequency is important, as is predictability for idioms. In the latter case only the more frequent word should show activation, given the segment 'capt-.' This claim is not in agreement with, for instance, the model of spoken-word recognition by Marslen-Wilson and collaborators, in which it is proposed that the initial cohort includes all the units with the same beginning phonemes, nor with the data brought forward by the same author (Marslen-Wilson, 1990, 1993). The issue is, however, still controversial also in this field.

5. IMPLICATIONS FROM THE PRESENT DATA AND CONCLUSIONS The experiments reported in this chapter show that in an on-line task, like selfpaced reading, there are important effects of the sentence context on the comprehension of ambiguous idiomatic expressions. When the idiom can have two plausible interpretations, one literal and one idiomatic, the latter interpretation is computed only when biased by the context. Both off-line and on-line evidence supports this claim. In experiment 1, it was found that the preferred completions depended on the context. When the context was neutral, however, the preferred interpretation was in terms of the literal meaning of the idiom. In experiments 2 and 3 there was evidence of activation of the idiomatic meaning only in the appropriate context. Another aspect that can be considered is related to the so-called key of the idiom, which is assumed to be the point at which the idiom's figurative meaning becomes active according to the configuration model (Cacciari and Tabossi, 1988; Tabossi and Zardon, 1993). Unfortunately, the notion of key is not yet as clearly defined as to allow to predict exactly what and where the key of an idiom should be. In fact, up to now this notion has been defined only in operative terms. However, other evidence by Tabossi (personal communication) using this notion of key is in agreement with the present data. In a cross-modal priming paradigm the two types of idioms, with an early or late key, were presented in neutral sentence contexts. For instance, one of the early-key idioms presented was avere un diavolo per capello, 'to have a devil for each hair,' to be compared to, say, avere la mani bucate, a late-key idiom. For early-key idioms there was facilitation for the target word related to the idiomatic meaning even at the offset of the first NP, although

Context and Ambiguous Italian Idioms

401

smaller with respect to the facilitation found at the end of the idiom. For late-key idioms, there was instead facilitation only at the end of the idiom, when a literal interpretation was no longer possible or likely. It should be pointed out that these types of idioms do not have a literal interpretation. For early-key idioms the literal interpretation becomes implausible after the first NP; for late-key idioms the last word makes the literal interpretation of the idiom implausible. These results are therefore in agreement with the present data; that is, they are compatible with the idea that the figurative meaning only becomes active when the literal interpretation becomes implausible. An interpretation congruent with the overall data, and with other results on ambiguity, can be conceived considering the figurative meaning of ambiguous idioms as a secondary meaning (at least with the present materials), which is therefore activated slower with respect to the dominant meaning and only by a biasing context. In the Rayner and Frazier (1989) study it was found that for ambiguous words with meanings of different frequency, only the dominant meaning was activated when the context was neutral or congruent with it. When the context biased the secondary meaning, or when the two meanings were equiprobable, gaze durations were lengthened. Their results were interpreted as showing that the dominant meaning is activated immediately but not the secondary meaning, whose activation rate is slow and necessitates the activation of a biasing context. Similarly, the present experiments can be interpreted assuming that the literal interpretation is always initiated by default (how can the processor know whether a figurative expression will be encountered?), and the figurative meaning becomes active only when either the context or some other cue can elicit it. The present results are not, however, compatible with those of Titone and Connine (1994), who showed evidence of idiomatic activation even at the penultimate position (NP) before the offset, even for ambiguous idioms, and for lowpredictable ones. There is no obvious way to explain the incompatible results. A first-look interpretation of the incongruent results might attribute the differences to the different languages in which the experiments were performed. Indeed, a recent study has compared the number of ambiguities in English and Italian for different types of ambiguities (Agnoli, Snow, and Hunt, 1997). This study shows that ambiguous words and expressions with a literal and a figurative meaning are much fewer in Italian than in English, as derived from a corpus based on Italian and English newspapers. Could it be that the disparity in the number of lexically ambiguous words and ambiguous idioms (assuming it is definitely proven) can explain the different results? Perhaps this can provide an interesting starting point for further research. The present results are also interesting when compared with the data on lexical ambiguity. As pointed out above, the influential view that had been predominant initially, favoring the idea of automatic activation of both meanings of an ambiguous word, has been followed in many studies and shown that the context, not just

402

Lucia Colombo

dominance of meaning, may have an early influence. The present results, in agreement with the latter view, are potentially relevant to this issue because they show that the meaning associated to a whole phrase, rather than to a single word, is only activated when the context provides a bias for the disambiguation. Initially, however, the processor seems to be oriented to process language in a literal way, which corresponds, in a way, to its dominant mode of sentence interpretation construction.

ACKNOWLEDGMENTS The author wishes to thank Daniela Guasconi for collecting the data of experiments 1 and 3, and Patrizia Tabossi and an anonymous referee for comments and suggestions. Thanks are also due to Sandzo Beltella for technical assistance.

REFERENCES Agnoli, F, Snow, L., and Hunt, E. (1997). A cross-linguistic examination into sources of ambiguity. Unpublished manuscript. Bobrow, S., and Bell, S. (1973). On catching on to idiomatic expressions. Memory and Cognition, 1, 343-346. Cacciari, C., and Tabossi, P. (1988). The comprehension of idioms. Journal of Memory and Language, 27, 668-683. Colombo, L. (1993). The comprehension of ambiguous idioms in context. In C. Cacciari and P. Tabossi (Eds.), Idioms: Processing, structure and interpretation. Hillsdale: Erlbaum. Duffy, S., Morris, R. K., and Rayner, K. (1988). Lexical ambiguity and fixation times in reading. Journal of Memory and Language, 27, 429-446. Fodor, J. (1983). Modularity of mind. Cambridge, Mass: MIT Press. Forster, K. I. (1979). Levels of processing and the structure of the language processor. In W. E. Cooper and E. C. T. Walker (Eds.), Sentence processing: Psycholinguistic studies presented to Merrill Garrett (pp. 27-85). Hillsdale, NJ: Erlbaum. Frazier, L., and Fodor, J. D. (1978). The sausage machine: A new two stage parsing model. Cognition, 6, 291-325. Frazier, L., and Rayner, K. (1987). Resolution of syntactic category ambiguities: Eye movements in parsing lexically ambiguous sentences. Journal of Memory and Language, 26, 505-526. Frazier, L., and Rayner, K. (1990). Taking on semantic commitments: Processing multiple meanings vs. multiple senses. Journal of Memory and Language, 29, 181-200. Gibbs, R. W. (1980). Spilling the beans on understanding and memory for idioms in conversation. Memory and Cognition, 8, 149-156.

Context and Ambiguous Italian Idioms

403

Gibbs, R. W., and Gonzales, G. (1985). Syntactic frozenness in processing and remembering idioms. Cognition, 20, 243-259. Gibbs, R. W., and Nayak, N. (1989). Psycholinguistic studies on the syntactic behavior of idioms. Cognitive Psychology, 21, 100-138. Gibbs, R. W., Nayak, N., and Cutting, C. (1989). How to kick the bucket and not to decompose: Analyzability and idiom processing. Journal of Memory and Language, 28, 576-593. Kellas, G., Paul, S. T., Martin, M., and Simpson, G. B. (1991). Contextual feature activation and meaning access. In G. B. Simpson (Ed.), Understanding word and sentence. Amsterdam: Elsevier. Marslen-Wilson, W. D. (1987). Functional parallelism in spoken word recognition. Cognition, 25, 71-102. Marslen-Wilson, W. D. (1990). Activation, competition and frequency in lexical access. In G. T. M. Altmann (Ed.), Cognitive models of speech processing: Psycholinguistic and computational perspectives (pp. 148-172). Cambridge, MA: MIT Press. Marslen-Wilson, W. D. (1993). Issues of process and representation in lexical access. In G. T. M. Altmann (Ed.), Cognitive models of language processing: The second Sperlonga Meeting (pp. 187-210). Hove: Erlbaum. McGlone, M., Glucksberg, S., and Cacciari, C. (1994). Semantic productivity and idiom comprehension. Discourse processes, 17(2), 167-190. Nunberg, G. (1978). The pragmatics of reference. Bloomington: Indiana University Linguistic Club. Onifer, W, and Swinney, D. A. (1981). Accessing lexical ambiguities during sentence comprehension: Effects of frequency of meaning and contextual bias. Memory and Cognition, 9, 225-236. Paul, S. T., Kellas, G., Martin, M., and Clark, M. B. (1992). Influence of contextual features on the activation of ambiguous word meanings. Journal of Experimental Psychology: Learning, Memory and Cognition, 18(4), 703-717. Rayner, K., and Duffy, S. A. (1986). Lexical complexity and fixation times in reading: Effects of word frequency, verb complexity, and lexical ambiguity. Memory & Cognition, 14, 191-201. Rayner, K., and Frazier, L. (1989). Selection mechanisms in reading lexically ambiguous words. Journal of Experimental Psychology: Learning, Memory and Cognition, 15(5), 779-790. Seidenberg, M. S., Tanenhaus, M. K., Leiman, J. M., and Bienkowski, M. (1982). Automatic access of the meanings of ambiguous words in context: Some limitations of knowledge-based processing. Cognitive Psychology, 14, 489-537. Simpson, G. B. (1981). Meaning dominance and semantic context in the processing of lexical ambiguity. Journal of Verbal Learning and Verbal Behavior, 20, 120-136. Simpson, G. B., and Burgess, C. (1985). Activation and selection processes in the recognition of ambiguous words. Journal of Experimental Psychology: Human Perception and Performance, 11, 28-39. Swinney, D. A. (1979). Lexical access during sentence comprehension: (Re)consideration of context effects. Journal of Verbal Learning and Verbal Behavior, 14, 645-660. Swinney, D., and Cutler, A. (1979). The accessing and processing of idiomatic expressions. Journal of Verbal Learning and Verbal Behavior, 18, 523-534.

404

Lucia Colombo

Tabossi, P. (1988). Accessing lexical ambiguity in different types of sentential contexts. Journal of Memory and Language, 27, 324-340. Tabossi, P., Colombo, L., and Job, R. (1987). Accessing lexical ambiguity: Effects of context and dominance. Psychological Research, 49, 161-167. Tabossi, P., and Zardon, F. (1993). The activation of idiomatic meaning in spoken language comprehension. In C. Cacciari and P. Tabossi (Eds.), Idioms: Processing, structure and interpretation. Hillsdale: Erlbaum. Tabossi, P., and Zardon, F. (1995). The activation of idiomatic meaning. In M. Everaert, E. Van der Linden, A. Schenk, and R. Schreuder, (Eds.), Idioms. Structural and psychological perspectives. Hillsdale, NJ: Erlbaum. Titone, D. A., and Connine, C. M. (1994). Comprehension of idiomatic expressions: Effects of predictability and literality. Journal of Experimental Psychology: Learning, Memory and Cognition, 20(5), 1126-1138. Wasow, T, Sag, I., and Nunberg, G. (1983). Idioms: An interim report. In S. Hattori and K. Inoue (Eds.), Proceedings of the XIIIth International Congress of Linguistics, Tokyo. Zwitserlood, P. (1989). The locus of the effects of sentential-semantic context in spoken word processing. Cognition, 32, 25-64.

COMPACTNESS AND CONCEPTUAL COMPLEXITY OF CONVENTIONALIZED AND CREATIVE METAPHORS IN ITALIAN CRIST1NA CACCIARI Department of Psychology University of Bologna Bologna, Italy

1. INTRODUCTION In many cultures, metaphors are used for describing persons, their appearance, attitudes, and personality. For the Western Apache, for instance, nominal metaphors (i.e., of the form "An A is a B") are a distinct speech genre associated with "wise" adult men and women (Basso, 1976). These "wise words" can be used only to specify one or more behavioral attributes "indicative of undesirable qualities possessed by the referents of the metaphor's constituents" (Basso, 1976, p. 104). In this culture wise words invariably refer to negative attributes; only some among the good talkers of Apache can use them, presumably those acknowledged by the community as "wisdom bearers." As an Apache consultant speaking to the anthropologist Keith Basso pointed out, the capacity to master the Apache language is identified with that of speaking metaphorically. It's too bad that you didn't try to learn about wise words before. When I was young, old people around here used to make them up all the time. Only a few people did it and they were the best talkers of all. . . . Those old people were smart. One of them would make a new one and right away other people would start to use it. ... Only the good talkers can make them up like that. They are

Syntax and Semantics, Volume 31 Sentence Processing: A Crosslinguistic Perspective

405

Copyright © 1998 by Academic Press All rights of reproduction in any form reserved. 0092-4563/98 $25.00

406

Cristina Cacciari the ones who really speak Apache. They are the ones who make up "wise words" and don't have to use someone else's.

In our culture the use of figurative language can sometimes reflect the need to maintain standards of politeness (Brown and Levinson, 1978), especially when a negative appraisal about someone's behavior, attitudes, or personality is intended. Literal language can be far too explicit and "face-threatening"; an ambiguous statement (be it ironic, idiomatic, or metaphorical) can always provide the speaker with an out ("you didn't understand me"). The figurative structure of many insults or euphemisms exemplifies such preference for indirectness when negative comments are involved. Figurative expressions, unlike explicit attributions, can in fact be ignored or misunderstood (cf. also Drew and Holt, 1988; Fussell, 1992; Leech, 1983; Levinson, 1978). Ortony (1980) suggested three different functions played by metaphors: they express things that are literally inexpressible (e.g., sensory experience, cf. Cacciari, in press); they provide a more vivid and image-evoking medium for expressing subjective experiences and emotional states (cf. also Labov, 1984); more relevantly for the purposes of the present study, metaphors represent a compact form of expression for complex ideas because they allow a predication of a bundle of properties in a condensed statement (sometimes a single word). Corresponding literal expressions (if there are any) can be quite prolix by comparison. Ortony explicitly refers to metaphors, but idiomatic expressions also possess this "summarizing property" (Drew and Holt, 1988, 1992). In this chapter, I report the results of a set of experiments concerning the use of conventionalized and innovative metaphors in persons description. Persons description provides a very interesting field for testing the compactness hypothesis proposed by Ortony: We can in fact describe someone using a metaphorical statement instead of a literal one, not so much to save face (a conventionalized metaphor can in fact be as "face-threatening" as a literal expression) but to communicate in a condensed way1 a set of attitudes, values, and appraisal toward him or her. Metaphorical meanings can be positively polysemous but also potentially ambiguous, depending upon the knowledge and beliefs shared by the interlocutors or specified by the linguistic context, the speaker's intentions, and the inferential capacity and knowledge base of the addressee. As Black observed (1979, p. 29), "ambiguity is a necessary by-product of the metaphor suggestiveness."2 What does it mean that a metaphor provides a compact form of predication? No attempt has been made so far to operationalize the notion of compactness proposed by Ortony. One possibility is to examine the number of words necessary to paraphrase a single metaphorical term. If it is true that nonliteral language allows a more compact form of expression than literal language, then more than a single word would be necessary to paraphrase a metaphorical term.

Metaphors in Italian

407

2. METAPHOR COMPACTNESS 2.1. Paraphrase Length In the first experiment, I asked three judges, unaware of the aim of the experiment, to write down as many nonliteral and literal terms commonly used to describe a person they can think of. I obtained 47 terms, 27 nonliteral terms and 20 literal terms (e.g., "fox," "volcano," "butcher," "wall," "benefactor," "optimist") (Tables 1 and 2 contain the list of the Italian terms with an English translation). Two lists were created, one formed by metaphorical descriptors and one by literal descriptors. The metaphorical and literal terms were embedded in sentences as those below (the first was used for metaphorical terms and the second for literal terms): If, metaphorically speaking, I say that a person is a metaphor, literally speaking I mean that she or he is a If speaking of a person I say that a person is a literal term, in other words I mean that she or he is a In order to prevent subjects (100 Italian university students) from interpreting the task as a test of their creative capacities, we asked them to write the shortest paraphrase they could come up with. The results obtained showed that literal terms were paraphrased more often with only one word (mean length of the paraphrase = 1.02) (e.g., "that cheats," "intriguer," or "meddler," for "dishonest") than metaphors (mean = 1.27) (e.g., "that is not capable," "does nothing good," or "acts with imprecise behavior" for "butcher"): only 4 metaphors out of 54 (7.4%) were in fact paraphrased with a single word as opposed to 16 literal terms out of 40 (40%). The meanings of the metaphorical and literal descriptors I used do not refer to a single semantic field. Some express mostly personality traits (e.g., "snake"), some behavioral characteristics (e.g., "clown"), and others refer to physical appearance (e.g., "skeleton") (Fiske and Cox, 1979; Fussell, 1992). Fraser (1979; see also Fussell, 1992) found a high variability in the interpretation that people assign to metaphorical terms when describing persons. In order to examine the interpretations given by the subjects, I asked two independent judges to sort the paraphrases as to how they reflect the following types of interpretations: 1. Personality interpretations (Pers.) that refer to characteristics a person stably possesses (i.e., how she or he is) 2. Behavioral interpretations (Behav.) that refer to observable behavioral patterns (i.e., what she or he does)

TABLE 1 LIST OF ITALIAN CONVENTIONALIZED METAPHORS WITH AN ENGLISH TRANSLATION Metaphors

English translation

Prestigiatore Pallone gonfiato Serpente Vipera Giocoliere Zingaro Bulldozer Terremoto Pagliaccio Filibustiere Macellaio Armadio Elefante Cairo armato Fantasma Frigorifero Bambino Incubo Iceberg Muro Pirata Condanna Camaleonte Bandito Scheletro Maledizione Strega Gazzella Vulcano Calamita Libellula Pantera Treno Perla Fulmine Piuma Usignolo Fiore Poeta Bocciolo Leone Rosa Pioniere Cerbiatto Colosso Stella Tigre Colombo Angela Aquila

Conjurer Puffed up ball Snake Viper Juggler Gipsy Bulldozer Earthquake Clown Filibuster Butcher Cupboard Elephant Tank Ghost Refrigerator Baby Nightmare Iceberg Wall Pirate Sentence Chameleon Bandit Skeleton Curse Witch Gazel Volcan Magnet Dragon-fly Panther Train Pearl Lightning Feather Nightingale Flower Poet Blossom Lion Rose Pioneer Fawn Colossus Star Tiger Dove Angel Eagle 408

TABLE 1 Continued English translation

Metaphors Miniera Roccia Cigno Volpe

Mine Rock Swan Fox TABLE 2

LIST OF ITALIAN LITERAL TERMS AND ENGLISH TRANSLATION Literal terms

English translation

Opportunista Contrabbandiere Misantropo Autoritario Ipocrita Ignorante Delinquente Sfruttatore Pettegolo Fanfarone Falsario Estremista Assassino Bugiardo Disonesto Nemico Persecutore Omicida Malfattore Spregiudicato Dirigente Gentiluomo Amico Erudito Benefattore Temerario Saggio Buontempone Ottimista Esteta Contemplativo Esperto Filantropo Estimatore Perseguitato Ostaggio Vittima Ospite Collezionista Prigioniero

Opportunist Smuggler Misanthrope Authoritative Hypocrite Ignorant Delinquent Exploiter Talker Braggart Forger Extremist Murderer Liar Dishonest Enemy Persecutor Murderer Malefactor Unbiased Manager Gentlemen Friend Erudite Benefactor Reckless Wise Happy Optimist Aesthete Contemplative Expert Philantropist Estimator Persecuted Hostage Victim Host Collector Prisoner

409

410

Cristina Cacciari

3. Physical Appearance interpretations (App.) that refer to a person's appearance (i.e., how she or he appears). Tables 3 and 4 show the percentages in each category. Consistent with the literature, I found a high variability in the interpretations assigned to metaphorical descriptors. Subjects were more consistent in paraphrasing literal terms: on average 23.9 different paraphrases were given for metaphorical terms as opposed to only 11.3 for literal terms. To assess whether such variability in a metaphor's interpretation decreased if one considered the fit between the semantic field to which the term belonged and that expressed in the paraphrase (that is, the extent to which Pers. metaphors elicited Pers. paraphrases, Behav. metaphors induced Behav. paraphrases, and App. metaphors induced App. paraphrases), the metaphorical paraphrases were again sorted by the same two judges. This allowed the assessment of fit between kind of metaphor and kind of paraphrase (the percentages are reported in Table 5). The lack of consistency slightly decreased with respect to the general analysis, but was still quite remarkable: overall, Pers. metaphors elicited 56.9% of Pers. paraphrases, Behav. metaphors elicited 56% of Behav. paraphrases, and App. metaphors elicited 62.9% of App. paraphrases. The possibility exists that such variability might depend on the low level of familiarity of the metaphorical descriptors used. Since no other means for controlling familiarity was available, I checked whether Italian dictionaries mentioned these metaphorical descriptors among the "figurative extensions" of the corresponding literal term. Almost 80% of the metaphorical terms were listed as extensions by Lo Zingarelli (12th ed., 1996), one of the most common Italian dictionaries. Thus, paraphrase variability cannot be due to a lack of familiarity with the metaphorical terms. TABLES PARAPHRASES CODED ACCORDING TO PERSONALITY, BEHAVIOR, AND APPEARANCE CATEGORIES FOR METAPHORS Category (%) Metaphors

Personality

Behavior

Prestigiatore Pallone gonfiato Serpente Vipera Giocoliere Zingaro Bulldozer Terremoto Pagliaccio

85.1 96 28 32 57.1 12 28 20 14.6

14.9 4 68 68 42.8 88 40 78 85.4

Appearance

4

32 2

continues

411

Metaphors in Italian TABLE 3 Continued Category (%) Metaphors Filibustiere Macellaio Armadio Elefante Carro armato Fantasma Frigorifero Bambino Incubo Iceberg Muro Pirata Condanna Camaleonte Bandito Scheletro Maledizione Strega Gazzella Vulcano Calamita Libellula Pantera Treno Perla Fulmine Piuma Usignolo Fiore Poeta Bocciolo Leone Rosa Pioniere Cerbiatto Colosso Stella Tigre Colombo Angela Aquila Miniera Roccia Cigno Volpe

Personality

Behavior

55.1

44.9

48 4

52

10.2

57.1

88 30 74 90 50 36

12 50 26 10 50 64

57.2 14.3 24.5 28.6

42.8 85.7 91.2 75.5 71.4

10.4

89.6

88 6 62 25 18 58

12 94 38 75 4 12

48.9

44.9 95.9

9.8

Appearance

94 32.6

20

100

4.1 6 2 8 81.6 36.7

94

94 2 92 24.5 18.4 63.3

6 4.3

78 30 6.1

96 75.5

95.7

86.9

13.1

44 10 8.2 40

24 10

32 80

61.2

30.6

94 60

54 2 6 28

64.6

35.4

36

60

12.2

14.3

73.5

2

2

91.8

96

6 6.1 12 4

412

Cristina Cacciari TABLE 4 PARAPHRASES CODED ACCORDING TO PERSONALITY AND BEHAVIOR CATEGORIES FOR LITERAL TERMS Category (%) Literal terms

Personality

Behavior

Pettegolo Fanfarone Falsario Estremista Sfruttatore Ipocrita Ignorante Delinquente Opportunista Contrabbandiere Misantropo Autoritario Assassino Bugiardo Disonesto Nemico Persecutore Omicida Malfattore Spregiudicato Benefattore Temerario Saggio Buontempone Vittima Ospite Prigioniero Collezionista Dirigente Gentiluomo Amico Erudito Filantropo Estimatore Perseguitato Ostaggio Ottimista Esteta Contemplativo Esperto

55.6 14.9 97.2 91.7 100 100 2.8 100 100 94.4 88.9 97.2 100 100 100 19.4 100 100 100 83.3 67.5 82.5 92.5 2.5 17.5 7.5 2.5 2.5 7.5 2.5 95 5 5 12.5

44.4 85.1 2.8 8.3

25 27.5 82.5 2.5

97.2

5.6 11.1 2.8

80.6

16.7 32.5 17.5 7.5 97.5 82.5 92.5 97.5 75 92.5 97.5 5 95 95 87.5 100 100 75 72.5 17.5 97.5

Metaphors in Italian

413

TABLE 5 FIT BETWEEN METAPHORS AND PARAPHRASES ACCORDING TO THE CATEGORIES OF PERSONALITY, BEHAVIOR, AND APPEARANCE Paraphrases (%) Metaphors Personality Behavior Appearance

Personality

Behavior

Appearance

56.9 38.0 8.9

38.8 56.0 28.1

4.2 6.0 62.9

A more reasonable explanation might be the polysemous nature of metaphors that might be amplified by the out-of-context condition we used in our study. What is in fact the role of context? It provides the information necessary for selecting the properties appropriate to the discourse topic; therefore context plays a major role in disambiguating the subset of properties intended by the speaker (Black, 1979; Cacciari and Glucksberg, 1994; Gildea and Glucksberg, 1983; Glucksberg, Gildea, and Bookin, 1982; Ortony, 1980; Shinjo and Myers, 1987). Consider a metaphorically intended term such as snake. It provides a set of candidate properties (many and different, as we saw in our data) that can be plausibly attributed to a person. Usually, mutual knowledge among the speakers and/ or the discourse context provide the relevant knowledge about the person to which such a vehicle ("snake") refers so that only a subset of properties will be attributed to him or her. Metaphorically speaking, a person can be a snake in several respects: she or he can be vicious, dangerous, hit you unexpectedly, and so on. Topics such as a lawyer or a supposedly good friend would require, for instance, the instantiation of a different set of attributive properties (or values) insofar as each of these two topics puts constraints on what is more likely to be attributed to them. In my experiments, the descriptors were presented out of context. Hence subjects did not have any reference point (i.e., specific discourse topics) that could constrain the interpretations to be assigned by setting the values of the attributes relevant to the referent. Subjects therefore listed all the candidate properties that one can reasonably predicate by saying that someone is, for instance, a snake. If one thinks of literality and metaphoricity as endpoints of a continuum, then the metaphors we used in this experiment can be located close to idiomatic expressions that are in turn the closest to the literality endpoint among figurative expressions. By asking subjects to paraphrase these conventionalized metaphors, I obtained a fairly wide range of the stereotypical properties associated with animals (e.g., vipers, chameleons), to particular human beings (e.g., pirates, witches), to natural events (e.g., icebergs, earthquakes), and to artifacts (e.g., walls, trains) that in our culture are attributed to persons.

414

Cristina Cacciari

As noted, the rationale behind Experiment 1 was to test the notion of compactness in terms of paraphrase length. The results showed that paraphrasing metaphorical descriptors required more words than paraphrasing literal descriptors. One might be tempted to conclude that indeed metaphors provide a more compact way for describing persons than literal terms. But caution should be exercised before drawing such a conclusion. That literal terms were paraphrased using fewer words might be due to factors other than compactness, for instance, the availability of synonyms. In order to test this possibility, I correlated the number of synonyms listed by an Italian dictionary (Pittano, 1987) for each term with the mean paraphrase length. We obtained a significant negative correlation (r = — .36,p < .02): the more synonyms available, the shorter the paraphrase. This result casts serious doubt on the appropriateness of paraphrase length as a measure of compactness. As an alternative, I hypothesized that the compactness of metaphorical descriptions might be measured in terms of conceptual complexity. By conceptual complexity I mean the number of different properties that a person can associate with a term. Experiment 2 was designed to investigate this hypothesis. 2.2. Conceptual Complexity Forty metaphorical and 40 literal terms were selected from among those used in Experiment 1. Eighteen subjects (who did not participate in Experiment 1) were asked to write down all the properties they could think of for each term. Some of the properties that subjects listed were synonymous or differed in minor respects. As Tourangeau and Rips (1991) noted, treating them as separate items would produce a misleading picture. In order to consider the degree of overlapping in the responses, we counted, as Tourangeau and Rips did, two properties as instances of the same category if they met the following criteria: (a) they were synonymous (e.g., "enormous" and "extremely big"); (b) abbreviations; (c) repetitions; (d) variations of intensity (e.g., "strong" and "very strong"); (f) specifications of the same predicate (e.g., "climbs" and "climbs mountains"); (g) differed in the level of abstraction (e.g., "watch dog" and "dog"). This procedure allowed me to distinguish between types of properties and tokens (i.e., instances of a same type of property). The conceptual complexity of a descriptor was therefore measured in terms of number of different types of properties listed. Significantly more properties3 (e.g., "dangerous," "intelligent," or "liar" for "snake") were listed for metaphorical than literal descriptors (e.g., "scorbutic," "egoist," or "hates women" for "misanthropist") (on average, 2.6 vs. 2.4, respectively4; (t (df 39) = 2.19, p = .03). In the highest frequency distribution range (namely that between 3 and 4 types of properties), we found 22.5% of the metaphorical terms and only 5% of the literal terms. The level of conceptual com-

Metaphors in Italian

415

plexity can hence be considered as a candidate for operationalizing the notion of compactness proposed by Ortony. One might wonder which factors contribute to determining the conceptual complexity of a metaphor used to describe a person. We considered three among such possible factors, namely Familiarity, Expressivity, and Comprehensibility. In a third experiment, we asked 25 subjects to rate each of the 40 metaphorical terms according to the following scales: a. Familiarity (1, not at all familiar, to 7, extremely familiar) b. Expressivity (1, not at all expressive, to 7, extremely expressive) c. Comprehensibility (1, not at all comprehensible, to 7, extremely comprehensible) These three dimensions overall correlated significatively (cf. Marschark, Katz, and Paivio, 1983; Katz et al., 1988). The highest correlation coefficients were obtained for Familiarity with Expressivity (r = .77), Familiarity with Comprehensibility (r = .82), and Expressivity with Comprehensibility (r = .89), all p < .001. The more familiar a metaphor was, the more expressive and comprehensible it was judged. Furthermore, the more expressive a metaphor was, the more comprehensible it was rated. To test whether the conceptual complexity of the metaphorical descriptors was influenced by the respective levels of Familiarity, Expressivity, and Comprehensibility, we split the set of 40 terms into subsets of 20 metaphors each: (a) for Familiarity, the subsets contained the higher rated 20 metaphors (the more familiar subset) and the 20 lower rated metaphors (the less familiar subset); (b) for Comprehensibility, the subsets contained the 20 more comprehensible and the 20 less comprehensible metaphors; and (c) for Expressivity, the subsets contained the 20 more expressive and the 20 less expressive metaphors.5 As to the Familiarity factor, a significantly higher number of properties was given for more familiar metaphors than for less familiar ones (t (df 19) = 3.27, p = .004) (cf. Marschark, Katz, and Paivio, 1983; Katz et al., 1988). A similar picture emerged for Comprehensibility: the more properties listed, the more comprehensible the metaphor (t (df 19) = 2.11, p = .04),6 a result consistent with some existing evidence (e.g., Marschark, Katz, and Paivio, 1983; Katz et al., 1988; Tourangeau and Rips, 1991). The level of Expressivity did not differentiate the conceptual complexity of the metaphorical descriptors (t (df 19) = .80). The extent to which a metaphor was comprehensible and familiar influenced its conceptual complexity as operationalized in terms of properties associated. What about the conceptual complexity of the literal descriptors? We considered the level of frequency of the literal terms. It is a well-known fact that frequent words are the most easily accessible from the mental lexicon, hence it makes sense to suppose that they are also those for which more properties are easily available. A significant correlation was in fact obtained between conceptual complexity and

416

Cristina Cacciari

word frequency.7 Subjects gave more properties for the more frequent among the 40 literal terms employed (r = .36, p = .02).

3. INNOVATIVE METAPHORICAL DESCRIPTIONS Good metaphors do something more than simply call our attention to some preexisting similarities. Metaphors force us to see things in a different perspective and to reconceptualize them accordingly. Such change of knowledge involves the representations of the single entities as well as that of the domains to which the metaphor belongs (cf. Lehrer, 1978; Steinhart and Kittay, 1994; Tourangeau and Sternberg, 1981,1982). This conceptual force has been considered to characterize conventionalized as well as truly innovative metaphors. To Sperber and Wilson (1986), for instance, conventionalized metaphors activate much more than a single extended meaning in the reader, providing access to an "encyclopedic schema" made up of different levels of implications. This is why even conventionalized metaphors cannot be concisely and exhaustively paraphrased.8 Indeed, we found that the interpretations assigned by subjects to standardized metaphors were not as stable and consistent as one might have supposed, given their frequent use in discourse.9 Very litttle is known about the comprehension and appreciation of novel, creative metaphors, with the possible exception of poetic metaphors (cf. Gibbs, 1994). To be able to conclude that metaphors provide a more compact and complex way of describing a person than literal language, one should consider not only conventionalized metaphors but also innovative ones (i.e., those whose interpretation is not yet part of our repertoire of linguistic meanings). For instance, one might wonder whether subjects were as fluent in listing candidate properties for innovative metaphorical descriptions as they were in producing the stereotypical attributes associated with walls, vipers, and earthquakes. The next experiments were devoted to investigating some of the factors potentially affecting the conceptual complexity of creative metaphorical descriptions of persons. In order to investigate the conceptual complexity of nonconventionalized metaphorical descriptions, we first needed a set of suitable materials. Therefore, we asked three colleagues to generate 10 original metaphorical ways to describe a person. The only constraint we put concerned the syntactic structure, as follows: "X is a PP + N" (or "X is a N + PP") (see Table 6 for the Italian descriptions and an English translation). Ten of the 30 descriptions were discarded because no agreement on their meaning was found. This left us with 20 metaphorical descriptions that were grouped into two lists, each formed by 10 descriptions in the nominal form ("is a") and 10 in a simile form ("is like a"). A description hence appeared only once in each list. Forty subjects (university students) participated in

Metaphors in Italian

417

TABLE 6 LIST OF THE ITALIAN CREATIVE METAPHORS WITH THE ENGLISH TRANSLATION Metaphors

English translation

Mare in tempesta Ragnatela awolgente Coltello affilato Cristallo trasparente Fiore appassito Fiume che scorre Filo spezzato Castello incantato Buio senza bagliori Onda in un campo di grano Lento sonno Giardino rigoglioso Falo nella notte Gatto che sonnecchia Specchio incrinato Logo prosciugato Pensiero martellante Vetro appannato Fiume in piena Betulla piegata dal vento

Sea in a tempest Winding spider's web Sharpened knife Transparent cristal Faded flower Flowing river Broken thread Enchanted castle Dark without gleams Wave in a wheat field Slow sleep Luxuriant garden Bonfire in the night Dozing cat Cracked mirror Drained lake Hammering thought Steamed window River in spate Birch tree folded by the wind

this experiment. Their task was to first write down the properties they could associate with each metaphorical description, and then to rate it according to the extent to which the description was as follows: a. Comprehensible (1, not at all comprehensible, to 10, extremely comprehensible) b. Creative (1, not at all creative, to 10, extremely creative) c. Good (1, not at all good, to 10, extremely good) Table 7 presents the mean ratings for nominal metaphors and similes. TABLE 7 MEAN RATINGS FOR THE THREE FACTORS OF COMPREHENSIBILITY, CREATIVITY, AND GOODNESS FOR NOMINAL METAPHORS AND SIMILES

Factors

Nominal metaphors

Similes

Comprehensibility Creativity Goodness

6.2 6.8 5.7

6.6 6.3 5.8

418

Cristina Cacciari

We used both the nominal and simile forms since there are good theoretical reasons to suppose, as we see further on, that they might be used differently by speakers in natural discourse (cf. Glucksberg and Keysar, 1990; Glucksberg and Manfredi, 1995; Johnson, 1996). Furthermore, the differences between metaphors and similes have been at the core of the debate on figurative language since Greek philosophers (Aristotle, to begin with) posited that metaphors were implicit similes. At the same time, as Johnson (1996) noted, there have been surprisingly few studies looking at similes in direct comparison to metaphors. We present first the results relative to nominal metaphors. We observed a significant negative correlation between the mean ratings of Comprehensibility and Creativity (r = —.38, p = .05): less creative metaphors were judged as more comprehensible than more creative ones. In contrast, a positive correlation was obtained between Comprehensibility and Goodness (r = .90, p < .001). The quality of more comprehensible metaphors was more appreciated than that of less comprehensible ones. That good metaphors are those that can easily be understood is confirmed by Tourangeau and Rips (1991; cf. also Marschark, Katz, and Paivio, 1983; Katz et al., 1988) but not by Gerrig and Healy (1983) who, instead, found a dissociation between the quality of the metaphorical sentences and online comprehension, as measured by reading times. Blasko and Connine (1993) found that also the level of perceived familiarity of a nominal metaphor, and not only its aptness, affected the ease with which it was comprehended; more precisely high-familiar metaphors were comprehended more rapidly than low-familiar metaphors that, in turn, were easily understood if highly apt. Now we turn to consider similes: a negative correlation between Comprehensibility and Creativity and a positive correlation between Comprehensibility and Goodness were found for similes as well (r = —.70, p = .001; and r = .94, p < .0001, respectively). But unlike nominal metaphors, similes also produced a significant negative correlation between Creativity and Goodness (r = —.63, p < .005): subjects appreciated most the quality of less creative similes. Whereas no difference emerged between the ratings of Creativity and Goodness assigned to metaphors and similes, metaphorical descriptions were rated as more comprehensible in the simile form than in the nominal form (6.6 vs. 6.2, respectively, t (df 19) = 2.62, p =. 02).10 Why is this so? In order to answer, one should be able to single out the differences between saying, for instance, "X is a cracked mirror" and "X is like a cracked mirror." Do they mean the same? According to Glucksberg and Keysar (1990, cf. also Black, 1979), they do not. The use of the "is a" form alerts a listener that "a specific relation is intended, not simply a general assertion of similarity" (Glucksberg and Keysar, 1990, p. 15). If someone says, "That actor is a real Bela Lugosi," 11 she or he intends to include such an actor in the category of those that are best exemplified by Bela Lugosi, so he "takes on all the properties of the type of actor, [italics added] not of the actor Bela Lugosi himself" (Glucksberg and Keysar, 1990, p. 16). Saying that someone is merely "like" Bela Lugosi is more generic and can, for instance, signal that the

Metaphors in Italian

419

actor overall tries his best to imitate Bela Lugosi's style (cf. Cacciari, 1995). But this leaves unspecified exactly what properties are intended in the comparison.12 As Turner (1991) pointed out, if someone says "She is a witch," we understand that she behaves in a witchy way not that she generically looks like a witch: "to indicate that someone looks like a witch, one must say explicitly 'she looks like a witch'" (Turner, 1991, p. 185). One reason that similes were judged as more comprehensible could be that similes, unlike nominal metaphors, simply assert that some nonliteral and generic similarity exists between a person (the generic X of the items) and, for instance, a cracked mirror. The simile form does not presuppose specific similarities between topic and vehicle that, in order to be successfully computable, would require the availability of a more complex set of information, not the least of which would be a discourse topic. The higher comprehensibility of similes might also be due to the fact that the presence of "like" makes it somewhat easier to identify the metaphor's meaning. If this is the case, and subjects accessed the meaning more easily, then a significant difference should be obtained between the number of properties listed respectively for similes and for nominal metaphors. We therefore employed the same typetoken distinction used for conventionalized metaphors and grouped together those properties that were synonymous or overlapping in some other sense. This allowed us to measure the conceptual complexity of the same nonliteral description when structured in a nominal or in a simile form. The results showed a perfectly overlapping picture: similes and nominal metaphors did not elicit a significantly different number of properties (on average, 1.63 properties for nominal metaphors and 1.7 for similes, t (df 19) = —.94). Thus, that similes were perceived as more comprehensible was not due to a higher availability of properties.

4. CONVENTIONAL AND INNOVATIVE METAPHORS Sperber and Wilson (1986) argued that the difference between a truly creative and a more standardized metaphor lies in the amount of predication (what they call "implicatures") that they respectively allow: "the wider the range of potential implicatures, and the greater the hearer's responsibility for constructing them, the more poetic the effect, the more creative the metaphor" (p. 236). Despite this claim, when we compared the conceptual complexity of the conventionalized metaphors to that of the creative metaphorical descriptions, we found that subjects listed a higher number of properties for conventionalized metaphors (2.6) than for innovative metaphors (1.6).13 Why is this so? That the metaphorical descriptions we used were indeed perceived as original is clear (the mean rating of creativity was 6.8 and 6.3 for the nominal form and the simile form, respectively). The innovative nature of these metaphors, together

420

Cristina Cacciari

with the lack of a specific topic, might have rendered the computation of their meaning more complex than that of conventionalized metaphors. It seems reasonable to suppose that the lack of a specific referent did not affect the comprehension of conventionalized and innovative metaphors to a similar extent. For metaphors whose meaning is already standardized, the problem is the selection of those properties, among all the possible ones, that are pertinent to describe a specific person. For innovative metaphors, readers are faced with the problem of property introduction, to use the distinction between "predicate promotion" and "predicate introduction" proposed by Ortony in 1980 (cf. also the notion of "emergent feature" proposed by Tourangeau and Rips, 1991). The difficulty of assessing which of the many possible concepts associated with a metaphorical vehicle such as cracked mirror, for instance, might be relevant and the lack of any preexisting interpretation might be responsible for the few properties listed, mostly basic semantic features characterizing the lexical units forming the vehicle (e.g., "once luminous," "distorted," "unsure"). But the suggestiveness of metaphorical meanings extends well beyond a simple composition of basic features. The possibility exists, however, that the nature of the task, and not the type of metaphors, might explain the difficulty experienced by subjects in listing properties. To test this hypothesis, we asked another group of 40 subjects (again University students) to simply provide a meaning paraphrase for each description. In order to eliminate any meaning overlap, we applied a procedure similar to that already used for distinguishing between types and tokens of properties. The results showed that, as in the case of properties, the paraphrases reflected only a few and repetitive meaning aspects (e.g., "gives a distorted image," "precarious," or "meaningless" for "X is a cracked mirror"). Assigning a meaning to these descriptions was difficult irrespective of the task: only a few aspects of meaning or a few properties were listed. The possibility exists that these metaphors contrast with the shared frameworks we ordinarily use for describing a person. As Quinn noted, "the invention of new metaphorical expressions is constrained ... by the necessity of finding metaphors that make sense in terms of the cultural models and by preference for metaphors that do a particularly good job of this" (Quinn, 1991, p. 79). In contrast, conventionalized metaphors such as those used in the first part of this study presumably represent what Quinn (1991, p. 79) calls "satisfying instantiations" of culturally shared models. Our results suggest once again the relevance of context for figurative language comprehension. This is particularly true for innovative metaphors: The more innovative or original a metaphorical description is, the more one should know who are we speaking about. In what respect is a person a cracked mirror, a drained lake, or a steamed window? What are the dimensions of attribution that are pertinent to refer to our best friend or to the least friendly colleague? To answer these problems, a revision of the notion of property attribution is needed, particularly for a problem that traditional matching models are unable to resolve, namely the

Metaphors in Italian

421

difference between "mere sharing" and "pertinent sharing" of features (Glucksberg and Keysar, 1990; Tyler, 1978). Glucksberg, Manfredi, and McGlone (1997), in describing their Property Attribution Model, discussed the example of the metaphor vehicle gold mine. This vehicle includes, as a concept, concrete properties (containing a precious metal, being a source of wealth, etc.) that are instantiations of a more general concept, that of richness. This general concept provides different values to any attribute that can vary along the dimension richness or poorness. If we can reasonably use the concept of richness to predicate something sensible about a specific discourse topic, then the concept is instantiated. So, a metaphor vehicle such as gold mine does not provide a specific property, but rather a value of an attribute that ranges along the dimension richness or poorness.14 Unsurprisingly so, different topics instantiate different values: libraries that are gold mines contain large and useful collections of books and manuscripts, while inventions that are gold mines will earn a great deal of money (Glucksberg, Manfredi, and McGlone, 1997). This implies a degree of flexibility in instantiating different attributes depending upon the metaphor topic: the salience and the values of the attributes are then dependent on the concepts highlightened in each metaphor (cf. Murphy, 1996). To go back to our results, if speaking of a person we describe him or her by saying that he or she is a cracked mirror, which concepts would we like to communicate to our reader? Many and different: that of a devastated self, that of an unlucky life, that of someone once happy and luminous and now sad and opaque, and so on and so forth. These interpretations are based on "constellations of expectations" (Quinn, 1991, p. 67) that reflect the cultural models that lie beyond our understanding of metaphors. These metaphors, much as poetic metaphors, require a creative uptake by the reader in that they allow the construction of different scenarios (to use Quinn's words, the "simultaneous apprehension of multiple relations," 1991, p. 77) and therefore a possible range of interpretations. We are not faced with the problem of a value to be given to an attribute than can range, as in the case of gold mine, along different levels of richness, but with a more complex and culturally dependent interpretative task.

5. CONCLUSIONS The experiments described in this chapter raised the question of what constitutes the compactness of metaphorical predication in the domain of person descriptions in Italian. We found that the more promising way to operationalize the notion of compactness is in terms of conceptual complexity: people tend to use fairly conventionalized metaphors for describing the behavior, personality, or appearance of someone because they allow the expression of a bundle of different properties in a condensed, single-word format.

422

Cristina Cacciari

Unsurprisingly, the most familiar metaphorical vehicles (e.g., "viper," "rock") were also considered as the most expressive and comprehensible ones, and those to which subjects attributed the highest expressive quality were also considered to be the easiest to understand. The meanings of original, creative metaphors were difficult to paraphrase and to break up in terms of attributing properties. The less creative a metaphor was judged, the more comprehensible it was perceived. In turn, understanding a metaphor was seen as a prerequisite for appreciating its quality. Our results suggest that similes were rated as more understandable than nominal metaphors (as reflected by a judgment of comprehensibility) because the presence of "like" makes more clear, and at the same time more generic, the nonliteral nature of the description. Specifying the set of more or less stereotypical properties that our culture associates with conventionalized metaphorical vehicles proved to be an easy task, even when no discourse referent was provided to specify the appropriate dimension of attribution. This happened not to be the case for newly created metaphors that behaved much as do poetic metaphors. They required in fact a complex kind of meaning computation that extended well beyond the simple collection of plausible features. Even listing properties can be difficult in this case because these metaphorical vehicles do not promote the salience of properties already potentially existing in the mental representation of how a person can be (either in terms of behavior, personality, or appearance). Describing someone as a cracked mirror, a drained lake, or a steamed window requires us to introduce new properties particularly difficult to specify when, as in the experiments described in this chapter, no referent specification was provided.

ACKNOWLEDGMENTS Portions of this work were presented at the 1993 Annual Meeting of the Psychonomic Society in Washington, USA. I want to thank Sam Glucksberg, Marcia Johnson, Phil Johnson-Laird, Tom Trabasso, Linda Ginzel and Boaz Keysar for their very helpful comments. I am grateful for the financial support provided by the Italian National Council for Research (CNR).

REFERENCES Barsalou, L. W. (1992). Frames, concepts, and conceptual fields. In A. Lehrer & E. F. Kittay (Eds.), Frames, fields and contrasts: New essays in semantic and lexical organization (pp. 21-74). Hillsdale, NJ: Erlbaum. Basso, K. H. (1976). 'Wise Words' of the Western Apache: Metaphor and Semantic Theory. In K. H. Basso (Ed.), Meaning in anthropology (pp. 93-121). Albuquerque: University of New Mexico Press.

Metaphors in Italian

423

Black, M. (1979). More about metaphor. In A. Ortony (Ed.), Metaphor and thought (pp. 19-43). Cambridge: Cambridge University Press. Blasko, D. G., and Connine, C. M. (1993). Effects of familiarity and aptness on metaphor processing. Journal of Experimental Psychology: Learning, Memory and Cognition, 19, 2, 295-308. Brown, P., and Levinson, S. (1978). Universals in Language Usages: Politeness phenomena. In E. Goody (Ed.), Questions and politeness. Strategies of social interaction (pp. 56-310). London: Cambridge University Press. Cacciari, C. (Ed.). (1995). Preface. In C. Cacciari (Ed.), Similarity in language, thought and perception (pp. 7-13). Turnhout (Bruxelles): Brepols. Cacciari, C. (in press). Why do we speak metaphorically? Reflections on the functions of metaphor in discourse and reasoning. In A. Katz (Ed.), Figurative language and figurative thought. Oxford: Oxford University Press. Cacciari, C., and Glucksberg, S. (1994). Understanding figurative language. In M. Gernsbacher (Ed.), Handbook of psycholinguistics (pp. 447-477). New York: Academic Press. Cohen, T. (1979). The cultivation of intimacy. In S. Sacks (Ed.), On metaphor (pp. 1-10). Chicago: University of Chicago Press. Drew, P., and Holt, E. (1988). Complainable matters: The use of idiomatic expressions in making complaints. Social Problems, 35, 4, 98-117. Drew, P., and Holt, E. (1992). Idiomatic expressions and their role in the organization of topic transition in conversation. In Proceedings of Idioms (pp. 171-185). International Conference on Idioms, Tilburg, The Netherlands. Fiske, S. T, and Cox, M. C. (1979). Person concepts: The effects of target familiarity and descriptive purpose on the process of describing others. Journal of Personality, 47,1, 136-161. Eraser, B. (1979). The Interpretation of Novel Metaphors. In A. Ortony (Ed.), Metaphor and thought (pp. 172-185). Cambridge, MA: Cambridge University Press. Fussell, S. (1992). Creating and interpreting personality metaphors. Poster presented at the Annual Meeting of the Eastern Psychological Association, Boston, MA. Gernsbacher, M. A., Keysar, B., and Robertson, R. R. W. (1995). The role of suppression in metaphor interpretation. Paper presented at the 36th annual meeting of the Psychonomic Society, Los Angeles, USA. Gerrig, R. J., and Healy, A. F. (1983). Dual processes in metaphor understanding: comprehension and appreciation. Journal of Experimental Psychology: Memory and Cognition, 9, 667-675. Gibbs, R. W. (1994). The poetics of mind. Figurative thought, language and understanding. Cambridge, MA: Cambridge University Press. Gildea, P., and Glucksberg, S. (1983). On understanding metaphor: The role of context. Journal of Verbal Learning and Verbal Behavior,, 22, 577-590. Glucksberg, S., Gildea, P., and Bookin, M. B. (1982). On understanding non-literal speech: Can people ignore metaphors? Journal of Verbal Learning and Verbal Behavior, 21, 85-98. Glucksberg, S., and Keysar, B. (1990). Understanding metaphorical comparisons: Beyond similarity. Psychological Review, 97, 1, 3-18. Glucksberg, S., and Manfredi, D. (1995). Metaphoric comparisons. In C. Cacciari (Ed.), Similarity in language, thought and perception (pp. 67-81). Turnhout: Brepols.

424

Cristina Cacciari

Glucksberg, S., Manfredi, D., and McGlone, M. (1997). Metaphor comprehension: An interactive property attribution model. Journal of Memory and Language, 36, 1, 50-67. Gregory, M., and Megler, N. (1990). Metaphor comprehension: In search of literal truth, possible sense, and metaphoricity. Metaphor and Symbolic Activity, 5, 151-173. Johnson, A. T. (1996). Comprehension of metaphors and similes: A reaction time study. Metaphor and Symbolic Activity, 11 (2), 125-149. Katz, A., Paivio, A., Marschark, M., and Clark, J. (1988). Norms for 204 literary and 260 nonliterary metaphors on 10 psychological dimensions. Metaphor and Symbolic Activity, 3, 4, 191-214. Labov, W. (1984). Intensity. In D. Schiffrin (Ed.), Meaning: form and use in context: Linguistic applications. Georgetown: Georgetown University Press. Lakoff, G. (1987). Women, fire and dangerous things. Chicago: University of Chicago Press. Leech, G. (1983). Principle of pragmatics. London/New York: Longman. Lehrer, A. (1978). Structures of the lexicon and transfer of meaning. Lingua, 45, 95-123. Levinson, S. C. (1978). Commentary on B. Beck, The metaphor as a mediator between semantic and analogic modes of thought. Current Anthropology, 19, 92. Marschark, M., Katz, A., and Paivio, A. (1983). Dimensions of metaphor. Journal ofPsycholinguistic Research, 12, 17-40. Murphy, G. L. (1996). On metaphoric representation. Cognition, 60, 173-204. Newsome, M. R., and Glucksberg, S. (1996). Do young and older adults suppress metaphor-irrelevant properties during metaphor comprehension? Poster presented at the 37th annual meeting of the Psychonomic Society, Chicago. Ortony, A. (1980). Some psycholinguistic aspects of metaphor. In R. P. Honeck and R. R. Hoffman (Eds.), Cognition and figurative language (pp. 69-83). Hillsdale, NJ: Erlbaum. Pittano, G. (1987). Sinonimi e contrari. Bologna: Zanichelli. Quinn, N. (1991). The cultural basis of metaphor. In J. W. Fernandez (Ed.), Beyond metaphor. The theory of tropes in Anthropology (pp. 56-93). Stanford, CA: Stanford University Press. Shinjo, M., and Myers, J. (1987). The role of context in metaphor comprehension. Journal of Memory and Language, 26, 226-241. Sperber, D., and Wilson, D. (1986). Relevance: Communication and cognition. Oxford: Basic Blackwell. Steinhart, E., and Kittay, E. (1994). Generating metaphors from networks: A formal interpretation of the semantic field theory of metaphor. In J. Hintikka (Ed.), Aspects of metaphor (pp. 41-94). The Netherlands: Kluver Academic Publishers. Tourangeau, R., and Rips, L. (1991). Interpreting and evaluating metaphor. Journal of Memory and Language, 30, 452-472. Tourangeau, R., and Sternberg, R. (1981). Aptness in metaphor. Cognitive Psychology, 13, 27-55. Tourangeau, R., and Sternberg, R. (1982). Understanding and appreciating metaphor. Cognition, 9, 27-55. Turner M. (1991). Reading minds. The study of English in the age of cognitive science. Princeton: Princeton University Press. Tyler, S. A. (1978). The said and the unsaid. New York: Academic Press.

Metaphors in Italian

425

Zingarelli, G. (1996). Lo Zingarelli. Dizionario delta Lingua Italiana, 12ma Edizione. Bologna: Zanichelli.

NOTES 1

They can serve other functions as well, for instance that of creating a sense of ingroupness (i.e., of intimacy, as Cohen, 1979, put it). 2 Not all metaphors are, of course, equal. We are referring to the most felicitous ones. In fact "metaphors are like jokes; good ones can be very successful, but bad ones can be disastrous"(Ortony, 1980, p. 364). 3 For expository reasons, I use the term property to refer to the types of property obtained by eliminating the overlap and not to tokens. 4 These values were obtained by dividing the overall number of types of properties listed for an item by the number of subjects who answered. 5 The six subsets were not formed by the same metaphors since the mean ratings relative to the three factors varied. 6 Marschark, Katz, and Paivio (1983) did not find any correlation between comprehensibility and number of alternative interpretations assigned to a metaphor. Such a discrepancy might be due to differences in the experimental conditions and types of stimuli they used. 7 Word frequency was measured on the basis of the number of occurrences listed in the Frequency Dictionary of Italian edited by IBM and the Istituto di Linguistica Computazionale) 8 See Sperber and Wilson's discussion of "This room is a pigsty"(1986, pp. 235-236). 9 The wide variation of interpretations did not imply that they were inconsistent. For instance, the positive and negative polarity of the metaphorical terms was always maintained: no positive paraphrase for a negative metaphor was ever provided nor was a negative paraphrase ever provided for a positive metaphor (for a similar conclusion see Fraser, 1979; Black, 1979). 10 The existing evidence is fairly controversial and difficult to compare due also to methodological differences. The comprehension difference we found contrasts the results obtained by Glicksberg and Keysar (1990) and Johnson (1996) and is instead partly confirmed by Gregory and Mergler (1990). 11 Bela Lugosi was the actor who played the vampire in many of the Dracula movies in the 1930's and 40's; his portrayal came to epitomize the sinister nature of vampires. 12 The choice between the "is a" and the "is like a" forms produces also an enhancement of some properties and, presumably, a suppression of others. On the role of suppression and enhancement mechanisms in metaphor interpretation, see Gernsbacher, Keysar, and Robertson (1995) and Newsome and Glucksberg (1996). 13 It should be noted that a different number of observations contributed to the two averages: 40 conventional metaphors as opposed to only 20 innovative metaphors. 14 The notion of attribute-value matching of Glucksberg and collaborators bears some interesting similarity to what Barsalou (1992) independently proposed for the organization of knowledge in memory (for a discussion see Cacciari, in press).

This page intentionally left blank

INDEX

A

Access context-free and context-relevant, 4-7 dominant and subordinate meanings, 12-14 form-driven, 253 lexical and coreference processing in Bulgarian, 167-180 and gap filling, 135-137 order, 13-14 Acoustical signal, sentence prosody tied to, 142 Activation idiomatic, 400-401 lexical and reactivation, 136-137 and typicality structure, 138-139 lexical ambiguity interpretations, 175, 177-178 Activation patterns officiality and directedness, 286-287 during sentence processing, 7-8 Adjectival phrase, attachment, 232-236 Adults, sentence comprehension Italian, 360-361 Serbo-Croatian, 354-356 Agreement assigned at local level by children, 371 cues Italian adults, 360-361 in real-time sentence comprehension, 373

main effects, Italian children, 364-365 in Serbo-Croatian, 357-359 subject-verb animacy, in Japanese, 118-122 syntactic or semantic, 113-115 Agreement errors with head-nonhead animacy mismatch, 121-122 pronoun, 122-126 subject-verb, in English, 116-118 Agreement system, impoverished Cantonese, 24 English, 17-18 Ambiguity attachment, 302-305 effect at head noun, 83-88 homophone, resolution by Chinese speakers, 49-50 Italian idioms, 379-402 lexical, see Lexical ambiguity Origo perspectives, 274 relative clause, resolution, 308 thematic, of head noun, 76-79, 81-83 verb, in English, 248-250 Ambiguity bias, pretest, 20, 172-173 Anaphor resolution, in German, 307-308 Animacy effect on error rate, 117-118 errors, English and Japanese, 124-125 main effects Italian adults and children, 360-366 Serbo-Croatian children, 356-358

427

428 Animacy (continued) noun, Chinese, 36-40 subject-verb, agreement in Japanese, 118122 Anomalies, types in sentences, 133 Anthropomorphism, Origos reflecting, 268-273 Apache talkers, 405-406 Argument structure, interaction with prosody, 237-239 Asymmetry, in mismatch effect, 119-124 Attachment adjectival phrase, 232-236 ambiguities, 302-305 dynamics, 240-243 parameterized head, 305-308 preferences, crosslinguistic study, 221-222, 315-318 prepositional phrase, 236-240 prosodic effects, 141 relative clause, 228-232 crosslinguistic variation, 315-318 referential account, 323 Attributes, richness-poorness dimension, 421

B

ba marker, in sentence interpretation, 37-43 Behavioral interpretations, metaphorical paraphrases, 407-414 bei marker, in sentence interpretation, 37-43 Bias corpus, discrepancy with parsing, 328-329 favoring argument interpretation, 235-236 homonym activation independent of, 380 SVO, 355-359, 364 Biasing context in ambiguity resolution, 27 Cantonese cross-modal naming study, 44-46 Cantonese word gating study, 47-49 creating, 173-174 effect on activation of multiple-meaning nouns, 134-135 idioms, 384-392 multiple access in spite of, 4-5 Bilinguals, cognate effects, role of form and meaning, 194-199 Boundaries, syntactic, prosodic effects, 141-142 Bulgarian antecedent reactivation via CMLP task, 162

Index language features, 169 lexical access and processing, 175-178 pretests ambiguity bias, 172-173 object-relative construction, 173-175 reaction times for probes, 173

C

Canonical Origo, 273-274 Cantonese cross-modal lexical ambiguity resolution, 24-27 cross-modal naming, 44-46 word gating, 46-49 Cartesian framework, Wilhelm von Humboldt, 1-2 Case lexical and structural, German, 255 -256 main effects, Serbo-Croatian children, 356358 morphology, Serbo-Croatian and Italian, 349-353 Catalan, and Spanish, bilinguals, 197-199 Categorization, syntactic, 137 Children Italian agreement morphology decoding, 366-370 sentence comprehension, 359-366 preferring local processing strategies, 371372 Serbo-Croatian case and animacy effects, 356-358 sentence comprehension, 353-354 Chinese importance in crosslinguistic studies, 34-35 Mandarin and Cantonese, ambiguity resolution, 18-27 sentence interpretation, 36-40 Choice responses Chinese sentence gating study, 41-43 Chinese sentence interpretation task, 37-40 Cleft sentence, 248-250, 259-261 CMLP, see Cross-modal lexical priming Cognate effect, relation to morphological effect, 199-201 Compactness, metaphor conceptual complexity, 414-416 paraphrase length, 407-414

429

Index Competition and convergence, cues in Chinese, 38-40 NVN and animacy, 362-363 Competition model cue-based interactive, 35 and cue validity, 370-371 word order, 347-348 Completion task, Italian idioms, 385-387 Complexity conceptual, metaphors, 414-416 thematic, German verbs, 257-259 Comprehensibility, metaphor, 415-417 Comprehension ambiguous idioms off-line task, 384-387 on-line tasks, 387-400 figurative language, 381-383 natural speech, 14-17 repair, in Japanese, 57-68 rule-based models, 285 sentence, Italian and Serbo-Croatian, 345374 Conceptual complexity, metaphor compactness, 414-416 Congruence to noncanonical sentence structures, 253-255 postcontextual sentence, 64-66 Constituent ordering German, 297-299 and thematic prominence, 300-302 and time course of phrase structure building, 299-300 Constraint contextual, 132-134, 147 grammatical, in sentence processing, 7-8 pragmatic, on crosslinguistic variation, 215217,222-223 Constraint-based theories, caveat, 331 Constraint satisfaction, 214-215, 256-257 lexicalist accounts, 320-321, 325 Construal hypothesis, 219-222, 229-230, 321323, 326, 329-330 Context independence, sentence processing, 134-143, 147 integration, preposed argument effect, 143144 neutral, literal, and figurative, 385-402 type, interactions with sentence type, 397 Context-dependency hypothesis, 33

Context-dependent models, 12-13 Context-independent models, 13-14 Contextual effects adjectival phrase attachment, 235-236 relative clause attachment, 229-231 Control structures, subject and object, Japanese and English, 99-102, 108 Convergence, and competition, cues in Chinese, 38-40 Coreferencing, reconstruction role, 170-171 Creativity, metaphor, 417-419 Cross-age analysis Italian study of interactions, 365-366 Serbo-Croatian agreement study, 358-359 Crosslinguistic differences, psycholinguistic explanation, 284-287 Crosslinguistic perspective, Origo setting, 280-284 Crosslinguistic research, language development sequences, 345 -346 Crosslinguistic variation pragmatic constraint, 215-217,222-223 psycholinguistic approach, 212-215 relative clause attachment, 315-318 syntactically based parsing strategies, 337-343 Cross-modal lexical priming, 8, 15-17, 134135,145-146, 156-160,171-172, 179180 Cue cost factor, 348-349 Cues acoustic, for prosody, 142 agreement for Italian adults, 360-361 in real-time sentence comprehension, 373 in assignment of sentence roles, 38-40 cross-age analysis in Italian, 365-366 extralinguistic, 233-234 hierarchy, in Bulgarian, 179 Cue validity, and competition model, 347-348, 370-371

D Decisions cross-modal lexical, 138-139, 146 immediacy, in lexical ambiguity resolution, 26-27 structure-based, delay, 85

430 Declarative sentence, processing, and minimal chain principle, 339-342 Deep structure access to, 170-171 different formats, 180 German sentence, 251-252 Delay expression of prosody and punctuation, 242-243 in latencies for figurative context, 388-390 structure-based decisions, 85 in use of agreement morphology, 366, 374 Dependencies discontinuous, 154-155 long-distance, processing in English, 160-162 Dimensional conception, object relations, 273-274 Dimensional prepositions, crosslinguistic differences, 265-287 Directedness, activation pattern, 286-287 Dominance, homophone meaning, Cantonese, 44-46 Dutch attachment preferences, 326-330 interpretations of "in front of" and "behind," 277-284

E Economy, listeners' repair strategies in Japanese, 67-68 Ellipsis, VP, gap filling in, 139-140 Empty category in argument position in Japanese, 73 in sentence processing, 8-9 Empty subject, processing English, 96-97 Japanese, 98-100 English agreement errors pronoun, 122-126 subject-verb, 116-118 ambiguities types, 401 verb, 248-250 attachment preferences, 316-317 crosslinguistic variation, pragmatic constraint, 215-217 genitive expression, 322

Index impoverished agreement system, 17-18 and Japanese, control structures, 100-102 minimal chain principle, 338-339 processing empty subjects, 96-97 long-distance dependencies, 160-162 word order and genitive, 218-222 Errors, agreement, English and Japanese, 116-126 Exhaustive access hypothesis, 33 Expectations, developed by listeners about argument, 239-240 Expressivity, metaphor, 415-416 Eye-tracking experiment German thematic prominence, 301-302 noun phrase attachment, 304

F Facilitation effect cognate words, 196-197 idioms, 383,400-401 morphologically related words, 188-189, 191-194 False friends pairs, 197-199 Familiarity, metaphor, 415-416 Fillers antecedent, 158-162 continued activation, 157 recent and distant, English and Japanese, 96-101 Fit between metaphor and paraphrase, 410 target with sentence, 159-160 Form of language, Humboldtian notion, 2 Forms, and meanings morphological relations reducible to, 186-194 role in bilingual cognate effects, 194-199 Fragments information, integration by Chinese listeners, 43 sentence, experimental manipulations, 220 Frame-by-frame self-paced reading task, 80-83 French adjectival phrase attachment, 232-236 dynamics of attachments, 240-243 interpretations of "in front of" and "behind, 277-284

431

Index prepositional phrase attachment, 236-240 relative clause attachment, 228-232 Frequency distribution, recognition response patterns, 62-63

G Gap in canonical word order, 154-155 relative clause, thematic role, 84-88 Gap filling in complex VP-ellipsis, 139-140 empty categories by listeners, 8 in English and Japanese, 96-101 and lexical access, 135-137 problem, in Bulgarian, 169-178 Garden-path effect English and Japanese, 107 German subordinate clause, 298 at matrix verb, 88-90 Garden-path theory, crosslinguistic variation, 212-213 Gender agreement, in Serbo-Croatian and Italian, 370-371 morphological, in Spanish, 189-190, 200 Genitive Spanish and English, 322 and word order, crosslinguistic study, 218222 German attachment ambiguities, 302-305 constituent ordering, 297-302 discontinuous verb particle combinations, 250-257 interpretations of "in front of" and "behind,' 277-284 minimal chain principle, 340-341 parallel constructions, 296-297 sentence processing, 305-306 sentence structure, 294-296 Goal as controller, 102-105 theta-role, assignment to noun phrase, 105107 Goodness, metaphor, 417-419 Grammatical constraints, sentence processing, 7-8 Grammatical markers, Chinese sentence interpretation task, 36-40

H Head noun ambiguity effect, 83-88 attachment, 305-308 thematic ambiguity, 76-79, 81-83 Homonyms activation independent of bias, 380 interpretation, and revising initial construal, 65-66 in Japanese comprehension repair study, 57-63 Homophones biasing context to, 47-49 density, in Cantonese cross-modal naming study, 44-46

I Idioms access to, 5-7 ambiguous off-line task, 384-387 on-line tasks, 387-400 comprehension, 381-383 early- and late-key, 399-401 Incongruity, recognition, and revising initial construal, 64-66 Information on availability of alternative structures, 221-222 from case-marking particle, 103 contextual, for assigning interpretations in Mandarin, 19, 23 contingent frequency, 214-215 fragment integration by Chinese listeners, 43 postcontextual, inconsistency, 65-66 theta-role, 106 Integration context, 143-145 lexical ambiguities, 4-6 target into sentence, 158-160 Interactionist model, language processing, 23 Interference task, cross-modal, with pseudowords, 16-17 Interpretation compatibility with postcontextual sentence, 58-60 idiomatic, 388-390, 393 "in front of and "behind," 277-284

432 Interpretation (continued) metaphorical paraphrases, 407-414 NVN and SV, adjustment by listeners, 42-43 sentence, in Chinese, 36-40 Intuition, about thematic role of head noun, 77-79 Italian ambiguity types, 401 ambiguous idiom comprehension, 379-402 attachment preferences, 317-318 children, agreement morphology decoding, 366-370 children and adults, sentence comprehension, 359-366 declarative sentence processing, 339-342 grammar overview, 350-353 innovative metaphorical descriptions, 416419 interpretations of "in front of" and "behind," 277-284 metaphors compactness, 407-416 conventional and innovative, 419-421 minimal chain principle, 338-339 rich agreement system, 18 selective access, 27

J

Japanese animacy errors, 125 attachment preferences, 317 complex sentence parsing, reanalysis ambiguity, 71-92 and English, control structures, 100-102 incongruity recognition, 64-66 lexical ambiguity reduction strategies, 56-57 listeners' repair strategies, economy, 67-68 processing of empty subjects, 98-100 repairing comprehension, 57-63 subject-verb animacy agreement, 118-122

K

kana, ambiguous word written in, 64-65 Kanji and Japanese word-processing software, 55-56 target word indication, 57

Index Kanji-encoding test, 61-63 Key, early and late, idioms with, 399-401

L

Language Bulgarian, specific features, 169 figurative, comprehension, 381-383 multistage model of production, 127 prepositional inventory, 280-284 psycholinguistic research, 184 Language-driven hypothesis, ambiguity resolution, 11, 18-19, 23-25 Language processing interactionist model, 23 time course, 15 Language-processing theory, interaction with research methodology, 155-160 Language processor, Humboldtian contributions, 3 Latencies delay, for figurative context, 388-390 naming, in Cantonese study, 45 plausibility response, 395-397 Lexical ambiguity activation of multiple interpretations, 175, 177-178 discourse integration, 4-6 reduction strategies in Japanese, 56-57 resolution context-dependent and -independent models, 12-14 Mandarin Chinese, 19-23 Lexical decision task, Spanish, 190 Lexical items corresponding to morpheme-like units, 202 flagged with special feature, 114-115 Listeners Chinese, integration of fragment information, 43 expectations about argument, 239-240 filling gap in empty categories, 8 German, verb meanings, 260 imagining speaker's point of view, 271-272 interpretations of "in front of" and "behind," 278-280 Italian, morphological cues, 351-353 Japanese, economy of repair strategies, 67-68 meanings, access and integration, 4-5

Index M

Mandarin ambiguity resolution: language-driven hypothesis, 18-24 sentence gating, 40-44 sentence interpretation task, 36-40 Markers ba and bei, canonical NVN orders with, 41-43 dative, 261 grammatical, in Chinese, 36-40 Matrix subject, control of subject gap, 86-87 Matrix verb, garden-path effect at, 88-90 Meanings ambiguous access in Cantonese, 24-25 activation, 175,177-178 assignment to metaphor descriptions, 420 discontinuous verb, 252-253 dominant and subordinate, accessing, 12-14 and forms morphological relations reducible to, 186-194 role in bilingual cognate effects, 194-199 literal and idiomatic, 6-7 primary and secondary, visual targets, 21 specific, homophones in Cantonese word gating, 48-49 Metaphor compactness conceptual complexity, 414-416 paraphrase length, 407-414 conventional and innovative, 419-421 functions, 406 innovative descriptions, 416-419 Minimal chain principle and declarative sentence processing, 339342 parsing, 338-339 Minimal revisions principle, 74-75, 90 Misconstrual, ambiguous word, revising, 64-66 Mismatch effect, 119-122, 124-126 Morphological effect, relation to cognate effect, 199-201 Morphological relationships, reducibility, 186-194 Morphology case, Serbo-Croatian, 349-353 as critical principle of organization, 200-201 role in mental lexicon organization, 185-186

433 Most Recent Filler strategy, 97-100 Multiple token reading, 123,126 N

Naming tasks Bulgarian, CMLP, 167-168 Cantonese, 25-26 cross-modal, 44-46 Nominal metaphors, 418-419 Noun, see also Head noun animacy, in Chinese, 36-40 Noun phrase attachment, in German, 303-305 Goal interpretation, 104-105 idiomatic, 386, 388, 390-396 modification in crosslinguistic study, 216-217 pseudo-grammatical relation, 87 theta-checking strategy, 105-107 Number errors, English and Japanese, 125-126 grammatical and semantic, mismatch, 115 verb and true subject, 114 NVN orders, with ba and bei markers, 41-43 O

Object relations, dimensional, and Origo perspective, 270-275 Object-relative construction, creation and pretest, 173-175 Officiality, activation pattern, 286-287 Off-line task, comprehension of ambiguous idioms, 384-387 On-line tasks comprehension of ambiguous idioms, 387-400 for data interpretation, 145-146 Origo inside perspective, 268-269 outside perspective, 269-275 spatial and relative positions, 267-268 Origo setting, determinants, 275-276 Orthographic effects, and morphological effects, differences, 187-192 P

Parallel constructions, German, 296-297 Parallel structure, favored by subject control of subject gap, 86-87

434 Paraphrase, length, and metaphor compactness, 407-414 Parser, human, models, 107-108 Parsing crosslinguistic variation, 313-332 experience-based mechanism, 213-214 Japanese sentences, reanalysis ambiguity, 71-92 structural ambiguities, 9 syntactic, 168-169, 242-243 syntactically based strategies, 337-343 theta-checking strategy, 105-107 Personality interpretations, metaphorical paraphrases, 407-414 Person marking, by Italian children, 369-370 Perspective, Origo, inside and outside, 268275 Phoneme, identification during sentence processing, 144-145 Phrase structure, building, and constituent ordering, 299-300 Physical appearance, interpretations of metaphorical paraphrases, 410-414 Plausibility, idiomatic sentence, 391-395, 397-398 Polar axis anthropomorphous pattern, 272-273 defining via speaker orientation, 269 Predicate proximity/recency theory, 319-320, 324-325, 331 Predictability, and idiom comprehension, 382-383 Preference in accounting for reanalysis, 74-75 attachment, crosslinguistic study, 221-222, 315-318 for object control, 101-102,109 SR or SOR, and indecisiveness, 85-90 Preference-based theories, 78-79 Prepositional inventory, languages, 280-284 Prepositional phrase attachment, 236-240 spatial, 265-266 Prepositions, dimensional, crosslinguistic differences, 265-287 Priming, see also Cross-modal lexical priming masked, 188,195-197 morphological semantic overlap effect, 192-194 within- and between-language, 202

Index Probes context-appropriate vs. -inappropriate, 133 for meanings of ambiguity, reaction times, 20,22 related and control, reaction times, 173 type, in Cantonese cross-modal naming study, 44-46 typical and atypical, reaction times, 138-139 visual, auditory sentence in relation to, 156157,176 Processing, see also Sentence processing coreference, in Bulgarian, 167-180 empty subject English, 96-97 Japanese, 98-100 long-distance dependencies in English, 160-162 morphological, within word recognition, 184-186 Processing load, measurement at different sentential positions, 80-83 Processing Load Unit, in accounting for reanalysis, 75-76 Pronoun, agreement errors, 122-126 Prosodic effects attachment, 141 syntactic boundaries, 141-142 Prosody interaction with argument structure, 237-239 role in adjectival phrase attachment, 234-235 Pseudowords, cross-modal interference task with, 16-17 Psycholinguistics approach to crosslinguistic variation, 212-215 explanation of crosslinguistic differences, 284-287 research on languages, 184

Q Questions, wh-, structural context effects, 140

R Rapid Serial Visual Presentation, 158-159 Reaction times for probes for meanings of ambiguity, 20,22 produced by typical and atypical probes, 138-139 for related and control probes, 173

Index Reactivation antecedent, 171 filler, 157,161-162 lexical, 136-137 Reading task self-paced crosslinguistic differences, 217 frame-by-frame, 80-83 insensitive to semantic activity, 398 word-by-word, comparison with CMLP, 145-146 Reading time ambiguous head nouns, 83-85 for four sentential positions, 81-83 literal-bias contextual sentences, 396 in postambiguity region, 392-393 whole-sentence, head noun role, 78-79 Reanalysis ambiguity in Japanese, 72-73 thematic information effect, 76-79 theories, 74-76 Recognition test, in Japanese comprehension repair study, 59-60, 62 Reference transfer, 115 Referential account, relative clause attachment, 323 Relative clause attachment, 228-232 crosslinguistic variation, 315-318 referential account, 323 Japanese, processing difficulty, 71-74 pragmatically ambiguous, 306-308 Relative pronoun drop, 326 in Dutch, 329 Relativized Relevance, 321-322 Research methodology, interaction with language-processing theory, 155-160 Response patterns Japanese comprehension repair study, 62-63 supporting context independence, 137 Russian, declarative sentence processing, 340

S Semantic overlap effect on morphological priming, 192-194 between translations, 198-199 Sentence auditory, in relation to visual probe, 156-157, 176

435 cleft, 248-250, 259-261 complex, parsing in Japanese: reanalysis ambiguity, 71-92 content, recognition test, 60-63 idiomatic, plausibility, 391-395, 397-398 object-relative construction, pretest, 173-175 postcontextual compatibility with interpretation, 58-60 recognized as incongruent, 64-66 Sentence comprehension, Italian and SerboCroatian, 345-374 Sentence gating, Mandarin Chinese, 40-44 Sentence interpretation task, Chinese, 36-40 Sentence processing access to idioms, 5-7 crosslinguistic studies, importance of Chinese, 34-35 declarative, and minimal chain principle, 339-342 empty categories, 8-9 German, 305-306 grammatical constraints, 7-8 initial, 146-147 lexical vs. syntactic, 168-169 parsing, 9 phoneme identification during, 144-145 Sentence role, assignment in Chinese, 35 role of cues, 38-40 Sentence structure German, 294-296 noncanonical, congruence to, 253-255 Sentential materials, object-relative construction, 173-176 Serbo-Croatian children case and animacy effects, 356-358 sentence comprehension, 353-354 grammar overview, 349-350 Simile, form of metaphor, 418-419 SOR sentence ambiguous or unambiguous, 81-83 easy, accommodation by preference-based theories, 78-79 garden path severity, 89 Spanish bilingual, cognate effects, 194-199 cognate and morphological effects, 199-201 crosslinguistic variation, pragmatic constraint, 215-217

436 Spanish (continued) genitive expression, 322 and word order, 218-222 morphological relations, reducibility, 186194 reactivation priming for verb, 163 relative clause attachment, 316 Spatial relations dimensional interpretation, 276 prepositional phrases, 265-266 Speech, natural, comprehension, 14-17 Speeded-selection hypothesis, ambiguity resolution, 11, 25-28 SR analysis, chosen by parser at head noun, 89-90 Stereotypical properties, used metaphorically, 413-414,422 Strategies lexical ambiguity reduction, 56-57 Most Recent Filler, 97-100 parsing, syntactically based, 337-343 repair, economy, 67-68 sentence processing, used by Chinese speakers, 50 theta-checking, for noun phrases, 105-107 word order-based, 372-373 Structural context effects VP-ellipses, 139-140 wh-questions, 140 Structural determinism, parser subject to, 75 Structural processing and discontinuous dependencies, 154-155 study, CMLP efficacy, 158-160 Subordinate clause, garden-path, 298 Symbiosis, linguistic variety and theory, 10 Syntactic boundaries, prosodic effects, 141-142 Syntactic categorization, 137 Syntax setting effects, relative clause attachment, 231-232

T

Talkers, Apache, 405-406 Target word, interpretation, Japanese, 59-60, 62-63 Thematic complexity, German verbs, 257-259 Thematic hierarchy, Goal role, 102-105

Index Thematic processor blocking by prosodic breaks, 239 and dynamics of attachment, 241-243 Thematic prominence, and constituent ordering, 300-302 Thematic role, head noun in Japanese, 76-79, 84-85 Theoretical interpretations, relative clause ambiguity resolution, 318-324 Theta-checking strategy, for noun phrases, 105-107 Theta-role German verbs, 257-259 information, 106 Third-person singular, mistakes by Italian children, 368 Top-down prediction, by parser, 107-108 Top-down search, in reanalysis, 79 Translations, cognate and noncognate, 194-200 Transparency, semantic, 193 Tree-lowering mechanism, for reanalysis, 74-75 Tuning, statistical, invoking experience-based parsing, 213-214 Tuning hypothesis crosslinguistic differences in attachment, 319 evaluation, 324 tested in Dutch, 327-328 Turkish basic word order hypothesis, 346-347 preference for SOV in children, 372 Typicality structure assessed by cross-modal lexical decision, 138-139 and context-integration process, 142-143

V

Vehicle, metaphor, 421-422 Verb ambiguities, English, 248-250 discontinuous, German, 250-257 integration, preposed argument effect, 143144 intransitive, Italian, 341-342 Italian paradigms, 367 positioning, German, 295-296 Visual target and normal sentence processing, 156-157

Index position, in lexical ambiguity resolution, 16-17,21 response times, 26-27 von Humboldt, Alexander, linguistic structural diversity, 2-3 von Humboldt, Wilhelm, language as creative act, 1-2

W

Word gating, Cantonese experiment, 46-49 Word order basic, Turkish, 346-347

437 canonical, effect on structural processing, 153-164 in Chinese sentence interpretation task, 36-40 cues, 352-353 and genitive, crosslinguistic study, 218-222 strategies based on, 372-373 strictness, as continuum, 163 SVO bias Italian children, 362-365 Serbo-Croatian, 355-359 Word recognition morphological processing within, 184-186 morphophonological, 3-4

This page intentionally left blank

SYNTAX AND SEMANTICS

Volume 1 edited by John P. Kimball Volume 2 edited by John P. Kimball Volume 3: Speech Acts edited by Peter Cole and Jerry L. Morgan Volume 4 edited by John P. Kimball Volume 5: Japanese Generative Grammar edited by Masayoshi Shibatani Volume 6: The Grammar of Causative Constructions edited by Masayoshi Shibatani Volume 7: Notes from the Linguistic Underground edited by James D. McCawley Volume 8: Grammatical Relations edited by Peter Cole and Jerrold M. Sadock Volume 9: Pragmatics edited by Peter Cole Volume 10: Selections from the Third Groningen Round Table edited by Frank Heny and Helmut S. Schnelle Volume 11: Presupposition edited by Choon-Kyu Oh and David S. Dinneen Volume 12: Discourse and Syntax edited by Talmy Givon Volume 13: Current Approaches to Syntax edited by Edith A. Moravcsik and Jessica R. Wirth

Volume 14: Tense and Aspect edited by Philip J. Tedeschi and Annie Zaenen Volume 15: Studies in Transitivity edited by Paul J. Hopper and Sandra A. Thompson Volume 16: The Syntax of Native American Languages edited by Eung-Do Cook and Donna B. Gerdts Volume 17: Composite Predicates in English Ray Cattell Volume 18: Diachronic Syntax: The Kartvelian Case Alice C. Harris Volume 19: The Syntax of Pronominal Clitics edited by Hagit Borer Volume 20: Discontinuous Constituency edited by Geoffrey J. Huck and Almerindo E. Ojeda Volume 21: Thematic Relations edited by Wendy Wilkins Volume 22: Structure and Case Marking in Japanese Shigeru Miyagawa Volume 23: The Syntax of the Modern Celtic Languages edited by Randall Hendrick Volume 24: Modern Icelandic Syntax edited by Joan Maling and Annie Zaenen Volume 25: Perspectives on Phrase Structure: Heads and Licensing edited by Susan D. Rothstein Volume 26: Syntax and the Lexicon edited by Tim Stowell and Eric Wehrli Volume 27: The Syntactic Structure of Hungarian edited by Ferenc Kiefer and Katalin E. Kiss Volume 28: Small Clauses edited by Anna Cardinaletti and Maria Teresa Guasti Volume 29: The Limits of Syntax edited by Peter Culicover and Louise McNally Volume 30: Complex Predicates in Nonderivational Syntax edited by Erhard Hinrichs, Adreas Kathol, and Tsuneko Nakazawa Volume 31: Sentence Processing: A Crosslinguistic Perspective edited by Dieter Hillert