Complex Predicates in Nonderivational Syntax, Volume 30 (Syntax and Semantics)

SYNTAX AND SEMANTICS VOLUME 30

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

PAULINE JACOBSON Brown University

PETER CULICOVER The Ohio State 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 30 Complex Predicates in Nonderivational Syntax

Edited by Erhard Hinrichs Seminar fur Sprachwissenschaft Universit'at Tubingen D-72074 Tubingen, Germany Andreas Kathol Department of Linguistics University of California, Berkeley Berkeley, California 94720 Tsuneko Nakazawa Language and Information Sciences University of Tokyo Tokyo 153, Japan

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: 0-12-613530-4 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 Preface

ix xi

Two Kinds of Composition in French Complex Predicates

1

ANNE ABEILLE, DANIELE GODARD, AND IVAN A. SAG 1. 2. 3. 4. 5. 6.

Introduction A Lexical Analysis of French Clitics Composition with Tense Auxiliaries Composition FAIRE Reflexive Verbs Conclusion References

Word Order Constraints on Verb Clusters in German and Dutch

1 5 14 17 23 33 39

43

GOSSE BOUMA AND GERTJAN VAN NOORD

1. 2. 3. 4. 5.

Introduction The Constituent Structure of German VPS Word Order within the German Verb Cluster Word Order in Dutch Verb Clusters Comparison and Conclusions References V

43 44 51 63 69 71

vi

Contents

Constraining Argument Merger through Aspect MIRIAM BUTT 1. 2. 3. 4. 5. 6. 7. 8. 9.

Introduction Previous Approaches to Complex Predicates Complex Predicate Formation Urdu Causatives Formal Considerations Causatives, Aspect, and Affectedness An Alternative Analysis of Causatives Passives and Unspecified Objects Conclusion References

Third Construction and VP Extraposition in German: An HPSG Analysis ERHARD W. HINRICHS AND TSUNEKO NAKAZAWA 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.

Introduction Incoherent Infinitive Constructions in German VP Extraposition Third Construction as a Special Case of VP Extraposition Dual Properties of Third Construction Triggering Verbs Argument Composition, ID Schemas, and Forcing Argument Raising in the Third Construction The Third Construction Lexical Rule LP Rules Long Passives A Comparison with Previous Analyses Open Problems Conclusion References

Argument Composition and Long-Distance Scrambling in Korean: An Extension of the Complex Predicate Analysis CHAN CHUNG 1. 2. 3. 4. 5.

Introduction Argument Composition and Auxiliary Verb Construction Long-Distance Scrambling and Argument Composition Theoretical Predictions Summary and Conclusions References

73 73 74 80 85 89 94 103 105 106 109

115 115 116 118 122 129 131 135 138 140 145 148 150 154

159 159 160 177 206 211 218

Contents

Constituency and Linearization of Verbal Complexes

vii

221

ANDREAS KATHOL 1. 2. 3. 4. 5. 6.

Introduction Argument Composition and Branching Structure Constituency Linearization A Radically Linearization-Based Theory? Concluding Remarks References

A Deductive Account of French Object Clitics

221 223 225 233 248 260 267 271

ESTHER KRAAK 1. 2. 3. 4. 5. 6. 7. 8. 9.

Introduction Categorial Logics Data Simple Clitic Constructions Clitic Order Clitics and Verbal Complexes Comparison with Other Analyses Conclusion Appendix: The Framework References

Italian Restricting Verbs: A Lexical Analysis

271 272 285 288 292 296 303 307 308 311 313

PAOLA MONACHESI 1. 2. 3. 4. 5. 6. 7.

Introduction Basic Facts The Syntactic Structure of Restructuring Verbs The Analysis An Alternative Analysis Restructuring Verbs in Salentino Conclusion References

Causatives and the Nature of Argument Structure

313 314 316 320 344 347 360 366 369

GERT WEBELHUTH 1. Introduction 2. The Data from Causatives

369 371

viii 3. Some Analytic Prerequisites 4. The Analysis of Causatives References Index

Contents

385 396 421 423

CONTRIBUTORS

Numbers in parentheses indicate the pages on which the authors' contributions begin.

Anne Abeille (1), TALANA, UFRL, Universite Paris 7, 75251 Paris cedex 05, France Gosse Bouma (43), Alfa-informatica & BCN, Groningen University, NL9700 AS Groningen, The Netherlands Miriam Butt (73), Fachgruppe Sprachwissenschaft, Universitat Konstanz, D-78457 Konstanz, Germany Chan Chung (159), Department of English Language and Literature, Dongseo University, Churye-Dong, Sasang-ku, Pusan 617-716, Korea Daniele Godard (1), CNRS, Silex, Universite Lille 3, BP149, 59653 Villeneuve d'Ascq, France Erhard W. Hinrichs (115), Seminar fur Sprachwissenschaft, EberhardKarls-Universitat Tubingen, D-72074 Tubingen, Germany Andreas Kathol (221), Department of Linguistics, University of California at Berkeley, Berkeley, California 94720 Esther Kraak (271), Utrecht Institute of Linguistics OTS, Utrecht University, Trans 10, 3512 JK Utrecht, The Netherlands Paola Monachesi (313), Utrecht Institute of Linguistics OTS, Utrecht University, Trans 10, 3512 JK Utrecht, The Netherlands Tsuneko Nakazawa (115), Language and Information Sciences, University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo 153, Japan Ivan A. Sag (1), Department of Linguistics, Stanford University, Stanford, California 94305 Gertjan Van Noord (43), Vakgroep Alfa-informatica & BCN, Rijksuniversiteit Groningen, NL-9700 AS Groningen, The Netherlands Gert Webelhuth (369), Department of Linguistics, University of North Carolina at Chapel Hill, Chapel Hill, North Carolina 27599

ix

This page intentionally left blank

PREFACE

Early transformational syntax was to a large degree centered around the notion of kernel sentence, that is, of a syntactic representation that consists of a verb—possibly accompanied by some tense and modal elements—and all and only its arguments at some underlying level. To a great extent, such kernel sentences contained elements that also formed parts of semantic units with prepositional meaning. Phenomena which involve a grouping in the syntax that is incongruent with such semantic predicate-argument relations were explained in terms of transformations that altered the original groupings and rearranged the syntactic configurations at a later stage in the derivation. Even though the term kernel sentence is no longer used, much of the work in current descendants of transformational grammar still implicitly implements the same idea. For instance, the role of Verb Phrase (VP) in much of recent transformational literature has to a great extent been to represent core predicate-argument relations, which also include subjects, at least in some variants (Diesing, 1992). Such representations serve as bases for the assignment of phrasal material to higher phrase-structural positions, mostly to specifier positions of various kinds. Alternatively, heads may move out of their immediate clausal cores to group directly with other heads, either in incorporation structures or in similar amalgams of nonphrasal (X°-level) elements, which we will refer to here as head clusters. The idea of deriving syntactic complexity from alterations of relatively simple representations of core predicate relations is shared by other derivational frameworks such as (early) Relational Grammar. However, there has been mounting evidence that the derivational perspective is inadequate for the description of head cluster phenomena. Mechanisms developed for dislocation of phrasal constituents can often be shown to lack descriptive adequacy when applied to these other construction types. Moreover, monostratal theories do not even avail themselves of different levels of syntactic description and therefore cannot readily implement analyses based on core predicate-argument representations. This has led to xi

xii

Preface

a line of investigation, mostly in nonderivational frameworks (but cf. Haider, 1993, for similar ideas in a transformational theory), that completely eschews the idea of core sentences and builds amalgams of nonphrasal constituents directly in the syntax. In many respects the resulting structures (or subparts thereof) behave in ways similar to lexical predicates, especially with respect to valence properties. As a result, the term complex predicates has emerged as a way to classify this line of inquiry. In the theory of Head-Driven Phrase Structure Grammar (HPSG) (Pollard and Sag, 1994) argument composition has been employed as the central mechanism in the analysis of complex predicates. Argument composition refers to the merging of the valence list of a lexical verbal category with the valence list of its verbal complement. Such an account of complex predicates was first proposed in HPSG by Hinrichs and Nakazawa (1989) (see also Hinrichs and Nakazawa, 1994) to account for the topicalization of verbal constituents and for the Oberfeldumstellung (also known as "auxiliary flip") construction in German. Subsequently, argument composition has been adopted to account for a variety of syntactic constructions in different languages, including clitic climbing in Italian (Monachesi, 1993), syntactic properties of auxiliaries in French (Abeille and Godard, 1994), and cross-serial dependencies in Dutch (Rentier, 1994). Argument composition in HPSG can be regarded as a lexicalized variant of the operation of function composition familiar from Categorial Grammar (Lambek, 1958; Geach, 1972). For the treatment of complex predicates function composition has been proposed by, among others, Moortgat (1984) and Moortgat and Oehrle (1994) to account for verb raising in Dutch and synthetic compounding and by Johnson (1986) to account for verbal clusters in German. Unlike HPSG or Categorial Grammar, the analysis of complex predicates in Lexical-Functional Grammar (LFG) (Kaplan and Bresnan, 1982) has focused on the level of functional structure (f-structure). Butt (1995) argues that cross-linguistically the phrase structure of a complex predicate can be either simple (i.e., monoclausal) or complex (i.e., biclausal), while at the level of f-structure complex predicates give rise to a single nuclear predicate (PRED). The present volume brings together recent research in the area of complex predicates in a variety of different languages (German, Dutch, Italian, French, Korean, and Urdu, among others) undertaken in a variety of nonderivational theories (HPSG, LFG, and Categorial Grammar). The contributions focus on diverse aspects of the syntax and semantics of complex predicate phenomena in these languages, including order variation, constituency relations, interactions with other construction types (cliticization, phrasal complementation), argument relations (causativization, passive), and the syntax-morphology interface, to name a few. Thus

Preface

xiii

this collection provides a representative sample of one of the most active research areas in nonderivational theories of grammar in recent years. The focus of the articles is decidedly on issues of linguistically adequate description; hence this volume should be of primary interest to linguists working on the phenomena mentioned, whether their focus is syntax, semantics, or morphology, regardless of theoretical orientation. But since the analyses are for the most part cast in highly formalized frameworks of grammatical descriptions, we believe that this collection should also be of considerable interest to researchers in neighboring fields such as computational linguistics and natural language processing. We now turn to a more detailed description of the individual contributions. Anne Abeille, Daniele Godard, and Ivan Sag investigate causative constructions in French involving FAIRE, with special attention to the behavior of pronominal clitics of various kinds. The distribution of such clitic elements presents crucial evidence that causative constructions have to be distinguished from other kinds of complex predicates, in particular those involving tense auxiliaries. While both construction types employ argument composition, they differ crucially with respect to whether it is the verbal argument's COMPS list or its ARG-ST list that contributes elements to the ARG-ST list of the complex predicate. The authors argue that causative FAIRE constructions belong to the first class and show how this assumption straightforwardly accounts for numerous differences in the climbing behavior of pronominal clitics. Aside from this novel analysis, which, among other things, is extended into a fine-grained account of subtle differences among French reflexive constructions, the authors also propose a number of innovations in the descriptive apparatus. Thus, while their analysis is crucially driven by distinctions made among different lexical variants of the verbs involved, the different lexical forms are not obtained by lexical rules, as has been the practice in much previous work. Instead, the entire system (apart from a derivational rule for medio-passive reflexives) functions in terms of constraints on lexical types. Gosse Bouma and Gertjan van Noord argue that the word order of German as well as Dutch verb clusters can be accounted for without attributing constituents status to them. Their analysis rests on the assumption that a single HEAD-COMPLEMENT schema exists and licenses phrases consisting of a lexical head and an arbitrary number of its complements. Word order is accounted for by means of linear precedence statements only. A set of such statements that accounts for the full range of ordering possibilities encountered in German and Dutch verb clusters is given. Furthermore, it is shown how the analysis leads to an improved account of partial VP fronting.

xiv

Preface

Miriam Butt provides an LFG analysis of the Urdu causative and permissive, which form a complex predicate in that a biclausal argument structure corresponds to a monoclausal syntactic structure. In addition to the f(unctional)-structure and c(onstituent)-structure, a(rgument)-structure is proposed as an independent level of representation relating grammatical functions to phrase structure representations of sentences. Arguments are "fused" in a-structure and in turn are linked to grammatical functions allowing for an analysis of complex predicates in which more than one argument corresponds to a single grammatical function. Unlike previous proposals, a-structure is viewed as a syntactic level of representation as it captures lexical properties of a predicate that help determine the syntactic realization of the predicate's arguments, i.e., their grammatical functions and case markings. The linking of thematic information to grammatical function is sensitive to aspectual information associated with each argument of a predicate in the a-structure. The alternation in the case marking of the causee, which is often attributed to the "affectedness" of the causee, is accounted for in terms of the aspectual relevance of arguments, thus avoiding the stipulation of semantic notions such as affectedness in the a-structure. Chan Chung's article investigates the syntactic properties of different kinds of finite and nonfinite complementation in Korean. After developing a version of Hinrichs and Nakazawa's argument composition approach for the analysis of Korean verbal complexes, Chung shows how this approach can be extended to longrdistance scrambling constructions. Such apparent displacements of arguments from VPS and finite clauses have traditionally been analyzed in terms of A'-movement (as in Government-BindingTheory (GB)) or as a result of SLASH passing (as in HPSG), but Chung argues that a superior analysis is obtained if instead the complex predicate approach is extended to these cases. Among the main evidence in favor of this proposal, Chung cites certain weak crossover effects and passivization constructions. Erhard Hinrichs and Tsuneko Nakazawa present an analysis of the so-called THIRD CONSTRUCTION in German. This construction type refers to a syntactic configuration in which a subject control verb such as versuchen ('try') splits its VP complement, e.g., das Auto zu reparieren, into discontinuous parts, as in daft Peter das Auto uersucht zu reparieren ('that Peter tries to repair the car'). What makes the THIRD CONSTRUCTION notoriously difficult to analyze is the fact that this construction does not fit neatly in the opposition between two types of infinitive constructions in German that, following Bech (1955), are commonly referred to as coherent and incoherent infinitive constructions. Rather, the THIRD CONSTRUCTION exhibits some of the properties of each of these two classes.

Preface

XV

Hinrichs and Nakazawa present an HPSG analysis of the THIRD CONSTRUCTION which makes use of argument composition, originally proposed for the treatment of German auxiliaries by the same authors (Hinrichs and Nakazawa, 1989). Argument composition allows control verbs to form complex predicates with their governed zw-infinitives; its nominal complements are "raised" to the matrix level, giving rise to the discontinuous word order of the elements of the VP complement. A wide range of related properties of the THIRD CONSTRUCTION and of the coherent-incoherent infinitive constructions are treated by a single lexical rule which alters the subcategorization framework of control verbs. Andreas Kathol's contribution parallels that of Bouma and van Noord in many respects. Like the latter authors, Kathol is concerned with the parameters of order variation seen in the verb cluster of various German and Dutch dialects. And while he also adopts argument composition as the driving force in the construction of head clusters, his approach to discontinuity effects involves not a flattening of the phrase structure—as proposed by Bouma and van Noord—but instead relies on a loosening in the relationship between constituency and linear order. This allows him to retain common assumptions about the constituency relations and to treat the verb cluster as a syntactic constituent. Embedded in the linearizationbased approach of Kathol (1995), the analysis extends naturally to the clause-initial occurrences of finite verbs in root environments in a purely linearly based fashion. In the final section, Kathol compares his approach to that of Reape (1993), which is often considered to be a competitor of the argument composition analysis. A number of shortcomings of Reape's proposals are pointed out which argue that a complex predicate view of German verb clusters is to be preferred over a purely linearization-based analysis. Esther Kraak's article presents a novel account of the distribution and placement of French object clitics in the framework of multimodal categorial grammar logic (Moorgat, 1997). Multimodal categorial grammar combines insights from resource-conscious logics, in particular Linear Logic, whose proof systems require that each assumption can and must be used only once, with categorial grammars in the tradition of the Lambek calculus. The resource-conscious aspects of multimodal categorial grammars is introduced through a set of control operators which can license and constrain the applicability of structural operations. These control operators can be related to one another by a set of inclusion principles, which order the application of different resource management modes, and by interaction principles, which legislate how different modes of combination can be mixed. For the purposes of natural language analysis, the resulting system is flexible in that it makes available different modes of

xvi

Preface

assembling linguistic structure and at the same time provides mechanisms to restrict the interaction between the different structural modes. Kraak shows how such a multimodal setup can be fruitfully applied to the highly complex empirical domain of French object clitics, and she compares her proposal to alternative analyses in categorial grammar (e.g., Nashida 1991) and HPSG (Miller and Sag 1995). Paola Monachesi presents an analysis of restructuring verbs of Italian, a class of verbs which includes modal, aspectual, and motion verbs. Restructuring verbs exhibit uniform morphosyntactic properties with respect to clitic climbing, long NP movement, tough constructions, and auxiliary selection. In particular, clitic climbing has often been used as one of the more reliable diagnostics for complex predicate formation. In this construction, a clitic which originates as a dependent of a complement verb can climb and attach to a restructuring verb. Monachesi shows that the mechanism of argument composition allows a uniform account of the different properties exhibited by restructuring verbs, in particular of citic climbing, long NP movement, and the apparent parallelism between these two phenomena. Furthermore, an explanation is given for the fact that the Italian tough construction, which is strictly bounded, appears to be unbounded in the presence of restructuring verbs. The differences and similarities of monosyllabic clitics and of the clitic loro ('to them') with respect to climbing are derived by assuming that they have a different status. Monosyllabic clitics show affix-like behavior, while loro exhibits word-like behavior. Finally, Monachesi demonstrates that her treatment is flexible enough to deal with dialectal variation. Her analysis is extended to deal with restructuring verbs in Salentino. In this southern Italian dialect, in contrast to Standard Italian, it is possible to have clitic climbing out of a finite clause. Gert Webelhuth's article gives an overview of the types of causative constructions found in the world's languages and illustrates that grammatical theory has to account for at least three different kinds of causatives: (i) those which behave in all respects as if they form a single clause; (ii) those which systematically behave as if they project two clauses; and (iii) those which show mixed behavior of the two previous types. In addition, the paper shows that each of the three types of causatives can be expressed synthetically or analytically, i.e., as one single word or a collection of words. Webelhuth goes on to develop a theory of causatives as a special case of an inventory of universally available lexical predicate constructions. The analysis exploits the lexical type hierarchy to express generalizations across the different causatives and allows individual languages to posit causative predicate types whose morphophonological spell-outs are Ian-

Preface

xvii

guage-particular but whose functional-semantic properties are inherited from a universally available type hierarchy for predicates. The conceptual structure of this analysis of causatives reflects the author's belief that languages are much more similiar in their functional-semantic than in their categorial-phrase structural properties.

ACKNOWLEDGMENTS

The editors thank the authors for submitting their articles and their indices, often under considerable time pressure and tight deadlines, as well as the external reviewers (you know who you are) for their timely and thorough reviews of the articles included in this volume. For proofreading and assistance in the preparation of the index the editors thank Ms. Stephanie Schwarz. We also acknowledge the assistance of the editors at Academic Press in the preparation of this volume and their reminding us and the authors that the value of any collection is intimately tied to its timely publication. Erhard Hinrichs Andreas Kathol Tsuneko Nakazawa

REFERENCES Abeille, A. and D. Godard (1994). The complementation of tense auxiliaries in French. In Proceedings of the Thirteenth West Coast Conference on Formal Linguistics, 157-172. Stanford University. CSLI Publications/SLA, Stanford, CA. Bech, G. (1955). Studien iiber das deutsche Verbum Infinitum. Danske Historiskfilologiske Meddelelser 35, 2. Butt, M. (1995). The Structure of Complex Predicates in Urdu. Dissertations in Linguistics. Stanford University. CSLI Publications, Stanford, CA. Diesing, M. (1992). Indefinites. Linguistic Inquiry Monographs, Vol. 153. MIT Press, Cambridge, MA. Geach, P. (1972). A program for syntax. In D. Davidson and G. Harman (eds.), Semantics of Natural Language, 483-497. Reidel, Dordrecht. Haider, H. (1993). Deutsche Syntax—generativ. Vorstudien zur Theorie einer projektiven Grammatik. Gunter Narr, Tubingen.

xviii

Preface

Hinrichs, E. and T. Nakazawa (1989). Flipped out: AUX in German. In Papers from the 25th Meeting, 193-202, Chicago. Chicago Linguistic Society. Hinrichs, E., and T. Nakazawa (1994). Partial VP and Split NP Topicalization in German: An HPSG Analysis. In E. Hinrichs, D. Meurers, and T. Nakazawa (eds.), Partial-VP and Split-NP Topicalization in German—An HPSG Analysis and Its Implementation. Arbeitspapiere des SFB 340, Vol. 58. Johnson, M. (1986). A GPSG account of VP structure in German. Linguistics 24, 871-882. Kaplan, R. and J. Bresnan (1982). Lexical Functional Grammar: A formal system for grammatical representation. In The Mental Representation of Grammatical Relations, 173-281. MIT Press, Cambridge, MA. (Also appeared in Dalrymple et al. (1995), Formal Issues in Lexical-Functional Grammar, 29-130. CSLI Publications, Stanford, CA.) Kathol, A. (1995). Linearization-Based German Syntax, Dissertation, Ohio State University. Lambek, J. (1958). The mathematics of sentence structure. American Mathematical Monthly 65, 154-169. Miller, P., and I. A. Sag. (1995). Une analyse lexicaliste des affixes pronominaux en francais. Revue Quebecoise de Linguistique 24, 135-171. Monachesi, P. (1993). Object clitics and clitic climbing in Italian HPSG grammar. In Proceedings of the Sixth Conference of the European Chapter of the Association for Computational Linguistics, 437-442, Utrecht. Moorgat, M. (1997). Categorial type logics. In J. van Benthem and A. ter Meulen (eds.), Handbook of Logic and Language. Elsevier, Amsterdam. Moorgat, M. (1984). A Fregean restriction on metarules. In C. Jones and P. Sells (eds.), Proceedings of NELS 14, 306-325. GLSA, Amherst, MA. Moortgat, M., and R. Oehrle (1994). Adjacency, dependency, and order. In P. Dekker and M. Stokhof (eds.), Proceedings of the 9th Amsterdam Colloquium. Nishida, C. 1991. A non-transformational analysis of clitic-climbing in Spanish. In Proceedings of the Tenth West Coast Conference on Formal Linguistics, 395-409, Stanford University. CSLI Publications/SLA, Stanford, CA. Pollard, C. J., and I. A. Sag (1994). Head-Driven Phrase Structure Grammar. CSLI Publications, Stanford, CA, and University of Chicago Press, Chicago. Reape, M. (1993). A Formal Theory of Word Order: A Case Study in West Germanic. Dissertation, University of Edinburgh. Rentier, G. M. (1994). A lexicalist approach to Dutch cross dependencies. In K. Beals, J. Denton, E. Knippen, L. Melnar, H. Suzuki, and E. Zeinfeld (eds.), Papers from the 30th Regional Meeting of the Chicago Linguistic Society, 376-390. Chicago Linguistic Society.

TWO KINDS OF COMPOSITION IN FRENCH COMPLEX PREDICATES ANNE ABEILLE* DANIELE GODARD IVAN A. SAG *TALANA, UFRL Universite Paris 7 75251 Paris cedex 05 France *CNRS Universite Lille 3 59653 Villeneuve d'Ascq France Department of Linguistics Stanford University Stanford, California 94305

1. INTRODUCTION

Romance causative constructions, and French causatives in particular, have well-known properties which set them apart from other kinds of infinitival complementation. Their most distinctive property perhaps is the fact that pronominal clitics interpreted as semantic arguments of the infinitival verbal complement appear attached to the causative verb. This phenomenon is illustrated in (Ib); le, the cliticized counterpart of the NP Proust in (la) and the semantic argument of lire, obligatorily occurs attached to the causative verb. Syntax and Semantics, Volume 30 Complex Predicates in Nonderivational Syntax

1

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

2

Abeille et al.

(1) a. Paul fera lire Proust aux eleves de terminate. Paul make-fut to-read Proust to-the students of senior-year 'Paul will make the senior year students read Proust.' b. Paul le fera lire aux eleves de terminale. Paul it make-fut to-read to-the students of senior-year 'Paul will make the senior year students read it.' c. *Paul fera le lire aux eleves de terminale. Paul make-fut it to-read to-the students of senior-year (same interpretation as (lb)) This situation contrasts with the classical complementation found in (2), where the pronominal clitic cannot attach to the main verb convaincra.1 (2) a. Paul convaincra Marie de lire Proust. Paul convince-fut Marie of to-read Proust 'Paul will convince Marie to read Proust.' b. Paul convaincra Marie de le lire. Paul convince-fut Marie of it to-read 'Paul will convince Marie to read it.' c. *Paul le convaincra Marie de lire. Paul it convince-fut Marie of to-read (same interpretation as (2b)) A similar situation holds for the French tense auxiliaries AVOIR and ETRE, which also exhibit "clitic climbing":2 (3) a. Paul a lu Proust. 'Paul has read Proust.' b. Paul I'a lu. Paul it has read 'Paul has read it.' c. *Paul a le lu. (4) a. Paul est parti a Rome. 'Paul has gone to Rome.' b. Paul y est parti. Paul there is gone 'Paul has gone there.' c. *Paul est y parti. These contrast with other auxiliaries such as aller (near future) or venir (near past) which do not allow clitic climbing in Modern French: (5) a. Paul va lire Proust. Paul is-going to-read Proust

Composition in French CPs

3

b. Paul va le lire. Paul goes it to-read 'Paul is going to read it.' c. *Paul le va lire. (6) a. Paul vient de partir a Rome. Paul comes of to-leave for Rome 'Paul has just left for Rome.' b. Paul vient d'y partir. Paul comes of-there to-leave 'Paul has just left for there.' c. *Paul y vient de partir. Both causative and tense auxiliary constructions have been analyzed as complex predicates in various frameworks. (For early proposals, see Emonds, 1978 (on French auxiliaries), Aissen and Perlmutter, 1983, and Zubizarretta, 1985, inter alia.) However, there are important differences between the two constructions. For example, the causative verb has its own role-assigned subject while the tense auxiliary can be analyzed as a subject-to-subject raising verb. Also, the causative verb introduces a new predication while the tense auxiliary only contributes tense and aspect information to the predicate denoted by the past participle. And, finally, some clitics can stay on the infinitival verb after FAIRE, while no clitics can be realized on the past participle. In this paper, we focus on this third contrasting property. We present an account of this difference between the two kinds of complex predicates in terms of two different ways the two predicates share grammatical information, or two different types of what we will call "argument composition." It has long been known (Kayne, 1975; Hyman and Zimmer, 1976; Rouveret and Vergnaud, 1980) that clitics are not always attached to the causative verb in French. In fact, while they must be so attached in most instances, e.g., (Ib), there are cases where they must attach to the infinitival verb, or else may occur in both positions.3 (7a) and (7b) illustrate the downstairs attachment of idiomatic and reflexive clitics, respectively: (7) a. Cette decision va faire en avoir marre This decision goes to-make of-it to-have enough aux Guyannais. to-the Guyanese 'This decision will make the Guyanese be fed up.' b. C'est une revelation brutale qui a fait It is a revelation sudden that has made

4

Abeille et al.

se convertir Claudel. se to-convert Claudel 'It is a sudden revelation that made Claudel be converted.' In construction with tense auxiliaries, on the other hand, these same clitics must occur attached to the auxiliary verb: (8) a. Les Guyannais en ont eu mane. the Guyanese of-it have had enough 'The Guyanese got fed up.' b. Claudel s'est convent. Claudel se is converted 'Claudel was converted.' We propose a solution to this set of contrasts in terms of two different ways that predicates share their arguments. In the framework of Head Driven Phrase Structure Grammar (Pollard and Sag, 1987, 1994), verbs have an argument structure (ARG-ST) specification: a list of elements (syntactico-semantic complexes called synsems) that are dependent upon the verb. In addition, a verb selects for the complements that it combines with syntactically via the COMPS feature, which is a sublist of the ARG-ST list. A (syntactic) complex predicate arises whenever the ARG-ST list of a verb A includes a predicative word B and B's arguments.4 Given these constructs, our proposal is that there are two ways that a verb can share information with its lexical argument (a participle or infinitive in the two cases we discuss here): the verb can either require that part of its argument structure be identical to that of the lexical argument or else it can require that part of its argument structure be identical to the COMPS list of the lexical argument. Since the combination of a predicate with a lexical argument is like composition of mathematical functions (a functor f1 combines with another functor f2, and the resulting functor f1 ° f2 then combines with the arguments of both f1 and f2), we may refer to analyses such as these as argument composition analyses. Let us further distinguish between the two kinds of argument composition just mentioned as follows: a-composition, if it involves the lexical argument's ARG-ST list; c-composition, if it involves the lexical argument's COMPS list.5 In the case of French, we will argue that the tense auxiliaries combine with their participial argument via a-composition. FAIRE (as well as other causative verbs), by contrast, combine with an infinitival form of a verb via c-composition. Thus, since the COMPS list is the vehicle for selecting overt complements, only complements that have a potential to be realized syntactically can climb in the composition causative construction.

Composition in French CPs

5

We analyze so-called pronominal clitics as inflectional affixes on the verb (Miller and Sag 1997) and propose that argument composition provides the basis for the account of clitic climbing (Abeille and Godard 1994, 1996a, Aissen and Perlmutter 1983, and Monachesi 1995). Indeed, argument composition makes it possible to systematically analyze the cliticized elements as arguments on the verb on which they appear. However, because infinitives bearing clitics are bona fide words in French, the argument composition analysis per se is not sufficient to rule out downstairs cliticization with causatives, even with the distinction between the two kinds of composition. Our analysis must include something more. The extra ingredient we posit is a distinction between two kinds of verb: basic and reduced.6 Intuitively, basic verbs are either those realized without clitics or else "intrinsic" clitic verbs, one of whose arguments must always be realized as a clitic. The basic verbs are thus those that realize SYNTACTICALLY all those arguments that CAN be realized syntactically. By contrast, the reduced verbs are those that realize AFFIXALLY at least one ARG-ST element that could have been realized syntactically. The intuition is that, in the case of most verbs, there is a basic mode of realization (arguments as overt subject or complements) and a secondary mode (one or more arguments is an affix). But because one of the arguments of an intrinsic clitic verb must always be affixal, their argument realization is not reduced or secondary in any way: intrinsic clitic verbs always exhibit a basic valence pattern. This distinction between basic and reduced verbs plays a crucial role in our account of discrepancies between the two kinds of complex predicates: composition causatives allow only basic verbs as their infinitival argument. In section 2, we expand the treatment of Romance pronominal clitics that has been developed in HPSG and distinguish between two types of words that can realize a given verbal lexeme: those with clitic morphology and those without it. In section 3, we analyze tense auxiliaries, providing an account of clitic realization on the auxiliary. In section 4 we discuss the infinitival complements of composition FAIRE, and explain how and why they are different from the complements of tense auxiliaries. This includes a discussion of intrinsic clitics. Section 5 focuses on the grammar of reflexives and leads up to an account of the "clitic-trapping" phenomenon.

2. A LEXICAL ANALYSIS OF FRENCH CLITICS

Before explaining how the two kinds of argument composition can account for clitic climbing with causative and tense auxiliary constructions,

6

Abeille et al.

we briefly summarize the analysis of cliticization developed by Miller and Sag (1997), which, unlike its HPSG predecessors, makes no use of a valence reducing lexical rule. 2.1. French Clitics as Pronominal Affixes Verbal clitics in French have two main properties, which together constitute an apparent paradox. On the one hand, they are affixes, or inflectionally attached affixal pronouns (Miller, 1991; Auger, 1994). This analysis is based on a number of properties: (9) a. high degree of selection with respect to a host b. arbitrary gaps in the set of combinations c. morphophonological idiosyncrasies d. rigid and idiosyncratic ordering e. lexical phonological rules applying to clitic-host combinations f. the impossibility of clitics attaching to a coordination of verbs On the other hand, as we know, clitics do not always occur on the verb which they are an argument of. Some of them must attach to a causative verb (as in (1)), and all of them must attach to a tense auxiliary (as in (3), (4), and (8)), rather than to the participle that selects for them. The intuitive solution to this paradox that we adopt is the argument sharing or "composition" analysis of FAIRE and the tense auxiliaries, according to which arguments of the verbal complement (infinitive or participle) are also arguments of the embedding predicate (FAIRE or the tense auxiliary). Realization of an argument as a pronominal affix on the embedding predicate is thus an alternative to realization as an overt complement. Before presenting the composition analysis, we sketch in broad outline the new treatment of cliticization in French, which builds directly on suggestions made to us by Gosse Bouma.7 2.2. Types of Words The affixal analysis of clitics led to an analysis in terms of a lexical rule (Miller and Sag, 1995; Monachesi, 1995). This LR allowed for an alternation between basic verbal forms (subcategorizing for overt complements) and verbal forms bearing clitics with reduced subcategorization frames. Here we propose a related, but distinct analysis stated in terms of constraints on lexical types. We begin by partitioning the type synsem (the type of the syntacticosemantic complexes that appear on all valence and ARG-ST lists in HPSG) into the various subtypes shown in (10).

Composition in French CPs

7

The most fundamental type of synsem here is canonical). A constraint on signs (i.e., on all words and phrases) requires that they must have canon synsems; thus canon is intuitively the type associated with all overt elements that combine syntactically with heads. This contrasts most directly with the type noncan(onical). And all the noncan synsem types in (10) correspond to arguments that are not realized as overt (sign) expressions, but rather as clitics (the type aff(ixal)) or as gaps (the arguments that are extracted in filler-gap constructions). The type aff is further divided into personal pronominal affixes (p-aff) and anaphoric affixes (a-aff) for reflexives. The choice of synsem type for a subcategorized complement is not in general restricted, i.e., words generally select for arguments without restricting them to being canon or noncan. HPSG is a lexicalist theory, incorporating the principle of Strong Lexicalism (see Scalise, 1984:101ff). We will assume here that each inflected word must belong simultaneously to three compatible types: (1) a (CLITIC-)REALIZATION (REALZN) type—either plain-word (pl-wd) or cliticized-word (cl-wd), with a further distinction between two subtypes of cl-wd: su(bject)-cl-wd and n(on)s(ubject)-cl-wd; (2) an INFLECTIONAL (INFLN) type, e.g., 3sg-presindic-ub, 2sg-imper-ub, etc.; and (3) a LEXEME (lxm) type that specifies the phonological stem, part of speech, argument structure, and meaning common to a family of inflected forms.8 The hierarchically organized verbal LEXEME types correspond to what are normally regarded as lexical entries. The hierarchy of words is thus partially described in terms of the three partitions indicated in (11).

8

Abeille et al.

Following Wechsler (1995) and Davis (1996), much of the information in the lexical description of a LEXEME type—in particular information about the linking of ARG-ST members to semantic roles—is predictable on semantic grounds. A lexemic description thus need include little more than a specification of phonology, grammatical category, and meaning. (12) illustrates the French lexeme LAYER. (Here and throughout, the feature TRANS is used to specify transitivity.)

The inflectional types of French are governed by constraints such as those shown in (13). (F3spi here designates the morphological function that maps each verbal lexeme to its 3sg-pres-indic-ub inflected form.)

The description of the lexeme LAYER in (12) can thus unify with the information in (13) to yield the more complete (but still partial) description in (14).

Composition in French CPs

9

Note that the ARG-ST list in (14) does not specify what subtype of synsem the arguments must be, yet constraints on the agreement properties of the subject argument have been "unified into" the partial word description for lave in (14), which specifies an I-FORM value in addition to the STEM value inherited from the lexeme. We assume further that inflectional information always combines in this way, i.e., monotonically. By contrast, any phenomenon involving a change in the number of arguments in the ARG-ST list will be treated as derivational, rather than inflectional, i.e., in terms of a relation between distinct lexemes. The lexeme description in (12) is quite sparse, making no mention of the valence features SUBJ and COMPS, whose values are systematically related to the ARG-ST list. Nor do the inflectional types fix this relation. Rather, this job is left for the third dimension of lexical structure in French, that of REALIZATION. The first type, pl-wd, is subject to the following constraints:9

The constraint in (15) guarantees that pl-wds have an argument structure list corresponding to the simple concatenation of the values of the valence

10

Abeille et al.

features SUBJ and COMPS; the SUBJ list must furthermore contain exactly one element. Since the SUBJ and COMPS values in general get "cancelled" as a head combines with overt complements and the subject, this constraint has the effect of ensuring that all the arguments of a pl-wd are realized syntactically, rather than affixally.10 As for the type cl-wd, this is subject to the constraints shown in (16).11

(16) first requires that all the members of the COMPS list be of type non-aff; i.e., these complements must either be gaps or else canons (overt complements). (16) also guarantees that SUBJ and COMPS lists add up to be the ARG-ST list, except that one or more ARG-ST elements of type aff must be absent from the SUBJ or COMPS list, i.e., "shuffled in" to constitute the ARG-ST list. Whenever an argument is of type aff and therefore does not belong to either the SUBJ or the COMPS list, the cliticized verb is realized with appropriate pronominal affixation. This effect is obtained via the function F^^, which determines the form of cl-wds only. PPRAF, defined roughly as shown in (17), takes as input an inflected form (as determined by the relevant inflectional type) and adds prefkal or suffkal clitics (or no clitic form in the case of past participles) to determine the FORM value of a given cl-wd:

The definition of AFF-CL given by Miller and Sag (1997) uses a template to fix affix position. The motivation for having FpRAF vacuously applying to past participle forms is discussed in section 3.

Composition in French CPs

11

The two subtypes of cl-type are subject to the following further constraints:

These constraints guarantee that a su-cl-wd (e.g., je-lave or je-le-lave) must have an empty SUBJ list and a first argument that is of type aff (allowing for the possibility that other arguments are also of type aff). A ns-cl-wd (e.g., le-lave), by contrast, must have an element on its SUBJ list (and hence combines with a canonical subject locally or else has an extracted subject) and an aff elsewhere in its ARG-ST list (because the ARG-ST list of all cl-wds must include at least one aff element that is not on any valence list). Finally, notice that the constraints in (15) and (16) make reference to an additional HEAD feature type distinction: every verb that is a pl-wd is specified as basic-verb (bas-vb), whereas the DEFAULT value for cl-wds is reduced-verb (red-vb).12 This distinction is crucial for our treatment of composition FAIRE. In virtue of the fact that it selects an infinitival argument specified as [HEAD bas-ub], it will only allow uncliticized verbs (pl-wds) as its argument, or else cliticized intrinsic clitic verbs (see below). In light of these constraints, a form of LAVER that is both 3sg-pres-indic-ub and pl-wd will obey not only the constraints in (14), but in fact all the constraints shown in (20).

12

Abeille et al.

This word (lexical description) enters into our analysis of sentences such as Jean lave les enfants ('Jean washes the children'). And a su-cl-word of this same inflectional and lexemic type obeys not only the constraint in (16), but also other constraints as in (21).

Here the SUBJ and COMPS values both are the empty list, and hence (21) cannot head a VP with an overt NP complement (complements must correspond to members of the head's COMPS list in HPSG); nor can it combine with an overt subject NP. The presence of the two p-aff elements in the argument structure of the verb triggers the desired morphological realization. There is a clear intuitive sense in which this cl-wd realization of the lexeme LAVER is a "reduced" realization of the lexeme (recall that the lexeme itself is underspecified with respect to syntactic-morphological realization). The situation is different in the case of intrinsic clitic verbs. 2.3. Intrinsic Clitics Intrinsic (idiomatic) clitics represent complements which can only be realized affkally. Some examples of such expressions are (with the intrinsic clitic in boldface): en vouloir a quelqu'un ('have a grudge against someone'), en avoir mane de quelque chose ('to be fed up with something'), en revenir de quelque chose ('to be disappointed with something'), en voir des vertes et des pas mures ('to have hard times'), n'y voir que du feu ('to be completely taken in'), y voir clair ('things are clear'), la bailler belle ('to tell a tall story'), le porter beau ('to look dapper'), se la couler douce ('to have it

Composition in French CPs

13

easy'), se serrer la ceinture ('to tighten one's belt'). Some examples are shown in (39): (22) a. Jean en veut a tout le monde. Jean of-it wants to everybody 'Jean is angry at everybody.' b. Le responsable n'y vena que du feu. the manager in-it see-fut only fire 'The manager will be completely taken in.' Intrinsic clitics by definition never alternate with canonical arguments. The lexeme for an intrinsic clitic verb thus differs from those of other verbs in that it specifies one of its arguments as being of type aff. In addition, these verbs state lexemically that they are of type cl-wd, which has as a consequence not only that realization as a pl-wd is ruled out, but also that intrinsic clitic verbs never allow aff elements on their COMPS list (this is one of the constraints on the type cl-wd given in (16) above). This effect is crucial, as we shall see below. Finally, the HEAD value of an intrinsic clitic verb is lexically specified as bas-vb. This will also play a role in our analysis: it will have the effect of overriding the specification [HEAD red-vb] that is given as a default for cl-wds in (16). For example, en-vouloir a ('be angry with') has the following lexemic description:

Thus the cl-wds from this lexeme, unlike those from most verbal lexemes, are specified as [HEAD bas-vb]. This property, as we will see, will endow such cl-wds with many of the distributional properties of pl-wds, including crucially the possibility of being an argument of composition faire. The lexeme in (23) is also transitive (although this lexeme does not have an

14

Abeille et al.

accusative complement on its ARG-ST list, (see note 19)), a fact that will determine the realization of the causee when an infinitive form of this lexeme is embedded under composition faire (see section 4).

3. COMPOSITION WITH TENSE AUXILIARIES

Abeille and Godard (1994, 1996a) defend an analysis of tense auxiliary constructions in French in terms of a flat structure where the auxiliary is the head and takes as complements the (bare) past participle and the complements of the latter. This analysis is motivated by several phenomena that set tense auxiliaries apart from control or other raising verbs (taking VP complements) in French: Constituency tests (pronominalization, question, cleft, VP preposing/ deletion) Tough constructions Position and scope of manner adverbs Coordination The proposed structure for a sentence like Paul a lave les enfants is the following:

In our present analysis, the auxiliary specifies a sharing of ARG-ST members (rather than COMPS members) with the participle. The use of a-composition rather than c-composition introduces no change in the phrase structure, but it is the key to accounting for the difference between the tense auxiliaries and FAIRE. Since tense auxiliaries involve a-composition, the arguments of the past participle are also arguments of the auxiliary verb. Hence clitic realization on the auxiliary (clitic "climbing") is accounted for straightforwardly in terms of the following (simplified) lexemic descriptions

h

of the tense auxiliaries:

The past participle arguments of tense auxiliaries share all of their arguments with the auxiliary, including the subject ( H ) (which is further constrained by various inflectional types, as we have seen). This analysis gives the right analytic results: the auxiliary "inherits" all the arguments of the participle (this is indicated by the two identities labelled 1 and GO in (25a, b)). In addition, the auxiliary and its participle argument share their semantic content ( m ), to which the auxiliary adds only temporal or aspectual constraints. The feature V-AUX is the vehicle for auxiliary selection. (25a, b) specify that the two tense auxiliaries take past participles bearing the specification [V-AUX avoir] and [V-AUX etre], respectively. By default, past participles are specified as [V-AUX avoir], thus most verbal lexemes give rise to past participles that cannot combine with EIRE. The past participle complements in (25a, b) are further constrained: they must

16

Abeille et al.

be [WEIGHT lite] (hence nonphrasal);13 and they must be [HEAD bas-vb]. This is enough in the general case to ensure that cliticization happens only on the auxiliary. In Jean les a laves ('Jean has washed them'), the form les-a has the following description (ignoring CONTENT):

The past participle is a pl-wd; hence its argument structure is the concatenation of its SUBJ and COMPS list. Furthermore, the only way the participle can possibly be a complement of the auxiliary is for it to be a pl-wd. Indeed, because LAVER is not an intrinsic clitic verb, this is the only way it can be specified as [HEAD bas-ub] (cf. (15)), which is necessary for it to be a complement of the auxiliary. If it were a cl-wd (with the p-aff present in the ARG-ST list but not the COMPS list), then it would be [HEAD red-vb], in accordance with the default constraint on cl-wds. In this case, it could not serve as complement of auxiliary avoir. On the other hand, if one or more arguments of the auxiliary is of type aff, then no finite words formed from the auxiliary can be a pl-wd: if such a word were a pl-wd, then it would also have an affix on its COMPS list and this would prevent the verb from combining with a corresponding overt complement (all overt complements are [SYNSEM canon]—see above). Thus when a past participle is not from an intrinsic clitic verbal lexeme, it must be a pl-wd; hence the clitics must be realized as pronominal affixes on the auxiliary, never on the participle.

Composition in French CPs

17

By contrast, when the participle is from an intrinsic clitic lexeme, it must always be a cl-wd. Yet, because intrinsic clitic verbs are lexemically specified as [HEAD bas-ub], they give rise to participles such as that in (27):

And this is compatible with AVOIR, which will compose all members of the ARG-ST list in (27) including affixes. The participle in (27) will not realize any clitics, however, because FPRAF attaches no affixes to past participles (see the definition of FPRAF in (17)). Once again, the finite auxiliary has to be a cl-wd realizing the clitics that are members of its argument structure via a-composition. The apparent result is the same, e.g., sentences such as (28), but the type assigned to the participle in this case is ns-cl-wd. (28) Jean en a voulu a tout le monde. Jean of-it has wanted to everybody 'Jean was angry at everybody.'

4. COMPOSITION FAIRE

Abeille et al. (1997; to appear), following a suggestion by Hyman and Zimmer (1976), contrast two types of causative structure. The first of these, illustrated in (1) (repeated here as (29)), is a composition structure, where the causative verb is the head of a flat VP and subcategorizes for an NP (the causee), an infinitive and (via composition) the complements of that infinitive.14

18

Abeille et al.

(29) a. Paul fera lire Proust aux eleves de terminate. Paul make-fut to-read Proust to-the students of senior-year 'Paul will make the senior year students read Proust.' b. Paul le fera lire aux eleves de terminate. Paul it make-tut to-read to-the students of senior-year 'Paul will make the senior year students read it.' c. *Paul fera le lire aux eleves de terminale. Paul make-fut it to-read to-the students of senior-year (same interpretation as (lb)) This contrasts with the second structure for causatives, similar to the VP complement structure usually posited for verbs like CONVAINCRE, as in (2) (repeated here as (30)): (30) a. Paul convaincra Marie [de lire Proust]. Paul convince-fut Marie [of to-read Proust] 'Paul will convince Marie to read Proust.' b. Paul convaincra Marie [de le lire]. Paul convince-fut Marie [of it to-read] 'Paul will convince Marie to read it.' c. *Paul le convaincra Marie de lire. Paul it-convince-fut Marie of to-read (same interpretation as (2b)) Here the verb combines with two complements: an NP controller and an infinitival VP. Abeille et al. argue that the double complementation analysis, which is widely accepted for LAISSER and perception verbs, should be extended to FAIRE as well. (See also Tasmowski, 1985, and Miller, 1991.) The VP complementation structure for (control) FAIRE, which is subject to considerable variation across speakers, differs from the composition structure in that the object controller must be pronominal: (31) a. *Paul fera les eleves de terminale [lire Proust]. 'Paul will make the senior year students read Proust.' b. Paul les fera [lire Proust]. Paul them make-fut [to-read Proust] 'Paul will make them read Proust.' The VP complementation structure also differs in freely allowing clitics to be realized on the downstairs verb heading the VP complement: (32) Paul les fera [le lire]. Paul them make-fut [it to-read] 'Paul will make them read it.'

Composition in French CPs

19

On the other hand, cases which we will show below to be unambiguously instances of composition FAIRE, such as (29b,c), only allow upstairs realization of the clitics on the causative verb.

4.1. Composition Causatives and Upstairs Cliticization If the double complementation analysis is correct, as we believe it to be, it changes the way we should look at the cliticization facts in causative constructions. One of our central claims is that composition causatives never allow downstairs cliticization of the sort shown in (32). From this perspective, so-called "clitic climbing" is not optional in French. Rather, there is an alternation between two different structures; and in each structure there is a determinate position for the realization of clitics. There are a number of other properties that distinguish the composition structure from the VP complementation structure. Most notable among these is the well-known fact that the realization of the causee depends on the transitivity of the infinitival verb: it is a dative NP (Np[dat]) if the v is transitive (as in (29) or (33a)); it is an accusative NP (NP[acc]) if the v is intransitive (as in (33b)):15 (33) a. Paul f e r a lire Proust aux/* les eleves de terminale. Paul make-fut to-lire Proust to-the/*the students of senior-year 'Paul will make the senior year students read Proust.' b. Paul fera travailler les/*aux enfants. Paul make-fut to-work the/*to-the children 'Paul will make the children work.' The flat structure analysis, where (an inflected form of) FAIRE, the infinitival and its complements, and the causee are all sisters in a single VP (Morin, 1978; Aissen and Perlmutter, 1983; Legendre, 1987; Miller, 1991; Koenig, 1994), provides the basis for a simple account of these properties (which are surveyed in more detail in Abeille et al., in preparation). In what follows, we will focus on the analysis of the composition causative. In our analysis, both the causee and the complements subcategorized by the infinitival v are complements of FAIRE.16 Composition FAIRE takes as its syntactic arguments (a) the causer, (b) the infinitival v, (c) the causee, and (d) the complements which the infinitival v subcategorizes for.17 Moreover, the causee is coindexed with the (unrealized) subject of the infinitival v, as in ordinary control constructions, and it is underspecified for case. The correct realization of the causee (accusative with intransitive infinitive;

20

Abeille et al.

dative with transitive infinitive) is guaranteed by the existence of the two distinct lexemes illustrated in (34).18

(34 ) a. Composition FAIRE ( intransitive complement):

b. Composition Faire (transitive complement):....................

In both of these patterns for composition FAIRE,19 the argument structure ends in a sublist (tagged El in both (34a, b)) that is identified with the COMPS list of the infinitive argument. Note that this differs crucially from the tense auxiliaries, whose analogous ARG-ST sublist is identified with the ARG-ST of the past participle. (35) and (36) show partial tree representations of the VPS in Paul fait lire Proust awe eleues de terminate ('Paul makes the senior year students read Proust') and Paul fait rire les enfants ('Paul makes the children laugh').20

Composition in French CPs

21

In these structures, the inflected form of FAIRE is a pl-wd—without clitic morphology. Clitic climbing is straightforwardly explained under the composition analysis, since the elements cliticized onto FAIRE are analyzed as arguments of FAIRE. In (37), for example, le-fait is a ns-cl-wd with four elements on its ARG-ST including the p-aff, but only two elements on its COMPS list, as shown in (38). (37) Paul le fait lire aux eleves. 'Paul makes the students read it.'

22

Abeille et al.

The infinitive lire in (37), on the other hand, is a pl-wd, with a p-aff on both its COMPS and ARG-ST lists (arising from unification with the relevant ARG-ST element of le-fait). Cliticization on the infinitival is ruled out in this case: if the complement v were a ns-cl-wd (with an aff element on its ARG-ST but not on its COMPS list), then it would be specified as [HEAD red-ub] (by the default constraint in (16)) and could not be a complement of FAIRE. There are other cases, however, where cliticization is intuitively "basic," and hence is allowed on the infinitive. It is to these that we now turn. 4.2. Composition Causatives and Downstairs Cliticization We can now deal with those clitics that occur on the infinitival v complement of composition FAIRE. These are either intrinsic clitic or reflexive clitic verbs.21 The examples in (39) illustrate the basic fact that intrinsic clitics are realized downstairs in the composition causative construction.22 (39) a. Une telle decision fera en vouloir a such decision make-fut of-them to-want a tout le monde a Jean. to everybody to Jean 'Such a decision will make Jean angry at everybody.' b. Cette manoeuvre a fait n'y voir que du feu this maneuver has made in-it to-see only fire au resonsable. to-the manager 'This maneuver caused the manager to be completely taken in.' c. La crise fait se serrer la ceinture a tout le monde. the crisis makes se-dat to-tighten the belt to everybody 'The crisis is making everybody tighten their belt.' It is clear that we have here instances of composition FAIRE, since the idiomatic verbs are transitive and the causee is realized as an Np[dat]. These facts follow from our analysis. Composition faire, like the tense auxiliaries, selects a bas-vb complement. As we saw earlier, the intrinsic clitic verbs and reflexive verbs are the only cl-wds that are nonetheless specified as [HEAD bas-vb]. Thus any infinitive formed from an intrinsic clitic verb or reflexive verb can function as the complement of faire. The key to explaining why the clitics of intrinsic clitic verbs cannot climb up to attach to composition FAIRE is the assumption that FAIRE involves c-composition. Given that the complements of intrinsic clitic verbs must all be non-aff(the constraint in (16) requires this of all cl-wds), intrinsic clitics

Composition in French CPs

23

can only belong to the ARG-ST list of the infinitive—never to its COMPS list. Since composition FAIRE inherits its complements from the COMPS list of the infinitival, it follows that it can never compose an intrinsic clitic into its argument structure and hence can never realize an intrinsic clitic as a pronominal affix. This is how we block (40): (40) *Qa en f e r a vouloir a tout le monde a Jean. 'That will make everybody mad at Jean.' Recall that this behavior is crucially different from what happens with tense auxiliaries, which involve a-composition (sharing of ARG-ST lists) rather than c-composition. The tense auxiliaries, as we saw in section 3, incorporate intrinsic clitics into their argument structure like any other clitics, and hence may realize them as pronominal affixes. Thus the difference between c-composition and a-composition accounts for the fact that intrinsic clitics climb onto tense auxiliaries, but not onto composition FAIRE.

The behavior of reflexive clitics is similar to that of intrinsic clitic verbs. Their properties are sufficiently complex, however, to warrant separate discussion.

5. REFLEXIVE VERBS

Reflexive clitics are the most discussed subset of clitics occuring downstairs with composition FAIRE. The data have been oversimplified in the literature (e.g., Kayne, 1975; Grimshaw, 1982), which assesses all reflexives as equally possible in composition FAIRE complementation. This is in fact an important area of variation, where speakers have no hesitation about their own judgments and express disbelief when learning that others may have different judgments. Although a complete description of reflexive verbs is outside the scope of this paper, it is nonetheless necessary to go into a bit of relevant background. Let us call reflexive verbs the verbs which host reflexive clitics and which have the form se + verb, where se stands for any reflexive clitic. Following Grevisse (1969:549-555), and leaving aside possible further semantic distinctions, we distinguish between three classes of reflexive verbs: (a) "True" reflexives, where se alternates with an accusative or dative argument (b) "Medio-passives," where se appears to be a grammatical marker, not an argument (c) "Intrinsic" reflexives, which only have a reflexive form

24

Abeille et al.

These are illustrated in (41):23 (41) a. Jean se lave. Jean se washes 'Jean washes himself.' b. Ce livre se vend bien. this book se sells well 'This book sells well.' c. Jean s' evanouit. Jean se faints 'Jean is fainting.' These reflexive clitics can all appear on the infinitives that function as complements of composition FAIRE:24 (42) a. Marie a fait se laver les enfants. Marie has made se to-wash the children 'Marie made the children wash themselves.' b. Le snobisme fait se vendre bien les classiques. the snobbery makes se to-sell well the classics 'Snobbery makes the classics sell well.' c. La chaleur a fait s'evanouir Paul. the heat has made se-to-faint Paul 'The heat made Paul faint.' Morphologically, the status of se is clear: it belongs to the set of verbal clitics, and has its position in the clitic template (Miller and Sag (1997)). However, it is not at all clear whether it should be treated as a pronominal affix or as a "marker"—the morphological reflex of a particular verbal type. In fact there are two families of analyses of Romance reflexives, which both have, as Alsina (1996) observes, certain strengths and weaknesses. In the first type of analysis, reflexive verbs do not differ in any essential way from other verbs that host clitics; the difference lies in the binding properties of the clitic, which must be coindexed with the verb's subject. Under the second approach, the reflexive clitic is simply an indicator of certain specific properties that make reflexive verbs different from other clitic-bearing verbs. It is natural that these two kinds of analysis should have arisen, as both are motivated by a desire to provide a unified analysis of a single morphological entity, in spite of the considerable syntactic and semantic heterogeneity of reflexive verbs. While we acknowledge the diversity in the behavior of reflexive verbs, we maintain that the reflexive morpheme in French corresponds systematically to an element of the verb's ARG-ST. The behavior of certain reflexive verbs in causative and impersonal constructions for many speakers, as well as the selection of auxiliary ETRE,

Composition in French CPs

25

may provide some insight into the recurrent intuition that reflexive clitics are somewhat different from other clitics. Certainly no other clitic causes a verb to change the auxiliary verb it selects (a change that is true for all reflexives in standard French). No other clitic changes argument structure (as medio-passives do: their subject is the complement of the input verb). And no other clitic alters the transitivity of a verb (as do medio-passives and certain cases of true reflexives). Finally, the reflexive clitic does not correspond to a semantic argument, at least in intrinsic reflexives and medio-passives. For all these reasons, it has been called a marker (Grimshaw, 1982; Descles et al., 1986; Koenig, 1994). However, other properties indicate that se corresponds to an element of the ARG-ST, just like other complement clitics. Consider for instance participle agreement. In most cases, the (active) past participle in French agrees in number and gender with a non-canonical accusative complement.25 With most reflexive verbs, it is impossible to decide whether the participle agrees with the subject or the reflexive element. However, dative reflexivization makes it clear that these participles follow the general pattern. Thus (43a), where the accusative reflexive induces participle agreement, contrasts with (43b), which contains a dative reflexive and a non-agreeing participle, and with (44), where the participle agrees with the gap, rather than the subject: (43) a. Marie s'est coupee. Marie-fsg se-acc-is cut-fsg 'Marie cut herself.' b. Marie s'est coupe le doigt. Marie-fsg se-dat-is cut-msg the finger 'Marie cut her finger.' (44) la maison qu'il s'est construite. the house-fsg that he-msg se-dat-is built-fsg 'the house that he built for himself In our analysis, a past participle shows agreement with a noncan accusative element on its ARG-ST list (Abeille and Godard, 1996a; Miller and Sag, 1997). Since the participle of reflexive verbs follows the same pattern, this provides a reasonably clear indication that the ARG-ST contains an element corresponding to the reflexive, just as it does for clitics in general. Here we analyze the reflexive element in terms of an anaphor (an a-aff) bound by the subject of the verb.26 We know that different anaphors may have diverse antecedent conditions within the same language as well as cross-linguistically (see, e.g., Dalrymple 1993). In HPSG, Binding Principles are constraints on the (syntactic) argument structure (Pollard and Sag,

26

Abeille et al.

1992, 1994; Manning, 1994). Within this theory, the obligatory agreement illustrated above indicates that the argument structure of reflexive verbs contains an element coindexed with the subject, which corresponds to the reflexive clitic. We formulate Principle A of Binding theory for French as follows, with a-aff being the synsem type for s(ubject)-anaphors: (45) "Principle A" ("parameterized" for French): A locally (s-)commanded (s-)anaphor must be locally (s-)bound. A type hierarchy which is relevant here concerns the types of content for nominal categories, which includes the content of aff elements. Nonreflexive clitics have personal-pro content, while reflexive clitics have content of type anaphor. One final argument in favor of the claim that reflexive clitic verbs require a uniform analysis in spite of their apparent diversity is the fact that all reflexive verbs in standard French select auxiliary etre, in spite of their differences. We may thus assume the following constraint on past participles:27

Here an ana-list is a list that contains an a-aff element. So what (46) says is that any past participle containing an a-aff element must be specified as [V-AUX etre].28 We now turn to a more detailed consideration of the individual categories: intrinsic reflexives, true reflexives, and medio-passives.29

5.1. Intrinsic Reflexive Verbs Intrinsic reflexive verbs, like the one in (47), do not alternate with nonreflexive verbs, at least synchronically.30 (47) a. Julie s'euanouit souuent. 'Julie often faints.' b. *Julie evanouit (son esprit). 'Julie faints (her mind).' Traditional intrinsic reflexives are intransitives: not only do they cooccur with an accusative causee in causative constructions (as in (42c)), they also

Composition in French CPs

27

alternate with an impersonal verb: (48) a. Plusieurs personnes s''evanouissaient sous several persons se-faint-imperf-3pl under l'effet de la chaleur. the-effect of the heat 'Several people fainted because of the heat.' b. Ils' evanouissait plusieurs personnes sous it se-faint-imperf-3sg several persons under I'effet de la chaleur. the-effect of the heat 'There were several people fainting because of the heat.' Our analysis of verbs of this category is the same as our treatment of intrinsic clitic verbs, except that the [V-AUX etre] constraint applies to all participles formed from this lexeme:

Like other intrinsic clitic verbs, intrinsic reflexive verbs can be complements of composition FAIRE because they are lexemically specified as [HEAD bas-vb], and this value is inherited by their cliticized verbal forms. They do not allow clitic climbing with composition FAIRE because their COMPS list disallows aff-type elements (again because of the constraints on cl-wd in (16) above). The intrinsic clitic se thus cannot be attached to composition FAIRE, because it cannot be part of the COMPS list of the infinitive and hence cannot be inherited as an argument of FAIRE. 5.2. True Reflexives With true reflexive verbs, the clitic alternates with an argument, either an accusative or dative NP complement, in the same way as pronominal nonreflexive clitics do so. True reflexive verbs can occur as complements of FAIRE only for certain speakers (who we will refer to as "A-speakers"). Thus, there are two descriptions of true reflexive verbs: one for A-speakers, which treats them as bas-vb; and one for B-speakers, which analyzes them as red-vb (like other cliticized verbs).

28

Abeille et al.

Reflexive verbs are often characterized as intransitive in the literature. However, whereas medio-passives are always intransitive (see below), true reflexives are somewhat more complicated, as we will explain. We make use of two main criteria for transitivity: impersonal and causative constructions. If a given verb alternates with an impersonal verb and makes the causee appear as an accusative NP (in the composition causative), then that verb is intransitive. For A-speakers, who accept them as complements of FAIRE, true reflexives with an accusative affix are intransitive, since they cooccur with an unmarked causee (as in (42a)) and alternate with an impersonal verb:31 (50) a. Quelqu'un se lavait dans la fontaine. someone se wash-imperf-3sg in the fountain 'Someone was washing himself in the fountain.' b. Il se lavait quelqu'un dans la fontaine. it se wash-imperf-3sg someone in the fountain 'There was someone washing himself in the fountain.' However, the transitivity of true dative reflexives is independently determined. These verbs have the same transitivity as their nonreflexive counterparts. Thus the verb S'OFFRIR is transitive because OFFRIR is: (51) a. Certains journalistes s'offriront un exemplaire some journalists se-give-fut-3pl a copy des Memoires of-the Memoirs 'Some journalists will buy a copy of the Memoirs for themselves.' b. *Il s'offrira un exemplaire des Memoires it se-give-fut-3sg a copy of-the Memoirs certains journalistes. some journalists 'There will be some journalists who buy a copy of the Memoirs for themselves.' c. Lew profession fera s'offrir un exemplaire their job make-fut se-give a copy des Memoires a certains journalistes. of-the Memoirs to some journalists 'Their job will make some journalists buy a copy of the memoirs for themselves.' By contrast, for B-speakers, who do not accept true reflexive verbs under composition FAIRE, the impersonal usage in (50b) is also unacceptable. For A-speakers, then, we define a lexical rule that derives intransitive reflexive lexemes like (52) (with an accusative a-aff) from the corresponding basic transitive lexeme.32 The result is lexemes like the one shown

Composition in French CPs

29

in (52):

Furthermore, A-speakers have the constraint (53) in their grammar, which makes reflexives of the type bas-ub, and allows them to appear as complements of composition FAIRE.

Thus for the A variety, all reflexive verbs override the red-vb default for cl-wds and hence all reflexive verbs may appear in the composition causative.33 In addition, binding theory, which all words must obey, then guarantees that the two indices i and j in (52) must be identical. Thus an inflected form like se-laver inherits all the information in (52) and hence is a possible complement of composition FAIRE for A-speakers. B-speakers lack this rule; for them, se-laver must obey the default [HEAD red-vb] specification. Hence in this variety, there is no possibility of embedding true reflexives under composition FAIRE. For dative reflexives, where no change in transitivity is involved in any variety of French we are familiar with, there is no need to posit a lexical rule. Their behavior in all varieties is correctly predicted in the minimally different grammatical systems just discussed. To conclude our discussion of true reflexives and composition FAIRE, notice that the a-aff element is a member of the infinitive's ARG-ST list, but not its COMPS list. Therefore, it is not part of the ARG-ST of FAIRE. This means that true reflexives behave like the intrinsic reflexives: they attach

30

Abeille et al.

to tense auxiliaries (under a-composition), but never to FAIRE (under c-composition). Our analysis also has consequences for binding theory. The lexical description of composition FAIRE in (34a) above requires that the unrealized subject of the v[inf] is identified with the object of faire, (as in raising constructions). Thus, the argument structure of se laver in (42a), repeated here in (54a), where "i" is the index of the direct object of fait, is as shown in (54).

And when an Np[dat] causee (indexed as "k") cooccurs with a dative reflexive, the argument structure of a verb like s'acheter in (55a) is (55b):

Our binding principles thus correctly guarantee that all such examples receive the correct semantic interpretation.

5.3. Medio-Passives Medio-passives appear under a variety of names in the literature, (middle-se, inchoative reflexives, ergatives, neuter) some of which pertain to only a subset of the phenomena we treat here. As a class, medio-passive verbs alternate with transitive verbs, whose direct object NP corresponds to the subject of the medio-passive.34 The acceptability of medio-passives in the impersonal construction and the fact that they occur under FAIRE with an accusative causee (as in (42b)) together show that they must be analyzed as intransitive:35 (56) a. Beaucoup de livres se vendront ce mois-ci. many of books se sell-fut this month 'Many books will be sold this month.' b. Il se vendra beaucoup de livres ce mois-ci. it se sell-fut many of books this month 'There are many books that will be sold this month.'

Composition in French CPs

31

Since the argument structure, the transivity, and the content of mediopassives and their corresponding basic lexemes are different, we propose a lexical rule for medio-passive lexemes which we formulate as follows:

This lexical rule supresses the first element in the ARG-ST of the input; the NP complement is thus promoted to first position in the output's ARG-ST. It also adds an a-aff complement (realized as se) in the output.36 There is a complex semantic effect of the medio-passive, but we make no attempt to describe it here. The transitive lexeme VENDRE ('to sell') thus undergoes the MedioPassive LR and the resulting lexeme gives rise to inflected words like the following:

32

Abeille et al.

This infinitive can function as the complement of composition FAIRE because it is classified as bas-vb. Koenig and Jurafsky (1994) point out that using lexical rules for medio-passive may raise a problem when one considers the interaction with the lexical rule used for impersonal constructions. Given that the subject is third person, and se does not distinguish between singular and plural antecedent, it is difficult to know whether the antecedent of se is the impersonal subject il or the complement NP which is the subject of the corresponding basic transitive verb. However, the difference appears when the subject of the verb is a first or second person NP. Consider then the impersonal verb in (59), whose argument toi has a metonymic interpretation ('your books,' rather than 'you'):37 (59) a.

Il ne se vend bien que toi it-3sg ne se sell-pres-3sg well only you-2sg dans ce salon. in this meeting 'The only one (whose books are) selling well in this meeting is you.' b. * Il ne te vend bien que toi it-3sg ne te-2sg sell-pres-3sg well only you-2sg dans ce salon. in this meeting

This example shows that the antecedent of the reflexive clitic is the impersonal subject il, which is a third person, while the complement is a second person pronoun. If the Medio-Passive LR output were to explicitly force coindexation (which entails sharing of person features) of its subject and the anaphor, then this would incorrectly require the agreement pattern of (59b) in the impersonal. But notice that the ouput of our Medio-Passive LR does not stipulate coindexation. This is unnecessary because precisely this coindexation is independently required (on words) by Principle A of the binding theory. The difficulty noted by Koenig and Jurafsky thus disappears given our analysis. In other words, the interaction between the Impersonal Subject Lexical Rule and the (one or two) lexical rules whose output is a reflexive verb is correctly described if none of these rules specifies the index of the a-aff element on the ARG-ST. Principle A, through its interaction with other independently motivated aspects of our grammar of French, thus plays a crucial role in the analysis of medio-passive constructions.38

Composition in French CPs

33

5.4. Clitic Trapping with Composition FAIRE We can now turn to one remaining question about downstairs clitics with FAIRE: clitic trapping. When a clitic stays on the infinitival, all complement clitics (including nonreflexive nonintrinsic ones) must stay downstairs as well:39 (60) a. Marie a fait s'en souvenir Jean. Marie has made se-of-it to-remember Jean 'Marie made Jean remember it.' b. *Marie en a fait se souvenir Jean. Marie of-it has made se to-remember Jean c. Qa lui fera leur en vouloir. this him-dat make-fut them-dat of-it to-want 'This will make him get annoyed at them.' d. *Qa leur fera en vouloir a Jean. this them dat make-fut of-it to-want to Jean 'This will make Jean get annoyed at them.' This puzzling behavior is straightforwardly accounted for in our analysis, given the constraint that the COMPS list of a cl-wd can never contain aff elements. This means that if a complement of a cl-wd is inherited by FAIRE, it cannot be of type aff, so it cannot be realized as a clitic on FAIRE. Thus the only clitics that may appear on the causative in these examples (lui in (60c)) correspond to arguments of FAIRE (e.g., the causee) that are not themselves part of the argument structure of the infinitival complement.

6. CONCLUSION Drawing on previous analyses of complex predicates in French, we have proposed an account of the contrast between tense auxiliaries and causative verbs with respect to downstairs clitic realization. We have defined a small set of verbal types. On the one hand we distinguish verbs bearing clitic morphology (Miller and Sag (1997)) from uncliticized verbs. Independent of this is our distinction between basic verbs (either uncliticized verbs or verbs realizing lexemes whose clitics never alternate with nonclitics) and reduced verbs (those which realize potentially nonaffixal arguments as clitics). Reflexive verbs, as we have seen, have a complex relation to this last distinction.

34

Abeille et al.

Both tense auxiliary and causative constructions are analyzed as complex predicates whose argument structures contain elements that are selected by the complex predicate's lexical argument (a participle or infinitive). Tense auxiliaries and the past participles with which they combine share elements of their argument structure. Composition FAIRE (and other causative verbs), by contrast, incorporate into their argument structure members of their infinitive's COMPS list. (These are always arguments that have a potential for nonaffixal realization.) Both tense auxiliaries and causative verbs constrain their lexical complement, requiring that it be a basic verb, but they differ with respect to the possibility of downstairs clitic realization. The distinction in composition type that we have posited (c-composition vs. a-composition) plays a fundamental role in our account of the fact that clitics always "climb" onto the tense auxiliary but intrinsic clitic verbs (and these alone) require downstairs realization in the composition causative construction.

ACKNOWLEDGMENTS We thank Gosse Bouma, Mike Calcagno, Ann Copestake, Francis Corblin, Tony Davis, Georgia Green, Erhard Hinrichs, Andreas Kathol, Bob Levine, Chris Manning, Jean-Marie Marandin, Philip Miller, Jean-Claude Milner, Tsuneko Nakazawa, Carl Pollard, Pierrette Thibault, Eun Jung Yoo, and our reviewers. This paper is part of a larger project on French syntax (see Abeille et al., in preparation) that has been supported in part by grants from the National Science Foundation (INT-94-16855) and CNRS.

NOTES *We analyse Romance clitics as pronominal affixes rather than syntactically independent words (see section 2.1). In this context then, the term "clitic" is a misnomer; nonetheless, we continue to use it for reasons of familiarity. 2 Other cases of clitic climbing in French are passive and copular constructions and genitive en. See Abeille and Godard (1994, 1996a) and Sag and Godard (1994) for compatible analyses in HPSG. 3 The possibility of attaching pronominal clitics to an infinitive is not a general trait of Romance causative constructions; Italian seems to disallow all downstairs clitics on the infinitival complement (e.g., Monachesi, 1995), while Spanish and Catalan allow more such attachments than French does, though there is some variation among speakers. (See Moore, 1990; Alsina, 1996.)

Composition in French CPs 4

35

Note that both the ARG-ST and the COMPS features belong to the syntactic description of items. LEXEMES specify how elements of the ARG-ST are linked to semantic roles. 5 The c-composition analysis plays a role in certain earlier HPSG analyses of German nonfmite auxiliary, raising, and certain control constructions. See Kiss (1994), Hinrichs and Nakazawa (1994), and the references cited there. 6 Actually, in the analysis we present below, the distinction in question is made as two types of verb HEAD values. Nothing essential hinges on this. 7 See Bouma (1997) and Bouma et al. (1997). 8 We draw freely here from the approach to lexical organization developed in Koenig (1994). Koenig uses "and-or" nets to constrain the interaction of types and type constraints. 9 The sign e means list concatenation (or append) and O designates the shuffle operation. In the more complete analysis, as presented by Miller and Sag (1997) and Abeille et al. (in preparation), there is the further possibility of an ARG-ST element of type gap appearing on the ARG-ST list, but not on the COMPS list. We will ignore issues of extraction here. 10 Further types may be required in order to allow for other words whose SUBJ value is empty (e.g., case-marking prepositions). 11"List(type)" designates a list of objects, all of which are of type type; "nelist" stands for "nonempty list." 12 The symbol "/" in (16) indicates a default type assignment for the value of the feature HEAD. For a conception of defaults particularly compatible with the present approach, see Lascarides et al. (1996), from whom we take the "/" notation. 13 See Abeille and Godard (1996b) for discussion. 14 To account for what seems to be "optional clitic climbing," Monachesi (1995), following a suggestion of Rizzi (1978), also proposes two different structures for restructuring verbs in Italian: a VP complement structure (with the clitics downstairs), and a flat structure (with the clitics upstairs). Under her proposal, the causative FAIRE only takes the flat composition structure. Moore (1990) posits two structures for a set of verbs in Spanish which includes causative and perception verbs, but they do not seem to completely correlate with clitic position. We ignore here the faire-par construction, which is also an instance of composition FAIRE. 16 There is also positive evidence that the causee is a complement of FAIRE, based on "quantitative en" distribution (Milner, 1978) and on heavy-NP shift (Miller, 1991). 17 In the Lexical-Functional Grammar framework, complements can be shared (in the functional structure) without being sisters of the causative verb (in the constituent structure). Zaenen and Dalrymple (1996), building on earlier work of Alsina (1996), represent causative constructions in French as hierarchical from the constituency point of view (FAIRE takes a VP complement) but "flat" from the functional point of view (the functional structure of FAIRE and the infinitival are fused).

36

Abeille et al. 18

We assume that composition FAIRE takes three semantic arguments when it combines with a transitive infinitive: the causer (ACTOR), the causee (UNDERGOER) and the event denoted by the infinitival v. By contrast, as argued in Abeille et al. (to appear), FAIRE plus intransitive infinitive combinations (including FAIRE-PAR) are semantically dyadic. For related justification of the raising analysis of FAIRE plus intransitive infinitive combinations, see also Koenig (1994). 19 It is of course tempting to try to collapse these two composition causative types into one, relying on the fact that French in general allows only one (role-assigned) accusative per argument structure to explain the distribution of accusative and dative causees. Such an analysis (first suggested, to our knowledge, by Liz Bratt—see Bratt, 1990) would postulate that composition FAIRE is transitive, relying on the fact that transitives take one and only one NP object to explain the distribution of causees. Our best attempts at executing this to date, however, have been thwarted by the fact that some transitives need not realize their object on the COMPS list (e.g., VOIR), other transitive verbs require an s or VP complement instead of an NP (e.g., PROMETTRE), and certain intrinsic clitic verbs (e.g., EN VOULOIR A) require an argument (the en) which, while arguably accusative, cannot be composed into the ARG-ST of FAIRE (as they are not realizable as affixes on FAIRE), and hence cannot contribute an accusative to the determination of FAIRE'S transitivity. Yet, in all these cases, despite the absence of an accusative NP on the ARG-ST of the lower verb, the causee is realized as dative. 20 Here we include functional annotations on the various branches (H for the head, C for complements). 21 For some speakers, quantitative en provides a third set of verbs that allow "downstairs" cliticization (Rouveret and Vergnaud, 1980): (i) Paul fera en acheter trois a Marie. Paul make-fut of-them to-buy three to Marie Taul will make Marie buy three of them.' Notice that the same en can also appear upstairs for the same speakers: (ii) Marie en fera acheter trois a Paul. Marie of-them make-fut to-buy three to Paul 'Marie will make Paul buy three of them.' Since quantitative en is introduced by a lexeme-to-lexeme lexical rule that constrains the accusative complement to be an indefinite pronominal complement and adds a p-aff en in the ARG-ST of the verb (see Abeille et al., in preparation), the HEAD value of the output can optionally be typed as bas-vb in our account of this dialect, thus allowing either the pl-wd acheter or the cl-wd en-acheter as the complement of composition FAIRE. 22 Certain expressions (such as en voir des vertes et des pas mures 'to go through hard times' (literally: 'to see green ones and unripe ones')) also allow their clitics to attach to FAIRE. In our analysis, this effect is achieved by lexemically restricting the relevant argument to be en-aff, but not requiring the verb to be of type cl-wd. This correctly allows for an infinitival pl-wd realization with the en-aff on the COMPS list. This en-aff can then be realized on the composition causative via c-composition.

Composition in French CPs

37

23 Some add benefactive or ethical datives which are not syntactically different from intrinsic reflexives, or inchoative reflexives (Ruwet, 1972; Koenig, 1994), which are not syntactically different from medio-passives. 24 A11 speakers accept reflexive verbs of classes (b) and (c) as complement of causative verbs, while only some of them accept class (a). See below for more on interspeaker variation. 25 This is true for all (active) verbs, except for a few intransitive verbs, which select etre, and agree with the subject. 26 Intriguing data in inversion and control structures indicate that this may be an oversimplification: the v does not agree in person with a subject or a controller that follows it, although it agrees in number (Marandin, 1997); the reflexive morpheme does not agree in person either:

(i) A la Fac se sont alors rencontres [Marie et toi], at the University se are 3pl then met [Marie and you]2jg qui etiez tons deux etudiants en medecine. who were2pl both students in medicine 'Marie and you, who were both medical students, met at the University. These facts can be interpreted as showing that person agreement should be distinguished from number (and gender) agreement; se may thus be thought of as agreeing with the verb's agreement features, rather than with the subject. We leave this problem for future research. 27 As argued in Abeille et al. (in preparation), it is impossible to account for the difference in auxiliary selection between the basic verb and the reflexive verb by appeal to any independent syntactic or semantic property in French, contrary to what has been proposed in a general way for Italian (Perlmutter, 1983; Burzio, 1986) or Dutch (Zaenen, 1993). Coindexation between the subject and a complement is certainly insufficient to induce etre, since the non-reflexive verb with a pronoun co-indexed with the subject has auxiliary avoir: (i) Paul n'a lave que lui-meme. 'Paul washed only himself.' The explanation based on the presence of a "trace" in object position (Burzio, 1986) is also problematic. It is unsupported for intrinsic reflexives (like S'EVANOUIR), as well as for medio-passives, since the reflexive clitic does not alternate with any overt constituent. In the unaccusative analysis of medio-passives (Grimshaw, 1990), there is a "trace" in object position. However, there are decisive arguments against this analysis. See Alsina (1996) for discussion. 98 This conditional formulation is easily replaced by one where the relevant constraints all "live on" types. That formulation appeals to two subtypes of past-part. 29 True reflexive verbs may have a reflexive or reciprocal interpretation (if the subject denotes an aggregate entity); see Melis (1987). We keep the name "reflexive" in spite of this slight inaccuracy. 30 They are called "inherent reflexives," "intrinsic reflexives," or "essential reflexives" in reference grammars; we use the term "intrinsic" for the whole class

38

Abeille et al.

of clitics (and verbs bearing them) which occur downstairs in causative constructions. 31 Cliticization by itself does not turn a transitive verb into an intransitive; there is no impersonal construction corresponding to un enfant le lavait dans la fontaine ('a child was washing it in the fountain'): (i) *Il le lavait un enfant dans la fontaine. it it wash-imperf-3sg a child in the fountain (putatively) 'There was a child washing it in the fountain.' 32

Note that in our analysis, ARG-ST cannot be used to predict transitivity, as there are intransitive verbs with an accusative NP on their ARG-ST list. See note 19. 33 There are speakers who accept (50b), but still do not allow true accusative reflexives in the composition causative (they do not accept (51c)). These speakers have the lexical rule creating intransitive acc-refl-lxms like (52), but not the constraint in (53). Accordingly, true accusative reflexives are intransitive (hence a possible input for the Impersonal Subject LR whose outputs appear in the impersonal construction), but they are not bas-vb (and hence are unacceptable as complements of composition FAIRE). 34 That only transitive verbs alternate with medio-passives is challenged by Descles et al. (1986) on the basis of such examples as le pain moisit/le pain se moisit ('bread goes moldy'), les fruits pourrissent/les fruits se pourriront vite ('fruit will get rotten quickly'), Jean guerira tres vite/Jean se guerira tres vite ('Jean will be cured very fast'). However, the corresponding transitive verbs exist (moisir 'make moldy', pourrir, 'make rotten'; guerir, 'cure'); they enter into two alternations: one with the medio-passive verb, and one with intransitive verbs in the examples above, often called "ergatives" (similar to the alternation of the two verbs break in English). 35 In spite of the fact that their complement list contains an accusative NP. See note 37. 36 The a-aff is analyzed as an accusative complement because of agreement facts. See the discussion at the beginning of section 5. 37 That toi is second person is shown by examples such as: Il ne se vend bien que vous qui avez toujours toutes les chances. it-3sg ne se sell-pres-3sg well only you-2pl who have-2pl always all the luck 'The only one (whose books are) selling well is you who always have all the luck.' 38 Koenig (1994:62ff), in discussing an earlier version of this analysis (Godard and Sag, 1995), argues that binding principles should play no role in the analysis of medio-passives. He suggests that binding principles, as part of a module of syntactic principles, should not be accessible to constructs that are part of the lexical module, e.g., principles of lexical well-formedness. We share Koenig's concern for processing autonomy, but binding principles must be constraints on words in HPSG, as words (more precisely their categories) are the only constructs that have ARG-ST lists. Thus it seems reasonable to interpret binding principles (as long as they constrain only coarguments) as lexically specified disjunctive constraints on ARG-ST lists that function monotonically in syntactic processing. Some such scheme is necessary in any case, as third person p-aff clitics (le, la, les, lui,

Composition in French CPs

39

etc.) must interact with Principle B of the binding theory in just the same way that we propose se interacts with Principle A. Koenig's other objection is based on his view that the medio-passive clitics are really subject agreement markers rather than reflexive anaphors. While he may be right that they mark agreement, inversion facts show that it is not straightforwardly subject agreement marking (see note 27). In any case, it would be easy to modify our Medio-Passive LR so that it induced the appropriate agreement leaving all other aspects of our analysis unchanged. 39 See Kayne (1975), Rouveret and Vergnaud (1980:153-154), Aissen and Perlmutter (1983), and Tasmowski (1985:330).

REFERENCES Abeille, A., and D. Godard (1994). The complementation of tense auxiliaries in French. In Proceedings of WCCFL 13, 157-172. CSLI Publications, Stanford, CA. Abeille, A., and D. Godard (1996a). La complementation des auxiliaires franc,ais. Langages 122, 32-61. Abeille, A., and D. Godard (1996b). French Word Order and Lexicality. Paper given at HPSG 96 Conference, Marseille. To appear in R. Borsley (ed.), Syntactic Categories. Academic Press, New York. Abeille, A., D. Godard, P. Miller, and I. A. Sag (1997). Bounded Dependencies in French. In S. Balari and L. Dini (eds.), HPSG in Romance. CSLI Publications, Stanford, CA. Abeille, A., D. Godard, P. Miller, and Ivan A. Sag (to appear). The Causative Structures of French. Unpublished manuscript, Paris and Stanford. Abeille, A., D. Godard, and I. A. Sag (in preparation). The Major Syntactic Structures of French. Unpublished manuscript. Aissen, J., and D. Perlmutter (1983). Clause reduction in Spanish. In D. Perlmutter (ed.), Studies in Relational Grammar 1, 360-403. University of Chicago Press, Chicago. Alsina, A. (1996). The Role of Argument Structure in Grammar: Evidence from Romance. Lecture Notes Series. CSLI Publications, Stanford, CA. Auger, J. (1994). Pronominal Clitics in Quebec Colloquial French, A Morphological Approach. Doctoral dissertation, University of Pennsylvania. (Distributed as IRCS Report 94-29.) Bouma, G. (1997). Monotonic Lexical Rules. Unpublished manuscript, Rijksuniversiteit Groningen. Bouma, G., R. Malouf, and I. A. Sag (1997). Satisfying Constraints on Extraction and Adjunction. Unpublished manuscript, Groningen and Stanford. Bratt, E. O. (1990). The French Causative Construction in HPSG. Unpublished manuscript, Stanford. Bresnan, J. (ed.) (1982). The Mental Representation of Grammatical Relations. MIT Press, Cambridge, MA.

40

Abeille et al.

Burzio, L. (1986). Italian Syntax. Reidel, Dordrecht. Dalrymple, M. (1993). Syntactic Constraints on Anaphora. CSLI Publications, Stanford, CA. Davis, A. (1996). Lexical Semantics and Linking in the Hierarchical Lexicon. Doctoral dissertation, Stanford University. Descles, J.-P., Z. Guentcheva, and S. Shaumyan (1986). Theoretical analysis of reflexivization in the framework of applicative grammar. Linguisticae Investigationes X, 1-65. Dobrovie-Sorin, C. (to appear). Impersonal si constructions and the passivization of unergatives. Linguistic Inquiry. Emonds, J. (1978). The verbal complex V'-V in French. Linguistic Inquiry 9, 151-175. Godard, D., and I. A. Sag (1995). Reflexivization and Transitivity in French. Paper presented at annual meeting of the LSA, New Orleans. Grevisse, M. (1969). Le Bon Usage. Duculot, Liege. Grimshaw, J. (1982). On the lexical representation of romance reflexive clitics. In J. Bresnan (ed.), The Mental Representation of Grammatical Relations, 87-148. MIT Press, Cambridge, MA. Grimshaw, J. (1990). Argument Structure. MIT Press, Cambridge, MA. Hinrichs, E., and T. Nakazawa (1994). Linearizing Aux's in German verbal complexes. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German Grammar in HPSG, 11-37. CSLI Publications, Stanford, CA. Hyman, L. M., and K. E. Zimmer (1976). Embedded topic in French. In Charles N. Li (ed.), Subject and Topic, 191-211. Academic Press, New York. Kayne, R. (1975). French Syntax: The Transformational Cycle. MIT Press, Cambridge, MA. Kiss, T. (1994). Obligatory coherence: The structure of German modal verb constructions. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German Grammar in HPSG, 71-107. CSLI Publications, Stanford, CA. Koenig, J.-P. (1994). Lexical Underspecification and the Syntax-Semantics Interface. Doctoral dissertation, University of California, Berkeley. Koenig, J.-P., and D. Jurafsky (1994). Underspecification and on-line type construction in the lexicon. In Proceedings WCCFL 13, 270-285. CSLI Publications, Stanford, CA. Lascarides, A., E. Briscoe, N. Asher, and A. Copestake (1996). Order independent and persistent typed default unification. Linguistics and Philosophy 19(1), 1-89. Legendre, G. (1987). Topics in French Syntax. Doctoral dissertation, University of California, San Diego. Manning, C. (1994). Ergativity: Argument Structure and Grammatical Relations. Doctoral dissertation, Stanford University. Marandin, J.-M. (1997). Dans le Titre se Trouve le Sujet; VInversion Locative en Franqais. Unpublished manuscript, Universite Paris 7. Melis, L. (1987). La Vote Pronominale. Duculot, Liege. Miller, P. (1991). Clitics and Constituents in Phrase Structure Grammar. Doctoral dissertation, Rijksuniversiteit Utrecht. (Published in 1992, Garland, New York.) Miller, P., and I. A. Sag (1995). Une analyse lexicaliste des affixes pronominaux en frangais. Revue Quebecoise de Linguistique 1, 135-171.

Composition in French CPs

41

Miller, P., and I. A. Sag (1997) French clitic movement without clitics or movement. Natural Language and Linguistic Theory, 15, 573-639. Milner, J.-C. (1978). De la Syntaxe a VInterpretation. Le Seuil, Paris. Milner, J.-C. (1982). Ordres et Raisons de Langue. Le Seuil, Paris. Monachesi, P. (1995). A Grammar of Italian Clitics. Doctoral dissertation, KUB, Tilburg. Moore, J. (1990). Reduced Constructions in Spanish. Doctoral dissertation, University of California, Santa Cruz. Morin, J.-Y. (1978). Une theorie interpretative des causatives en franqais. Linguisticae Investigationes 11(2), 363-417. Perlmutter, D. (1983). Personal vs impersonal constructions. Natural Language and Linguistic Theory 1, 141-199. Pollard, C., and I. A. Sag (1987). Information-Based Syntax and Semantics, Vol. 1. Lecture Notes Series. CSLI Publications, Stanford, CA. Pollard, C., and I. A. Sag (1992). Anaphors in English and the Scope of Binding Theory. Linguistic Inquiry 23, 261-303. Pollard, C., and I. A. Sag (1994). Head Driven Phrase Structure Grammar. University of Chicago Press, Chicago, and CSLI Publications, Stanford, CA. Rizzi, L. (1978). A restructuring rule in Italian syntax. In S. J. Keyser (ed), Transformational Studies in European Languages. MIT Press, Cambridge, MA. Rouveret, A., and J.-R. Vergnaud (1980). Specifying reference to the subject. Linguistic Inquiry 11(1), 97-202. Ruwet, N. (1972). Les constructions pronominales neutres et moyennes. In Theorie Syntaxique et Syntaxe du Franqais, 87-125. Le Seuil, Paris. Sag, I. A., and D. Godard (1994). Extraction of de-phrases from the French NP. In Proceedings of NELS 24, 519-541. University of Massachusetts, GLSA, Amherst, MA. Scalise, S. (1984). Generative Morphology. Foris Publications, Dordrecht. Tasmowski, L. (1985). Faire infinitif. In L. Melis, L. Tasmowski, P. Verluyten, and D. Willems (eds.), Les Constructions de la Phrase Franchise. Invitation a la Reflexion sur le Passif, le Pronominal, l'Impersonnel, et le Causatif. Communication and Cognition, Gent. Wechsler, S. (1995). The Semantic Basis of Argument Structure. CSLI Publications, Stanford, CA. Zaenen, A. (1993). Unaccusativity in Dutch: Integrating syntax and lexical semantics. In J. Pustejovsky (ed.), Semantics and the Lexicon, 129-161. Kluwer, Dordrecht. Zaenen, A., and M. Dalrymple (1996). Les verbes causatifs 'polymorphiques': les predicats complexes en fran^ais. Langages 122, 79-95. (Nouveaux Raisonnements Syntaxiques, A. Abeille and D. Godard (eds.).) Zubizarreta, M. L. (1985). The relation between morphophonology and morphosyntax: The case of romance causatives. Linguistic Inquiry 16, 247-289.

This page intentionally left blank

WORD ORDER CONSTRAINTS ON VERB CLUSTERS IN GERMAN AND DUTCH GOSSE BOUMA GERTJAN VAN NOORD Vakgroep Alfa-informatica & BCN Rijksuniuersiteit Groningen Groningen The Netherlands

I. INTRODUCTION

There is a broad concensus among researchers working within the paradigm of HPSG that complement inheritance of the kind proposed in Hinrichs and Nakazawa (1989, 1994) is an essential operation in the analysis of the verb cluster in German and Dutch. Both languages have a class of verbs (including auxiliaries and modals) that subcategorize for a (possibly) unsaturated verbal complement, and for all the complements on the coMPS-list of this verbal complement. Most analyses of German have assumed that these complement inheritance verbs combine with their verbal complements to form a phrase consisting of (lexical) verbs only. This phrase is usually referred to as the verbal complex. In this paper, we argue that the word order of German as well as Dutch verb clusters can be accounted for without introducing a verbal complex.1 Our analysis rests on the assumption that a single HEAD-COMPLEMENT schema exists, which licenses phrases consisting of a lexical head and an arbitrary number of its complements. This schema allows a complement inheritance verb to combine with its verbal complement, as well as the complements of this complement, in one step. A consequence of this Syntax and Semantics, Volume 30 Complex Predicates in Nonderivational Syntax

43

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

44

G. Bouma and G. van Noord

analysis is that there is no room within the verb phrase for partial VPS or a verbal complex. The advantage of such an account is that there is no need to distinguish between a rule schema for verbal complexes and for (partial) VPS. Furthermore, a "flat VP" implies that phrase structure does not impose any constraints on word order. Therefore, the full range of word order possibilities found in German and Dutch verb clusters is captured by a single schema. Of course, the main challenge for this "flat VP" analysis is to demonstrate that it can do so without leading to vast overgeneration. This is the main topic of the current paper. In the next section, we introduce the German data and discuss the analysis of Hinrichs and Nakazawa (1989, 1994) as well as a number of related approaches. Next, we present our analysis of German. It uses a general HEAD-COMPLEMENT schema in conjunction with linear precedence statements. We demonstrate that the proposed set of LP-statements accounts for all ordering possibilities encountered in the German verb cluster. Furthermore, we argue that our analysis leads to an improved account of partial VP fronting (Nerbonne, 1994). The account of word order adopted here is considerably more sophisticated than the proposal in van Noord and Bouma (1996), in which an account of the Dutch verb cluster was presented which relied primarily on ordering in terms of obliqueness. In section 4, we demonstrate that the improvements that were necessary in order to account for German also lead to a smoother account of some of the more problematic Dutch data.

2. THE CONSTITUENT STRUCTURE OF GERMAN VPS German subordinate clauses containing a cluster of a main verb and modal and/or auxiliary verbs give rise to a nesting (or embedding) dependency word order: (1) dafi er das Examen bestehen konnen wird. that he the exam pass be-able-to will 'that he will be able to pass the exam' Hinrichs and Nakazawa (1989, 1994) argue that the phenomenon known as Oberfeldumstellung or auxiliary flip (2) provides evidence for the fact that the complement of auxiliaries and modals (i.e., wird and konnen in (2)) is not a full VP, but a constituent consisting of verbal material only: (2)

dafi er das Examen wird bestehen konnen.

Word Order Constraints on Verb Clusters

45

If konnen were to select the VP das Examen bestehen as complement, giving rise to a constituent which is itself the complement of wird, the word order in (2) would be completely unexpected. If konnen and wird are "complement inheritors," however, an analysis suggests itself in which bestehen konnen, but not das Examen bestehen konnen, is a constituent. The proposed lexical entries for konnen and wird are given in (3):2

The NPCOMPS feature plays a role in the following two HEAD-COMPLEMENT schemata:

The first rule in (4) licenses the derivation of a verbal complex, whereas the second rule licenses the derivation of (partial) verb phrases. Note that, in VPS with standard word order, these two schemata give rise to a binary, left-branching tree:3

46

G. Bouma and G. van Noord

Auxiliary flip is accounted for by means of a binary head feature FLIP. The following linear precedence statement expresses that verbal complements marked [FLIP + ] must follow the head:

Since main verbs never induce flipped word order, they are considered to be marked [FLIP — ]. Infinitival modals such as konnen are unspecified for the head feature FLIP. Thus, given the constituent structure in (5) and the LP statement in (6), it is predicted that both the word order in (1) and in (2) are allowed. A complication arises from the fact that konnen can also function as Ersatzinfinitiu (substitute infinitive).4 In those cases, flipped word order is obligatory. This fact is accounted for by assigning the Ersatzinfinitiu konnen the value [FLIP + ]:

This accounts for the contrast in (8): (8) a. daft er das Examen hat bestehen konnen. that he the exam has pass be-able-to 'that he has been able to pass the .exam' b. *dafi er das Examen bestehen konnen hat. The auxiliary haben, finally, functions as a "trigger" for flipped word order only in case it appears in "flipped" position itself. This is illustrated in (9): (9) a. dafi er die Lieder wird haben singen konnen. that he the songs will have sing be-able-to 'that he will have been able to sing the songs' b. * dafi er die Lieder haben singen konnen wird. c. dafi er die Lieder gesungen haben wird.

Word Order Constraints on Verb Clusters

47

In (9a), haben must precede singen konnen and therefore is a trigger for flipped word order itself. In (9c), haben does not appear in flipped position and consequently cannot trigger flipped word order for wird. This can be accounted for by assuming that haben inherits the FLiP-value of its verbal complement:

An essential aspect of the Hinrichs and Nakazawa account of word order within the verb cluster is the distinction between verbal complexes and other partial verb phrases expressed by the feature NPCOMPS. First, in sentences with "normal" word order, selection for a [NPCOMPS — ] complement prevents ambiguity. That is, in those cases the [NPCOMPS -] specification on the complement ensures that the inheritance verb combines with a verb or verbal complex, rather than a full VP. Second, in sentences with "flipped" word order, the NPCOMPS feature ensures that the auxiliary "flips" only over the verbal complex, and not over a full or partial VP. This is illustrated in (11): (11) a. wenn er dem Professor die Studenten hdtte vorstellen durfen. if he the professor the students had introduce be allowed 'if he had been allowed to introduce the students to the professor' b. *wenn er die Studenten hdtte dem Professor vorstellen durfen. c. *wenn er dem Professor hdtte die Studenten vorstellen durfen. d. *wenn er hdtte dem Professor die Studenten vorstellen durfen. While the examples in (llb-d) are judged grammatical by some speakers, other speakers reject such examples. The fact that auxiliaries and modal verbs are complement inheritance verbs has been widely accepted. Questions concerning the constituency of phrases headed by such, however, have not been answered uniformly. In Kiss (1994), for instance, it is argued that the verbal complex is right-

48

G. Bouma and G. van Noord

branching instead of left-branching:

This analysis has the advantage that it is no longer necessary to have two HEAD-COMPLEMENT schemata, one for creating verbal complexes and a second rule for creating (partial) VPS. The feature NPCOMPS can dispensed with as well. Instead, complement inheritance verbs select a verbal complement which is marked +LEX. A problem for this kind of analysis is obviously that there is no easy way to account for auxiliary flip. Proposals for a nonbinary branching analysis have been put forward as well. In Nerbonne (1994), for instance, it is argued that instances of partial VP fronting (13) are best accounted for by making minimal assumptions about the internal structure of VPS: (13) a. Das Buck lesen wird er schon konnen. the book read will he already can 'He'll surely be able to read the book.' b. Lesen konnen wird er das Buch schon. For transformational accounts, which assume a correspondence between a fronted element and a trace in the remaining clause, such examples are problematic. Only an analysis that assigns multiple bracketings to das Buch lesen konnen can account for the two examples in (13). However, such an analysis must accept spurious ambiguity in examples without fronting. Nerbonne provides an alternative, nontransformational and traceless, account, in which a complement extraction lexical rule shifts an element from COMPS to SLASH. A special feature of this rule is that the requirement that a complement must be [NPCOMPS -] or +LEX is not carried over to SLASH. This essentially allows a complement inheritance verb, which normally selects for a verbal complex or lexical verb, to have an element on SLASH matching an arbitrary partial VP. Nerbonne emphasizes that, given his analysis, fronting of partial VPS is no longer an argument for assigning constituent status to such phrases in

Word Order Constraints on Verb Clusters

49

nonfronted positions as well. Consequently, there seems to be no reason why a verbal head could not combine with all its complements in one step. Although Nerbonne does not address this issue in detail, he suggests that an account of auxiliary flip still necessitates the introduction of a verbal complex. That is, an auxiliary could have a verbal complex as one of its arguments:

However, even the existence of verbal complexes can be questioned. In Baker (1994), for instance, the example above is assigned the following structure:

Flat structures of this kind can be obtained if complement inheritance verbs select a lexical complement and, furthermore, the HEAD-COMPLEMENT schema imposes the constraint that the head must be lexical (i.e., of type word). It should be noted, however, that Baker tries to account for auxiliary flip by assuming that in those cases the auxiliary takes a partial VP as argument. As is admitted by Baker herself, such an account overgenerates, as it not only accepts ordinary cases of auxiliary flip, but also all of the examples in (11) and even such sequences as (16), which are completely ungrammatical: (16)

*dqfi er bestehen wird das Examen konnen. that he pass will the exam be-able-to 'that he will be able to pass the exam'

The proposal of Hinrichs and Nakazawa (1994) deals successfully with auxiliary flip in the German verb cluster. Alternatives which also explicitly address this issue, such as Baker (1994), are not without problems. This does not imply, however, that alternatives are not worth considering. For one thing, while the analysis of Hinrichs and Nakazawa leads to a smooth account of auxiliary flip, it cannot easily account for some other word order patterns.

50

G. Bouma and G. van Noord

In Dutch subordinate clauses, for instance, the presence of auxiliaries and modals also leads to a cluster of verbs in clause-final position. The standard order in this case, however, is one where the governing verb precedes the verb it governs: (17) dat Jan het boek moet hebben gelezen. that John the book must have read 'that John must have read the book' As in German, a certain amount of word order variation is permitted. The participle in (17), for instance, may also be the first element in the cluster: (18)

dat Jan het boek gelezen moet hebben.

An analysis which assigns constituent status to gelezen hebben cannot easily account for this possibility. Of course, it may be that constituency within the verb cluster in Dutch is simply different than in German. However, there are also German examples which appear to be problematic. Meurers (1994) has drawn attention to the following examples, which he refers to as cases of Zwischenstellung: (19) zu dem Zeitpunkt an dem ich mich entscheiden hdtte mussen at the point at which I me decide had must 'at the point at which I should have made a decision' The analysis of Hinrichs and Nakazawa (1994) considers entscheiden mussen as a constituent. The fact that hdtte may, for a considerable number of speakers, "flip" over the modal verb only in this case, is clearly problematic for their analysis. Below, we develop an alternative analysis for the German verb cluster. It not only handles auxiliary flip, but also Zwischenstellung. An advantage of our "flat" analysis is that it uses a single HEAD-COMPLEMENT schema to derive both full and partial VPS, without spurious ambiguity. In previous work (van Noord and Bouma, 1996), we argued that this means that a single rule can account for the derivation of full VPS as well as the kind of partial VPS encountered in examples of partial VP fronting and partial extraposition (third construction) in Dutch. Thus, the alternative developed below accounts for a wider range of data, and does so by using fewer rules, than previous proposals. Moreover, there is no need for a feature such as NPCOMPS. This is advantageous because the percolation of this feature in previous analyses is rather peculiar and not subject to one of the ordinary feature percolation constraints (such as the head feature principle).

Word Order Constraints on Verb Clusters

51

3. WORD ORDER WITHIN THE GERMAN VERB CLUSTER

In this section, we present our analysis of German verb clusters. We present a HEAD-COMPLEMENT schema which allows both full and partial verb phrases to be derived. Furthermore, we present our approach to word order, which relies on LINEAR PRECEDENCE statements.5 Linear precedence within phrases is determined by three principles: Directionality. Directionality determines the position of a complement daughter relative to the head daughter. Topology. Topological constraints determine in which "topological field" a daughter appears. Instead of adopting the full range of topological fields usually assumed for German syntax, we will only distinguish between "inner zone" and "outer zone" positions, where the "inner zone" contains the verb cluster, and the "outer zone" contains all other elements of the VP. Government. The direction of government determines the position of a complement relative to its governor, where government is defined in terms of argument structure. The distinction between the verb cluster and other parts of the VP is expressed by means of the distinction between inner and outer zone, and the order of elements within the verb cluster is determined by the direction of government. 3.1. The Head-Complement Schema

We assume the following, general, HEAD-COMPLEMENT schema:

Schema (20) states that a phrase may consist of a lexical head daughter and one or more complement daughters. The requirement that the head daughter must be of type word, while the mother is a phrase rules out recursive rule application, i.e., no phrase derived by means of the HEAD-COMPLEMENT schema can be the head of a larger phrase also derived by means of the HEAD-COMPLEMENT schema. The requirement that there must always be at least one complement means that there can be no nonbranching derivations. Both constraints are required to rule out spurious derivations.

52

G. Bouma and G. van Noord

The HEAD-COMPLEMENT schema is subject to the VALENCE Principle, which is similar to the Subcategorization Principle of Pollard and Sag (1994:34):6 (21) Valence Principle: The coMPS-list of the head daughter is the append (0) of the coMPS-list of the mother and the list of SYNSEM-values of the complement daughters. Note that valence does not distinguish between a case where all complements are selected, and cases where one or more complements are not selected. The latter allows partial phrases to be derived.7 This is essential for our account of partial VP fronting and partial extraposition (or third construction) verbs. The order of complements relative to the head daughter is determined by the feature DIR. A head daughter may specify its complements as either [DIR ->], in which case the complement must precede the head, or as [DIR ( ^>) (the governee must (immediately) precede its governor), and ] instead of [GVOR -> ]. The restriction that the governor of a main verb must be right-adjacent rules out the derivation in (38), but not the derivations in (36) and (37). The final issue to be discussed concerns Ersatzinfinitiv and the auxiliary haben. If haben governs a modal verb, this verb will appear in its infinitival form, instead of its participle form. Following Hinrichs and Nakazawa (1994), we will assume that the infinitive form of a modal may be marked [VFORM psp]. Furthermore, since haben must obligatorily precede the modal verb in this case (40), it is specified as [GVOR -» gi is injective (one-to-one and into) (Bresnan, 1982:163). A notable exception to this generally accepted view is found in Jackendoff (1990), who explicitly argues for the need of having two arguments link to one and the same syntactic NP. In fact, overwhelming evidence from languages such Chinese, Dutch, German, and Hindi/Urdu (Li, 1995; Neeleman, 1994; T. Mohanan, 1994a; Butt, 1995; Alsina, 1995) points to

78

Miriam Butt

the fact that the restrictions expressed by Function-Argument Biuniqueness or the Theta Criterion must be weakened. At the very least, they cannot apply in quite the way envisioned to date. For example, Alsina (1993:81-82) instead proposes an Argument-to-Function Uniqueness condition. I adopt this condition here.5 Argument-to-Function Uniqueness. There cannot be two distinct syntactic functions with the same argument value. The adoption of this more flexible constraint entails that more than one argument may correspond to a single syntactic function, thus opening the door for argument fusion. 2.4. Argument Fusion The assumption of the weaker principle of Argument-to-Function Uniqueness frees Alsina (1993) from the need to attempt a strict one-to-one mapping from arguments to syntactic positions. Following T. Mohanan (1988) for Malayalam, he posits both the traditionally accepted two place causative, and an additional three place causative. These are illustrated in (5) and (6). The two place causative accounts for constructions with optional oblique causees in Romance, while the three place causative is taken to form the causatives in (1) and (2). Note that in the three place causative two arguments are assumed to map to the same syntactic constituent. This is indicated by the bar connecting them and can be thought of as Argument Fusion, which combines two arguments, rendering only one of them visible for linking purposes.

The case marking alternations observed in (1) and (2) now follow from independent principles of LFG's linking theory in which arguments are linked to grammatical functions, and the further assumption that in Romance the dative case is assigned to the thematically most prominent internal argument (Alsina, 1993).

Constraining Argument Merger

79

This nonstipulative account of case marking is a nice result of the type of Argument Fusion or Composition engendered by (6). As already noted, it does violate the Principle of Function-Argument Biuniqueness in LFG. And it is exactly because (6) would violate the corresponding constraint in GB that S. Rosen avoids the notion of Argument Fusion and the concomitant responsibilities of formulating appropriate restrictions on it. Similarly, approaches to Romance complex predicates within HPSG such as Monachesi (1995) or Manning and Sag (1998) avoid the question of Argument Fusion. These approaches rely on the mechanism of Argument Composition, first introduced by Hinrichs and Nakazawa (1990) for a treatment of verb clusters in German, which is similar to the original RG clause union analyses and the notion of Argument Merger first introduced by Grimshaw and Mester (1988) in that a matrix predicate (e.g., a restructuring verb in Romance) is simply postulated to subcategorize for a further predicate and, by extension, for all the arguments of that predicate. The possibility of identifying two arguments with one another, as illustrated above, never arises.6 However, as Alsina (1993) and Butt (1995) have argued for Romance causatives and Urdu complex predicates, respectively, a satisfactory analysis of the wide range of data involved cannot be achieved without such a notion. Once this position is granted, some questions which arise immediately must necessarily be considered. How can Argument Fusion be thought of formally? What restrictions operate on Argument Fusion: can any argument simply be fused with any other argument, and if so, is this supported by linguistic phenomena? In this paper, I make a first attempt at formulating a restrictive theory of complex predicate formation. Building on earlier work, I propose defining characteristics of complex predicates, and formulate restrictions on the types of argument merger that are available. Ultimately, the enterprise should yield an intuitive and simple account of complex predicates crosslinguistically. Within the scope of this paper, I first briefly recapitulate evidence for Argument Fusion from the Urdu Permissive, which serves as an excellent case in point for the need to recognize argument structure as an independent level of representation (see Butt, 1994, 1995, for an extensive discussion). Building on the insights provided by the Urdu Permissive, I then examine Urdu causatives in light of Alsina's (1992, 1993) approach and propose a restrictive view of Argument Fusion in conjunction with the integration of aspectual information into the theory of linking, which has the desirable result of rendering some of the linking mechanism more intuitively transparent.

80

Miriam Butt

3. COMPLEX PREDICATE FORMATION 3.1. The Urdu Permissive Data from agreement, anaphora, and control very clearly show that the Urdu Permissive in (7)7 must be analyzed as a complex predicate. Furthermore, a comparison of the Permissive with the superficially similar Instructive in (8) shows that phrase structure cannot be taken to determine complex predicate status. (7) anjum = ne saddaf = ko haar Anjum.F = Erg Saddaf.F = Dat necklace.M = Nom banaa-ne di-yaa make-Inf.Obl give-Perf.M.Sg 'Anjum let Saddaf make a necklace.' (8) anjum = ne saddaf = ko [haar Anjum.F = Erg Saddaf.F = Dat necklace.M = Nom banaa-ne = ko] kah-aa make-Inf.Obl = Acc say-Perf.M.Sg 'Anjum told Saddaf to make a necklace.' The contrast between these two constructions demonstrates that complex predicate formation is best analyzed at the level of argument structure. Note that this point has also been made by Manning (1992) and Alsina (1993) for Romance complex predicates, and by T. Mohanan (1994a) for Hindi N-V complex predicates. However, the Urdu Permissive contrasted with the Instructive provides some of the clearest and most forceful evidence in favor of an argument structure analysis. The evidence comes from control, anaphora, and agreement phenomena. Only the agreement data are recapitulated here, with respect to which the Permissive robustly behaves like a simple predicate, while the Instructive must be analyzed as a matrix verb with an embedded infinitival complement. In Urdu a verb can only agree with one of its own nominative arguments (see T. Mohanan, 1994a, for details). If both the subject and the object are nominative, then the verb agrees with the subject.8 When there is no nominative argument in the clause, the verb is inflected with the default masculine singular -aa. A verb can never agree with an argument outside of its own domain of predication. The lack of agreement between the feminine object cittii 'letter' and the finite verb of the Instructive in (9) indicates that the object cannot be a matrix argument. Rather, it must be the argument of the embedded infinitive 'to write.'

Constraining Argument Merger

81

(9) anjum = ne saddaf =ko [citthii Anjum.F = Erg Saddaf.F = Dat note.F = Nom likh-ne = ko] kah-aa write-Inf.Obl = Acc say-Perf.M.Sg 'Anjum told Saddaf to write the note.' The Permissive in (10) does not display the same pattern as the Instructive. In both (l0a) and (l0b) the finite verb agrees with the object. (10) a. anjum = ne saddaf = ko citthii Anjum.F = Erg Saddaf.F = Dat note.F = Nom likh-ne d-ii write-Inf.Obl give-Perf.F.Sg 'Anjum let Saddaf write a note.' b. anjum = ne saddaf = ko xat Anjum.F = Erg Saddaf.F = Dat letter.M = Nom likh-ne di-yaa write-Inf.Obl give-Perf.M.Sg 'Anjum let Saddaf write a letter.' Because the finite verb agrees with citthii 'letter' in (l0a), this noun cannot be an embedded object, but must be analyzed as a matrix object.9 The Permissive thus behaves as if it were a simple clause headed by a single predicate. The agreement, anaphora, and control facts remain constant under scrambling as well. While this in itself is a strong indication for a nonstructural (and nonderivational) analysis of complex predicates, even more interestingly, phrase structural properties such as scrambling, negation, and coordination do not differentiate between the two constructions. A purely structural and derivational account could still, no doubt, be formulated; however, the more fruitful avenue of research would seem to lie in pursuing an analysis which allows for an independent level of argument structure, and which relates phrase structure representations to grammatical functions nonderivationally, as is the case in LFG. 3.2. Complex Predicate Analysis The essential, defining characteristics of complex predicates, as evidenced by the Permissive, can be stated as follows. (11)

Defining Properties of Complex Predicates: • The argument structure is complex (two or more predicative heads contribute arguments).

82

Miriam Butt

•

The grammatical functional structure is that of a simple predicate. In particular, it contains only a single subject. • The phrase structure may be either simple or complex. It does not necessarily determine the status of a complex predicate. As LFG encodes grammatical relations at a separate level from phrase structure, the differences and similarities between the Permissive and the Instructive can easily be accounted for in terms of f(unctional)-structure and c(onstituent)-structure properties. The fact that the Permissive is a complex predicate while the Instructive is not is expressed through the relationship between a-structure and f-structure. As shown in the abbreviated f-structure10 representation in (14) for the Permissive, a single domain of predication at f-structure corresponds to complex predication at a-structure.11 Thus, there is only one subject, one object, and one indirect object: the Permissive is flat at f-structure, but complex at a-structure. (12) anjum = ne saddaf = ko haar Anjum.F = Erg Saddaf.F = Dat letter.M = Nom banaa-ne di-yaa make-Inf.Obl give-Perf.M.Sg 'Anjum let Saddaf make a necklace.'

The f-structure in (17) for the Instructive, on the other hand, contains two argument taking PREDS. The PRED 'say' takes an XCOMP as one of its arguments. This XCOMP represents a second, embedded domain of predication of the PRED 'write.' The Instructive is thus represented by a complex a-structure which corresponds to a complex f-structure.

Constraining Argument Merger

83

(15) anjum = ne saddaf = ko citthii Anjum.F = Erg Saddaf.F = Dat letter.F = Norn likh-ne = ko kah-aa write-Inf.Obl = Acc say-Perf.M.Sg 'Anjum told Saddaf to write a letter.'

(16)

(17)

As LFG defines f-structure and c-structure as independent, but mutually constraining levels of representation, a given sentence may have more than one c-structure realization, as long as the requirements of completeness and coherence are met at f-structure (see Bresnan, 1982). The data from agreement, anaphora, and control, which indicated that the Permissive and the Instructive differ from one another, are completely accounted for at f-structure, while the data from coordination, scrambling, and negation can be accounted for at c-structure. 3.3. Light Verbs Under the approach to complex predicates taken here, argument structure must be viewed as an interface between lexical semantics and syntax. I follow T. Mohanan (1994a) in assuming that a-structure, while linked to lexical semantics, is not itself a semantic level of representation. This is in line with more structural approaches, such as Hale and Keyser (1993). However, as discussed previously, I assume that the encoding of arguments

84

Miriam Butt

at argument structure is not necessarily isomorphic with their syntactic realization. A further idea found in many analyses of complex predicates, and one that is adopted here as well, is that complex predicate formation is triggered by light verbs, a term first used by Jespersen (1954) to characterize verbs which appear to have been semantically "bleached." The predicates are incomplete and subcategorize for another predicate (Alsina, 1993; Butt, 1995). I assume that the predicates are incomplete in the sense that the event structure denoted by the light verb is incomplete and must be combined with the event structure of the main verb. While it is beyond the scope of this paper to delve into the semantics of events, the idea of an incomplete event is supported by the fact that the two events involved in causation or permission are cognitively separable and yet difficult to modify individually, as illustrated by the contrast in (18): (18) a. ??anjum = ne saddaf = ko aaj haar Anjum.F = Erg Saddaf.F=Dat today necklace.M = Nom banaa-ne kal di-yaa make-Inf.Obl yesterday give-Perf.M.Sg. 'Anjum yesterday let Saddaf make a necklace today.' b. anjum = ne saddaf = ko aaj haar Anjum.F = Erg Saddaf.F = Dat today necklace.M = Nom banaa-ne = ko kal kah-aa make-Inf.Obl = Acc yesterday say-Perf.M.Sg 'Anjum told Saddaf yesterday to make a necklace today.' Light verbs thus trigger a merger of argument structures which results in a single predicational element, with a single, albeit internally complex, event structure. I now return to the question: Why Argument Fusion? Why postulate a potentially unrestricted formal device as part of complex predicate formation instead of sticking simply to the notion of argument structure merger without the concomitant identification of two arguments? The position taken in this paper is that one needs both concepts, just as both the fundamental syntactic notions of Equi (control) and Raising are needed. In particular, as Neeleman (1994) also observes for Dutch and German consider constructions, an adequate account of the Urdu Permissive must take into account that the permittee in (18a) plays two roles: the recipient (goal) of the permission, and the agent of the embedded action. Note that the Permissive de in fact derives from the ditransitive main verb 'give.' As all goals are marked by dative case in Urdu, including "experiencer" subjects (K. P. Mohanan and T. Mohanan, 1990), the dative case marking

Constraining Argument Merger

85

on the permittee follows immediately from an account allowing for a three place permissive light verb in conjunction with Argument Fusion. It could be argued that in the process of "semantic bleaching" the permissive form might have lost all of its arguments except the permitter. However, this is not supported by the behavior of another light verb derived from 'give,' illustrated in (19): (19) anjum = ne ghar banaa di-yaa Anjum.F = Erg house.M = Nom make give-Perf.M.Sg 'Anjum built a house.' As an Aspectual light verb, de is used to signal telicity, and selects the case marking on the subject (see Butt, 1995). Unlike the Permissive, it does not retain any other arguments. As shown in Butt (1995), the Aspectual light verbs in fact provide further evidence for an Argument Fusion approach, since an alternation between ergative and nominative case marking on subjects of complex predicates formed with these light verbs can be explained straightforwardly under the assumption of Argument Fusion.

4. URDU CAUSATIVES

A further case in point for the adoption of an Argument Fusion approach to complex predicates comes from causatives. In this section, I first detail the rather complicated pattern in Urdu and present the relative strength of an account which allows for Argument Fusion (Alsina and Joshi, 1991; Alsina, 1992). However, the particular formulation of Alsina and Joshi's account raises a question as to the viability of such an essentially unconstrained process. I address some of the formal issues involved and propose an alternative account of Urdu causatives. This account relies on the notion of Argument Fusion as well, but does so in a constrained and systematic fashion. 4.1. The Data Hindi/Urdu has two causative morphemes -aa- and -vaa-, which are historically derived from the Sanskrit allomorphs aya and apaya (Masica, 1991). Saksena (1980, 1982a, b) details a complex system of causativization and analyzes the variation in case marking on causees (accusative ko vs instrumental se) in terms of whether the causee is an affected agent or not. The fact that there are two causative morphemes available in Hindi/Urdu has led some researchers (e.g., Kachru, 1980) to posit two

86

Miriam Butt

stages of causativization, the first represented by -aa-, and the second by -vaa-. However, as Saksena (1982b) shows, the data in Hindi/Urdu do not actually allow themselves to be sorted into such a neat pattern. Rather, some notion of involved vs noninvolved causer, plus a notion of affectedness of the causee must be taken into account. In the data presented below, I abstract away from a number of complications in the system which determine when -aa- vs -vaa- is chosen (e.g., morphophonological reasons, historical gaps, the notion of involved or non-involved causer) and concentrate on the patterns Saksena details with regard to the choice of ko or se case marking on the causee. The basic pattern of causativization is illustrated in (20) and (21).12 In these cases,13 the causees (house, Saddaf) always appear as direct objects and can be either nominative (unmarked) or accusative (ko). This case alternation on direct objects in Urdu is in fact associated with specificity (Butt, 1993b).14 (20) a. makaan ban-aa house.M = Nom be made-Perf.M.Sg 'The house was built,' 'Das Haus entstand.' b. anjum = ne makaan ban-a-yaa Anjum.F = Erg house.M = Nom be made-Caus-Perf.M.Sg 'Anjum built a house.' (21) a. saddaf bhag-ii Saddaf.F = Nom run-Perf.F.Sg 'Saddaf ran.' b. anjum = ne saddaf = ko bhag-a-yaa Anjum.F = Erg Saddaf.F = Ace run-Caus-Perf.M.Sg 'Anjum chased Saddaf away (made her run).' Most transitives display a different pattern. The causee here is not directly affected by the caused action and can only appear as an oblique argument with the instrumental -se. (22) a. anjum = ne paodaa kaat-aa Anjum.F = Erg plant.M = Nom cut-Perf.M.Sg 'Anjum cut a/the plant.' b. anjum = ne saddaf = se/*ko paoda Anjum.F = Erg Saddaf.F = Inst/Acc plant.M = Nom kaat-a-yaa cut-Caus-Perf.M.Sg 'Anjum had Saddaf cut a/the plant.'

Constraining Argument Merger

87

For another set of transitives (other members of this class include drink, learn, see) the agent can be said to be affected by the action, and only a direct (affected) object causee is allowed. (23) a. saddaf = ne khaanaa kha-yaa Saddaf.F = Erg food.M = Nom eat-Perf.M.Sg 'Saddaf ate food.' b.anjum = ne saddaf = ko/*se khaanaa Anjum.F = Erg Saddaf.F = Acc/Inst food.M = Nom khil-a-yaa eat-Caus-Perf.M.Sg 'Anjum made Saddaf eat food (gave Saddaf food to eat).' Finally, there are some transitives which allow both oblique and direct object causees, depending on how "affected" the causee is judged to be from situation to situation (other members of this class are read, write, sing). (24) a. saddaf = ne masaalaa cakh-aa Saddaf.F = Erg spice.M = Nom taste-Perf.M.Sg 'Saddaf tasted the seasoning.' b. anjum = ne saddaf = ko masaalaa Anjum.F = Erg Saddaf.F = Acc spice.M = Nom cak^-va-yaa taste-Caus-Perf.M.Sg 'Anjum had Saddaf taste the seasoning.' c. anjum = ne saddaf = se masaalaa Anjum.F = Erg Saddaf.F = Inst spice.M = Nom cakh-va-yaa taste-Caus-Perf.M.Sg 'Anjum had the seasoning tasted by Saddaf.' The reading that (24b) must receive is that there was a situation in which Anjum wanted Saddaf in particular to taste the seasoning. In (24c), on the other hand, the emphasis is placed on the seasoning: it is important that the seasoning be tasted, whether this is done by Saddaf or some other random person is immaterial. Thus, in (24b) it is Saddaf who is "affected," while in (24c) it could instead be the seasoning. 4.2. Parameters on Argument Fusion The data presented above may appear to come in a bewildering variety, but the account proposed by Alsina and Joshi (1991) for very similar data from Marathi and Chichewa can in fact be extended successfully to the

88

Miriam Butt

Urdu data. This approach actually appears to be the only analysis to date which can account for the full range of data in Hindi/Urdu. Furthermore, it also manages to integrate Saksena's (1982b) original intuition as to the relevance of an "affected agent" into a formal theory of linking. The key to the account posited by Alsina and Joshi (1991) and Alsina (1992) is the assumption of a three place causative in conjunction with the possibility of Argument Fusion. Alsina and Joshi also continue to assume the traditional two place causative for a case alternation which encodes the relative affectedness of the causee in Malayalam causativized unergatives. (25) Kadzldzi a-na-phika maungu. (Chichewa) owl suBJ-PAST-cook pumpkins 'The owl cooked pumpkins.' (26) a. Nungu i-na-phik-itsa kadzldzi maungu. OBJECT CAUSEE porcupine SUBJ-PAST-cookcAus owl pumpkins 'The porcupine made the owl cook the pumpkins.' b. Nungu i-na-phik-itsa maungu kwa porcupine SUBJ-PAST-cookcAus pumpkins by kadzldzi. OBLIQUE CAUSEE owl 'The porcupine had the pumpkins cooked by the owl.' Like Urdu, Chichewa allows a choice in the realization of the causee. Again, the relevant distinction to be made between (26a) and (26b) appears to be in terms of the affectedness of the causee. In (26a), the affectedness of the direct object owl is emphasized; in (26b), on the other hand, the pumpkins must be interpreted as being the most affected argument. Alsina and Joshi analyze these effects in terms of different possibilities for the fusion of arguments. The possible parameters that languages may set are shown in (27). The analyses for (26a) and (26b) are shown in (28a) and (28b), respectively. (27) Parameters The patient of the causative predicate may fuse with a. the logical subject of the base predicate or b. the logical object (affected argument). (28) a. phik-itsa CAUSE < ag pt COOK < ag pt > > (OBJECT CAUSEE) COOk-CAUS

b. phik-itsa CAUSE < ag pt COOK > (OBLIQUE CAUSEE) COOk-CAUS

Constraining Argument Merger

89

In (28a), the causee is also the patient of the cause predicate. Hence, it qualifies as an affected causee and is realized as an object. In (28b), on the other hand, fusion has taken place according to Parameter b of (27). The causee here is now not also simultaneously the patient of causation, so that it is realized as an oblique argument and the pumpkins surface as the direct object. Note, however, that there is an inconsistency in the use of the three place causative. In (28a) the causer acts upon the patient of causation in order to bring about a certain event. That is, (26a) is reasonably interpreted as describing a situation in which the porcupine acts upon the owl, so that the owl then goes and cooks the pumpkins. In (28b), on the other hand, the causer porcupine acts upon the patient of causation, which is also the patient of the embedded predicate, the pumpkins. That is, according to (28b), (26b) should actually be interpreted as the porcupine acting upon the pumpkins directly in order to bring about their cooked state, leaving the owl, the actual cooker of the pumpkins, somewhat in limbo. This point is actually indicative of a larger, technical problem with the theory of argument composition presented by Alsina and Joshi. The postulation of Parameters on Argument Fusion opens the door to any kind of combinations of arguments: one could imagine additional parameters sensitive to volitionality, benefaction, or experiencerhood. Such semantic factors governing the parameters on argument fusion preclude the formulation of any restrictive theory of complex predicate formation. Furthermore, while the notion of affectedness plays a large role, it is not defined in a satisfactory manner. What does it mean to be affected? In fact, as much recent work on the interaction between aspect and argument structure has actually focused on just this question (e.g., Tenny, 1987; Krifka, 1992), the investigation of a connection to possible aspectual influences in causativization is not only desirable, but necessary. Before moving on to a reanalysis of Urdu causatives, I address the issues raised by Argument Fusion and resolve the technical problems that may be brought up with respect to such a seemingly nonmonotonic addition to an otherwise monotonic framework. I then go on to present a very restrictive view on Argument Fusion which finds a parallel in the notion of syntactic control. 5. FORMAL CONSIDERATIONS 5.1. Linking

An immediate objection to the notion of Argument Fusion can be made from both a formal and a semantic point of view in that the notion may be

90

Miriam Butt

incoherent with respect to both of these areas. However, I argue that Argument Fusion must actually be understood as the argument structure analog to syntactic control, and be viewed as an instance of the type of ©-identification proposed by Higginbotham (1985). The notion is coherent and formally unproblematic under this view. Both here and in the bulk of the literature on argument structure, its representation has been in the form of a list of roles. Consistent with general LFG architecture, argument structure can, however, alternatively be viewed as an Attribute-Value Matrix (AVM). In this paper, I propose a representation which naturally integrates the principles of linking theory as formulated in LFG (Bresnan and Kanerva, 1989; Alsina and Mchombo, 1993; Bresnan and Moshi, 1990; Bresnan and Zaenen, 1990).15 The linking theory relies primarily on the Thematic Hierarchy shown in (29), which is based on argumentation from Kiparsky (1987). (29) Thematic Hierarchy: agent < beneficiary < experiencer/goal < instrument < patient/theme < locative The thematic arguments are associated with the features [±o(bject)] and [+r(estricted)J. These features in turn provide the constraints for the mapping from a(rgument)-structure to f-structure. The correspondence between grammatical functions and features is shown in (30): (30) Grammatical Functions SUBJ OBJ OBJ0 OBL0

Features [ -r, — o] [ -r, + o] [+r, + o] [+r, — o]

The association of the features with thematic roles is governed by two classification systems: Intrinsic and Default Role Classifications, whereby the Intrinsic Classification has come to be viewed as part of the argument structure representation in more recent developments within linking theory (Bresnan and Zaenan, 1990; Alsina, 1993): Intrinsic Role Classifications Patientlike roles: [—r] Secondary patientlike roles: [+o] All others: [ —o] Default Role Classifications Highest thematic role: [-r] All others: [+r]

Constraining Argument Merger

91

The essential function of the features is to ensure that agents are usually mapped onto subjects ([ —o]), while patients and themes ([ —r]) are mapped onto objects. Thus, the [±o] feature sets the subject apart from other arguments by marking all the roles that might be mapped onto objects. The [+r] feature refers to the restrictedness of the semantic interpretation of a given grammatical function: subjects and direct objects are not restricted in terms of the thematic roles they may be associated with, while indirect objects and obliques do have a semantically restricted interpretation. Now, given that the initial set of classificatory features is considered to be intrinsic, i.e., part of the predicate-argument structure, a reasonable rendering of an argument structure representation as in (31), taken from Bresnan and Zaenen (1990:48) could be as in (32).

(31)

(32)

While linking theory was originally viewed as applying to lexical items within the lexicon, it has been successfully extended to apply beyond the lexicon (e.g., Bresnan and Zaenen, 1990, for an account of resultatives). In particular, it can also be applied to dynamic complex predicate formation in the syntax, as argued for in Butt (1995) and Alsina (1993,1996) for Urdu and Romance, respectively.16 The argument structure of the Permissive and Argument Fusion between the permittee and the agent of the embedded predicate can thus be rendered as in (33):17

(33)

92

Miriam Butt

Argument Fusion is thus taken to be the a-structure analog to functional control, the formal properties of which are well defined and understood within LFG (Kaplan, 1995; Kaplan and Bresnan, 1995). Note that the embedded agent role in (33) does not have an inherent specification. In addition to the [ -o] option, which functions as the elsewhere case for the determination of intrinsic features, I assume that, when a role is subject to control as a-structure as in (33), it may not carry a separate specification. 5.2. Parallelism to Syntactic Control and Raising Having arrived at a formal understanding of Argument Fusion, the next immediate question which must be addressed is the potentially unrestricted nature of the combinatory possibilities that become available. If, as in Alsina and Joshi's approach to causatives, any argument from one a-structure could potentially be fused with any argument from another embedded a-structure, is it still possible to formulate a restrictive theory which allows the statement of strong and potentially falsifiable predictions? The answer here must clearly be in the negative, as any number of parameters in addition to the affected argument of Alsina and Joshi (1991) may be subsequently introduced. In this section, I put forward some very strong restrictions on the combinatory possibilities of argument structures. As suggested above, I propose that complex predicate formation at argument structure is analogous to (and probably historically derived from) syntactic control and raising (see Sells, 1993, for a similar idea). Accordingly, only the highest embedded argument should ever be available for fusion. This restriction is also in accordance with Neeleman's (1994) structural approach to complex predicates, in which he motivates 0-percolation only for the highest 0-role (based on independent argumentation from Williams, 1980). In our terms, the restriction argued for by Neeleman can be formulated as in (34): (34) Restriction on Argument Fusion: Only the highest 0-role may escape its domain of predication, and thus become eligible for Argument Fusion. Parallel to syntactic control and raising, I thus admit the following two combinatory possibilities for argument structures. Argument Fusion. The lowest matrix argument Argument Raising. The with the arguments of

highest embedded argument is fused with the (analogous to syntactic control (PRO)). arguments of the matrix a-structure are unified the embedded a-structure (cf. Grimshaw and

Constraining Argument Merger

93

Master's, 1988, Argument Transfer for Japanese sum, Monachesi's, (1995), use of Argument Composition within HPSG for Romance restructuring verbs). Both Argument Fusion and Raising are amply documented. The Urdu Permissive, for example, is an instance of Argument Fusion. Neeleman (1994) argues for a complex predicate, rather than a small clause, analysis for Dutch and German consider-constructions and thus provides a welldocumented example of Argument Raising, as well as positing the equivalent of Argument Fusion as defined here for Dutch resultatives. The table in (35) summarizes the correlations between complex predicate formation, argument structure combination, and syntactic control and raising:

(35) Syntax (f-structure) a-structure

Control PRO controlled Argument controlled (fusion)

Raising Exceptional Case Marking Arguments unified (raising)

Complex predicate No Yes

Note that although I cite the notions of Argument Transfer and Argument Composition as being closely related to the notion of Argument Raising, the phenomena for which these analyses were originally proposed do not necessarily provide well-documented instances of Argument Raising as formulated here, i.e., as leading to complex predicate formation (see (11)). Matsumoto (1996), for example, strongly challenges an Argument Transfer analysis for Japanese sum. Similarly, it is not clear whether Romance restructuring verbs should be viewed as an instance of Argument Raising, or rather, as argued in Butt (1995), as an instance of Argument Fusion. The phenomena of German verbal complexes for which Argument Composition was originally introduced (Hinrichs and Nakazawa, 1990), in fact, show no evidence of being instances of complex predicate formation (see Rambow, 1996). Also note that I make a very clear distinction between what I have described as complex predicates in terms of (11), and further constructions such as applicatives, passives, or serial verbs. While these constructions must also involve argument structure combinations, there does appear to be a difference as to the kind of combination. Applicatives and passives involve argument addition and reduction, respectively, while serial verbs are most commonly characterized by object sharing (Baker, 1989) and

94

Miriam Butt

therefore are not subject to the restriction in (34). It appears that the crucial factor differentiating the constructions must be stated in terms of event structure: passives and applicatives only involve one event, unlike the Urdu Permissive, or causatives, in which there is an event of causing-letting and the caused-permitted event. Serial verb formation, on the other hand, seems to be governed by restrictions on complex eventbundling (Durie, 1993). For example, in Sranan buy take fish is a good "normal-expected" complex event, but sell take fish is not. This issue of differences in event structure needs to be explored in much more detail and remains subject to further research. Within the scope of this paper, I now move on to a reexamination of Urdu causatives and propose an analysis which builds on the original insights of Alsina and Joshi (1991) and Saksena (1982b), but avoids the unconstrained nature of Alsina and Joshi's approach by incorporating the restricted notion of Argument Fusion presented above.

6. CAUSATIVES, ASPECT, AND AFFECTEDNESS 6.1 Interactions

An alternative approach to the notion of affectedness, which both Saksena (1982b) and Alsina and Joshi (1991) demonstrated to be directly relevant for an analysis of causatives, is offered by the literature on clausal aspect. The notion of affectedness has been shown to play a large role in the determination of clausal aspect (e.g., Tenny, 1987), and considerable effort has been put into formalization over the years (e.g., Dowty, 1979; Verkuyl, 1972; Krifka, 1992). Interestingly enough, other recent work on complex predication has noted an interaction between aspect, affectedness, and the licensing of arguments (e.g., English have (Ritter and S. Rosen, 1993) and Chinese resultatives (Li, 1995)). The accounts posited for these phenomena independently arrive at very much the same conclusions: they suggest that the licensing of arguments in the syntax is determined through an interaction of aspect with an underspecified thematic structure (compare Grimshaw's, 1990, Thematic and Aspect Tiers). While these approaches provide interesting evidence for the role of aspect in complex predication, a precise integration of aspectual information into a theory of projection (linking in our terms) has not so far been undertaken. This is not surprising, as the representation of aspectual information within the syntax is as yet not well understood. However, this

Constraining Argument Merger

95

is precisely the issue taken on in Ramchand (1997). Based on data which illustrate an interaction between aspect and argument structure in Scottish Gaelic, Ramchand extends Krifka's (1992) system of role types (defined primarily for verbs of creation and consumption) to account for a broader field of predication. This system is further combined with a theory in which verbs provide an underspecified 0-grid whose arguments are licensed in the syntax through roles assigned by the aspectual head of the construction (situated in AspP). Note that this is exactly within the spirit of the approach argued for independently by Li (1995) and Ritter and S. Rosen (1993). In particular, Ramchand also sketches an analysis for Urdu causatives, in which the affectedness of causees is accounted for through the interaction between aspect and argument structure. The crucial difference between Ramchand's approach and the analysis presented in Alsina and Joshi (1991) and Alsina (1992) is that Ramchand does not rely exclusively on lexical information provided by the predicates, but rather has lexical (thematic role) information interact with structurally determined aspectual roles. Ramchand, in fact, maintains that an approach within linking theory, where lexical information is placed in correspondence with grammatical functions cannot account satisfactorily for the full range of interactions between case marking, syntactic position, and semantic interpretation of argument NFS. In this section, I propose to show that lexically determined thematic information can be placed into a relationship with clause-based considerations of aspectual interpretation within the already existing LFG formalism, without necessarily presupposing an explicitly structural approach as that advocated by Ramchand, or assumed by Ritter and S. Rosen (1993). As LFG assumes an architecture of grammar in which several differing levels of representation are independent, but copresent in the sense that they are explicitly related to one another by the projection architecture (Kaplan, 1995; Kaplan and Wedekind, 1993), a very natural analysis can be developed for Urdu causatives in which the linking of lexical thematic information to grammatical functions is also sensitive to aspectual information encoded at a different, semantic, level of representation. 6.2. Lexical and Compositional Aspect It has long been noted that verbs can be classified into several different types. The verb break, for example, entails a result state in which some object is broken where it previously was not, while a verb like run simply denotes an activity without necessarily entailing a resultant state. Vendler's (1967) classification of verbs in terms of the Aktionsart types state, activity, achievement, accomplishment, along with the very careful tests Dowty

96

Miriam Butt

(1979) formulates have come to represent a standard in examinations of verbal aspect. However, it has also long been recognized that verbal aspect cannot be based on the lexical semantics of the verb alone, but must take into account clausal influences such as adverbials, quantifiers, or the semantics of NP arguments (Verkuyl, 1972). A well-known example with respect to the latter is the alternation between (36a) and (36b), where the former must be interpreted as an achievement, and the latter as an activity. (36) a. Yassin ate an apple (*for an hour)/(in an hour). TELIC b. Yassin ate apples (for an hour)/(*in an hour). ATELIC Verkuyl (1993) and Krifka (1992) have proposed an explicit semantics for the interaction between verbal predication and argument NPs, with special emphasis on verbs of creation and consumption. The basic intuition behind their respective formalisms is a mapping of created or consumed objects onto a spatial-temporal axis. Notions like affectedness are then captured by constructing a mapping between an event denoted by a verb and its object. While a precise characterization of the respective semantic formalisms employed by Krifka and Verkuyl goes beyond the scope of this paper, the underlying intuition can be summarized roughly as follows. Internal arguments are mapped onto the event structure of a predicate. In (36a), every subpart of the event can be seen as corresponding to a subpart of the apple being consumed: the event is telic, and the object apple is affected. In (36b), on the other hand, such a strict mapping between events and objects is not possible, and a different set of event to object correspondences captures the atelic nature of the event. 6.3. Urdu Causatives Revisited The affectedness effects found for causees in Urdu do in fact interact with aspect. In (37) the causee is realized as a direct accusative object. If the causee here is indeed "affected" in the aspectual sense of Krifka (1992), the construction as a whole should be telic and therefore incompatible with a durative adverbial like for a long time, but compatible with a time-span adverbial like in ten minutes. As (37a) and (37b) illustrate, this is indeed the case. (37) a. *anjum = ne xansaamo = ko masaalaa Anjum.F = Erg cook.M.Pl = Acc spice.M = Nom barii dheer = ke liye cakh-va-yaa big time = for taste-Caus-Perf.M.Sg 'Anjum had the cooks taste the seasoning for a long time.'

Constraining Argument Merger

97

b. anjum = ne xansaamo = ko masaalaa Anjum.F = Erg cook.M.Pl = Ace spice.M = Nom das mint = me cakh-va-yaa ten minutes = in taste-Caus-Perf.M.Sg 'Anjum had the cooks taste the seasoning in ten minutes.' In (38), the causee is realized as an oblique. Under a reading where the causee is not affected, i.e., is not implicated at each stage of the caused event, the tasting event should be atelic. Recall that the oblique cook here may be only one of many tasters. As such, this construction should be compatible with both the durative adverbial in (38a) and the time-span adverbial in (38b). Again, the data bear out the prediction.18 (38) a. anjum = ne xansaamo = se masaalaa Anjum.F = Erg cook.M.Pl = Inst spice.M = Nom barii dheer = ke liye cakh-va-yaa big time = for taste-Caus-Perf.M.Sg 'Anjum had the cooks taste the seasoning for a long time.' b. anjum — ne xansaamo — se masaalaa Anjum.F = Erg cook.M.Pl = Inst spice.M = Nom das mint = me cakh-va-yaa ten minutes = in taste-Caus-Perf.M.Sg 'Anjum had the cooks taste the seasoning in ten minutes.' Having established the interconnectedness between aspect, affectedness, and case marking in Urdu causatives, I now offer an alternative view of linking, which allows an incorporation of aspectual considerations, while remaining true to the constrained theory of complex predicate formation articulated above. 6.4. Rethinking Linking 6.4.1. THEMATIC ROLES A strong view of argument structure is that the information represented at this level must be considered to be of a semantic nature, and that a statement of semantic entailments ought to be possible at this level. This expectation is exactly what has given rise to the general dissatisfaction with the use of thematic roles as a representation of argument structure. As Dowty (1991) points out very forcefully, there is simply no general consensus as to which and how many thematic roles there ought to be. It is also not clear exactly what it means to be an agent vs an instrument, or a patient vs a theme.19 As a realistic alternative, Dowty formulates the notion of protoroles, which allows the properties of a given argument to vary over

98

Miriam Butt

a collection of semantic entailments, thus allowing greater flexibility in the assignment of thematic roles, and dispensing with the need to invent a new role for each slightly differing collection of semantic entailments. Alsina (1993) imports Dowty's protoroles in a version of LFG's Linking Theory and reformulates the linking principles accordingly. In this paper, I propose to take a different approach and follow Zaenen (1993:147) in assuming that lexical meaning cannot always be characterized directly in terms of entailments. Zaenen takes the collection of Dowty's proposed agent and patient properties and uses them simply to determine whether a given argument's intrinsic feature classification will be [— o] (protoagent), or [—r] (protopatient), without worrying about semantic entailments until semantic interpretation at the level of the clause. Zaenen (1993) dispenses with a hierarchy of thematic roles. However, as she is primarily considering intransitives, this is easy to do. I propose to adopt the reduced thematic hierarchy in (39), which is essentially the hierarchy put forward in Kiparsky (1987). (39) Reduced Thematic Hierarchy: agent < good < instrument < theme < locative In comparison to the standardly assumed hierarchy in (29) within LFG's Linking Theory, I take goal to be the prototypical instantiation of beneficiary and experiencer, and theme as the label for a protopatient. The purpose of the hierarchy continues to be as standardly assumed: to rank the arguments of a predicate. The role names, however, as already argued in Grimshaw (1990), must be seen as abstract labels attached to a ranked list of arguments. I could try to dispense with the labels altogether, but as Grimshaw's (1990) attempt in this direction showed, the labels are not only handy, but also allow distinctions that cannot be achieved through a mere employment of alphabetical items such as x, y, or z. The thematic roles are viewed as fulfilling a prototypical role in a given action schema, but there is no direct evaluation of the roles in terms of semantic entailments. Rather, the parameters involved, such as affectedness, volition (with agents), or instrumentality, are properties which can be shown to have a direct syntactic influence, for example in terms of case marking or grammatical function realization. These syntactic realizations may, in turn, give rise to differing semantic interpretations. However, the semantic input from the lexical properties of a predicate takes an indirect, rather than a direct form. The primary purpose of argument structure, under this view, is to help determine the syntactic realization of a predicate's arguments in terms of grammatical functions and case marking. In this sense, then, argument structure must be viewed as a primarily syntactic level of representation.

Constraining Argument Merger

99

6.4.2. ASPECTUAL INTERPRETATION OF INTERNAL ARGUMENTS Given that aspectual considerations appear to influence the mapping from a-structure to f-structure, as evidenced by the Urdu causatives, how should aspectual notions be integrated? Grimshaw (1990) sketches a proposal for a Causation Hierarchy based on aspectual dimensions in addition to a Thematic Hierarchy, and explicitly assumes an interaction between the two. Similarly, Ramchand (1997) proposes that the projection of arguments into the syntax must take into account both an underspecified thematic hierarchy provided by the verb and a system of aspectual roles provided by the aspectual head of the clause. Based on Krifka's and Verkuyl's pioneering work centering on verbs of creation and consumption, Ramchand (1997) defines a number of differing aspectual roles for internal arguments,20 for a broader spectrum of verb types, and details their interaction with basic verb types (e.g., states, activities) to yield the aspectual interpretation of the entire clause compositionally. The differing aspectual roles express the nature of the mapping that has taken place (parts of the object mapped onto each stage of the event for creation or consumption verbs; locations of an object mapped onto stages of the event for change of location verbs, etc.). Ramchand also distinguishes between two basic types of internal arguments: those which enter into a calculation of compositional aspect (©,„,), and those which are aspectually inert (®mod). A number of different types of aspectual arguments are proposed and formally defined for result predicates such as break, predicates of movement such as push, and predicates of consumption or creation such as eat. The precise distinctions made do not interest us here. Relevant is the very explicit claim made that notions such as "affectedness" can only be introduced by an aspectual role(0//u), not by an aspectually inert one (®mod). This follows from the formal difference in how the two types of objects are taken to enter into the calculation of the semantic interpretation. In (40), the object is represented as an explicit argument of the predicate, as an individuated variable. In (41) the object is taken to be related to the predicate through a kind of modification (cf. de Hoop's, 1992, predicate modification). The crucial distinction is that only in (40) can the object be related to a time index of the verb and allow a mapping from object to event structure. In (41) there is no explicit representation of the object as an entity that may be related to the event structure of the verb, so the object is aspectually inert. (40)

(41)

100

Miriam Butt

This abstract distinction may become easier to follow through the presentation of two examples from Scottish Gaelic, which very clearly support the essential argument being made.21 In (42a) and (43a), the direct object is marked with direct case, occurs preverbally, must be interpreted as being affected, and gives rise to a telic reading of the sentence. (42) a. Tha Calum air na craobhan a ghearradh. be-Pres Calum Asp the trees-Dir OAgr cut-VNOUN 'Calum has cut the trees.' b. Tha Calum a' ghearradh nan craobhan. be-Pres Calum Asp cut-VNOUN the trees-Gen 'Calum is cutting the trees.' (43) a. Tha mi air am ball iarraidh. be-Pres I Asp the ball-Dir want-VNOUN 'I have acquired the ball.' b. Tha mi ag iarraidh a'bhuill. be-Pres I Asp want-VNOUN the ball-Gen 'I want the ball.' In (42b) and (43b), on the other hand, the direct object appears postverbally in the genitive, cannot be interpreted as being affected, and the sentence has an atelic reading. In the (a) examples the object is therefore taken to be aspectually relevant, as expressed in (40); in the (b) examples, it must be analyzed as being aspectually inert, as expressed in (41).

6.4.3. LINKING In some languages, like Scottish Gaelic, the contribution of the aspectual head can clearly be shown to be distinct from that of the predicate. In other languages, like English, these two sources of information appear to be collapsed into the verbal predicate. Precisely where the information may be encoded does not, however, detract from the fact that thematic information is provided exclusively by the lexical predicate, while aspectual information is sensitive to a collection of information at the clause level. Aspectual information thus should not be encoded as a direct part of linking theory. However, I would like to adopt the fundamental distinction made by Ramchand between aspectually relevant and aspectually inert arguments. In particular, I would like to take into account the fact that only the former can give rise to an interpretation in which the object is affected. The need for a special, stipulative, feature such as [+affected] can thus be avoided at a-structure.

Constraining Argument Merger

101

Looking back at the specifications for the intrinsic classification of 0-roles, repeated below, note that these primarily seek to make a distinction between differing types of internal arguments. In particular, the distinction between patientlike [—r] roles and secondary patientlike [+o] roles was primarily posited to account for the rather complicated pattern of double objects in applicative Bantu constructions. While I believe that the analyses presented for these constructions should also ultimately be reanalyzed within the approach adopted here, such a reanalysis is clearly beyond the scope of the paper. Intrinsic Role Classifications Patientlike roles: [-r] Secondary patientlike roles: [ +o] All others: [—o] The intrinsic property of [ +o] for these applicative objects can simply be retained for the moment, as it does not interact directly with the particular set of data examined here. None of the following proposals for the linking of objects in causatives thus has an immediate impact on the analyses presented for Bantu in Bresnan and Kanerva (1989), Bresnan and Moshi (1990), or Alsina and Mchombo (1993). What I propose is to make a finer distinction between the set of "patientlike" roles, basically themes in terms of the reduced thematic hierarchy in (39), and to incorporate the idea that some arguments may be aspectually contentful, while others are aspectually inert: Revised Intrinsic Classifications Themes: [—r] (aspectually inert) or [+r] Secondary patientlike roles: [+o] All others: [ -o] This slight amendment to the linking theory has the effect of linking object themes to either OBJ ([— r]) or OBJS ([+r]). Recall that the [-r] in linking theory was intended to indicate semantic unrestrictedness, while [+r] indicates that the object is restricted to express some particular semantic information. Standardly, this type of information has been expressed in terms of thematic roles, so that one could differentiate between simple OBJS, OEJbeneficiary, or OBJgo. Thus, the semantically restricted object was annotated to indicate which precise thematic role it was linked to. Alsina (1993) argues very clearly that this method of labeling semantically restricted objects is redundant. Given the architecture of LFG, where the various levels of representation are copresent and related to one another via the projection architecture it is simply redundant to express information already represented at a-structure again at f-structure, i.e., in terms of

102

Miriam Butt

grammatical functions like OBJO. Alsina thus dispenses with the distinction and allows only OBJS. I am essentially in agreement with Alsina's line of argumentation, but retain the notational distinction between OBJ and OBJS at f-structure for ease of exposition. I assume along with Ramchand that objects come in two differing general varieties: those that contribute a semantics to the compositional semantic interpretation of the clause and those that do not (aspectually inert). These are taken to correspond to [+r] (OBJS) and [—r] (OBJ), respectively. Note, however, that it is not the thematic roles themselves which contribute the semantics of the object. Rather, the [+r] of a semantically restricted OEJO simply indicates that this is an object which may enter into the compositional semantics of the clause. The precise semantic interpretation of the clause is not represented at a-structure, or f-structure, but at s(emantic)-structure (see Halvorsen, 1983, or Dalrymple et al., 1993, for some proposals on the representation of semantics in LFG) and the interaction between argument structure, the syntactic realization of arguments, and the interpretation at s-structure is modeled through LFG's projection architecture. A simple example of a contrast is given in (44) and (45). (44)

Yassin ate an apple.

(45) a-structure EAT Default GF

(46)

Yassin ate apples.

(47) a-structure EAT Default GF

In both cases the theme could carry either a [+r] or [—r] feature. However, only one of the feature realizations can lead to a well-formed semantic interpretation: in the atelic case in (47) it is the OBJ; in the telic sentence in (45) it is the OBJe.

Constraining Argument Merger

103

7. AN ALTERNATIVE ANALYSIS OF CAUSATIVES At this point, the necessary ingredients are in place for an analysis of the Urdu causative that is also in keeping with restrictive theory of complex predicate formation. With Alsina and Joshi (1991), I assume both the three place and two place predicate for causatives shown in (48):

(48) a. CAUSE < ag th PRED < ag... > > b. CAUSE < ag PRED < ... > >

The two place causative conies into play when the causer does not act upon somebody to cause an event, as in a three place causative, but initiates the action directly.22 In (48a) Argument Fusion applies and must obey the principle in (34) so that only the highest embedded argument fuses with the lowest matrix argument. The linking from arguments to grammatical functions obeys the revised intrinsic classifications outlined in section 6.4.3. The realization of a given argument as a given grammatical function must ultimately be compatible with the aspectual interpretation of the clause. This interaction can be taken to explain the interesting alternation observed between the causee as object or oblique for certain transitives in Urdu and avoids both a stipulative reference to affectedness at a-structure, and the concomitant assumption of parameters on Argument Fusion. (49) Object Causee anjum = ne saddaf — ko masaalaa Anjum.F = Erg Saddaf.F = Acc spice.M = Nom cakh-va-yaa taste-Caus-Perf.M.Sg 'Anjum had Saddaf taste the seasoning.'

(50) a-structure CAUSE Default

GF Case

104 (51)

Miriam Butt

Oblique Causee anjum = ne saddaf — se masaalaa( = ko) Anjum.F = Erg Saddaf.F = Inst spice.M = Nom( = Ace) cakh-va-yaa taste-Caus-Perf.M.Sg 'Anjum had the seasoning tasted by Saddaf.'

(52) (52) a-structure CAUSE Default

GF Case

Case marking in Urdu interacts with semantic interpretation as well. As shown in Butt (1993b), the accusative ko marks specific objects in Urdu. Being specific, these objects are semantically restricted and must be realized as OBJ0, else semantic interpretation at the clause level will be ill-formed.23 Nominatives in Urdu are compatible with either a specific or nonspecific interpretation. In (50) there are two linking possibilities for both of the themes. However, only one of these possibilities is ultimately compatible with the necessary semantic aspectual interpretation of the clause, and only that possibility is compatible with the case marking on the objects: the causee must be realized as an accusative OBJe and hence become eligible to be interpreted as the affected object of the clause. This leaves the 'spice' to be linked to OBJ. In (52), on the other hand, the causee does not enter the aspectual semantics, and is realized as an oblique. This leaves the 'spice' to be realized either as an OBJ or an OBJ0, depending on the aspectual interpretation (if the 'spice' is to be considered affected, it must be realized as an OBJ0). Thus the alternation in the case marking on causees and the interaction with the notion of "affectedness" now follows. In (50), but not in (52), the causee interacts with the aspectual semantics of the clause. The other cases in the causative data presented above also follow as simply from Argument Fusion, the three place causative, and the interaction of aspect and argument structure. In particular, the class of "ingestives" always

Constraining Argument Merger

105

requires an accusative direct object causee because the causee there has no option of being other than affected by the matter that is ingested.

8. PASSIVES AND UNSPECIFIED OBJECTS The above discussion should have demonstrated very clearly that I do not take OBJ and OEJ0 to differ in terms of syntax. Along with Alsina (1993), who abolishes the distinction at f-structure, I take the two types of objects to have the same general syntactic properties. There is, however, one area where the two types of objects have traditionally been taken to differ: [+r] objects (OBJ0) are precluded from undergoing passivization, since these could never be linked to a subject ([—r, — o]). The question which arises is whether the current proposal with respect to OBJ vs OBJe makes the prediction that only clauses containing an OBJ, but not an OBJe can be passivized. This is by no means the case. Passivization with the advent of Linking Theory no longer applies directly to grammatical functions, but to the highest argument at a-structure. This argument is simply suppressed. (53) Passive

Given that most themes will always have either the [—r] or the [+r] option for linking, if passivization applies, it can only apply felicitously to the [ —r] option. Given that the theme argument in this case is not linked to an object, and thus cannot enter aspectual semantics as an object, it is only logical that the [+r] option should never be chosen. The interaction of subjects (external arguments) with aspectual semantics must also be taken into account, but is beyond the scope of this paper (see Ramchand, 1997, for a discussion).24 Finally, the distinction drawn here between OBJ and OBJ0 may be useful in terms of the unspecified object deletion observed for Bantu in Alsina and Mchombo (1993), which was accounted for by (54): (54)

Theme Suppression

106

Miriam Butt

The observation that unspecified objects may be optional follows in my account simply from the idea that OBJ, as opposed to OBJ^, is nonspecific because it is aspectually inert, i.e., does not carry any semantic aspectual information with it which may enter into the compositional semantics of the clause. Thus, in parallel with the affectedness effects observed for causatives, the deeper semantic explanation needed for unspecified object deletion is squarely placed at the level of representation most suited to encode it: s(emantic)-structure. At a-structure and f-structure the option between two differing types of objects is merely made available, but not evaluated.

9. CONCLUSION The analysis of Urdu causatives presented here takes a step beyond the original insights formulated on the basis of Clause Union. For one, the analysis is in keeping with a restrictive theory of complex predicates in which the pertinent characteristics of a complex predicate are very clearly defined, and in which restrictions on argument structure composition parallel the restrictions found for syntactic control and raising constructions. The clear formulation of a theory of complex predicates also provides a firm basis for the investigation of similar constructions cross-linguistically, and for a comparison to other constructions, such as applicatives and serial verbs, which involve argument structure processes, but differ in their overall event structure. The recognition of argument structure as a syntactic level of representation in conjunction with a consideration of aspectual semantics has allowed an alternative view of mapping and linking. In this approach, it is not attempted to code semantic notions such as affectedness or (non)specificity at the already semantically incoherent level of argument structure. Rather, semantic interpretation of a clause is taken to be represented at s(emantic)-structure, and an interaction between semantics and the differing linking possibilities for objects is defined. The encoding of semantic notions at the level where they can be most usefully dealt with, at s-structure, allows a clearer analysis of case alternations on causees, as observed in Urdu causatives. In particular, a very constrained version of Argument Fusion can now apply within a larger theory of complex predicate formation, and allows a principled reformulation or loosening of statements such as Function-Argument Biuniqueness or the Theta Criterion, whose strict requirements on a one-to-one correspondence between

Constraining Argument Merger

107

predicate-argument structure and the syntactic realization of arguments do not allow a lucid treatment of complex predication.

NOTES 1

I particularly thank Alex Alsina, Joan Bresnan, Mary Dalrymple, Tracy King, Gillian Ramchand, and members of the LFG-list for many interesting discussions and input, Tilman Hohle and Dieter Wunderlich for asking skeptical stimulating questions, and two anonymous reviewers for further helpful comments and criticisms. 2 See also Evers (1975) for an analysis of German coherent constructions in terms of Clause Union within an early transformational approach. 3 The South Asian language Urdu is closely related to Hindi, which is mostly spoken in India. In this paper, the data are primarily drawn from the dialect of Urdu spoken in Lahore, Pakistan, and that of Hindi-speaking informants from New Delhi, India. 4 Contrary to a reviewer's comment that stipulations about transitivity may be needed in all argument structure approaches, the analysis proposed by Alsina (1993) involving argument fusion and LFG's linking theory overcomes a number of disadvantages of previous approaches and does not need to stipulate anything special for transitive vs intransitive embedded predicates (see section 2.4). It is beyond the scope of this paper to summarize the rather extensive and detailed argumentation of Alsina (1993, 1996); however, I should note that examples in which intransitive vs transitive versions of the same verb, e.g., manger 'eat,' are embedded under a causative by no means present a counterexample to an argument structure approach. There is nothing inherent in the argument structure approach which prevents one from modeling optionally intransitive predicates at argument structure. As will become clear in the section on aspect, I do not subscribe to the view that argument roles enter directly into a process in which semantic entailments are calculated. Argument structure under the view presented in this paper is very much a syntactic level of representation. 5 Alsina (1995) goes even further and proposes to abolish even this unidirectional constraint based on data from Catalan reflexives, but this more extreme position is not taken in this paper. Note that the RG approach does not constitute an instance of Argument Fusion because there is no instance in which a given argument is introduced by more than one predicate. Contrary to a reviewer's reading of (3), despite the fact that las herramientas 'tools' in (2) ends up being represented as the 2-argument of the complex predicate, it is only introduced to the unified clause by the embedded predicate, not by the causative hacer 'make.' This is in marked contrast to the three place causative adopted by Alsina. 7 Abbreviations used are as follows: F = feminine; M = masculine; Erg = ergative; Nom = nominative; Gen = genitive; Dat = dative; Ace = accusative;

108

Miriam Butt

Inst = instrumental; Loc = locative; Inf = infinitive; Obl = oblique; Perf = perfect; Impf = imperfect; Stat = stative; Pron = pronoun; Sg = singular; P1 = plural. A "-" indicates a morpheme boundary, while an " = " separates a clitic from a lexical item. 8 The nominative case marker in Urdu is phonologically null. 9 For details on infinitive agreement see Butt (1993a, b). Infinitives generally involve what appears to be "long-distance agreement," but which must actually be analyzed as an instance of successive local agreement. The agreement facts of the Permissive cannot be analyzed similarly. 10 The f-structure is abbreviated in that it displays only some basic information. A complete f-structure would list values for number, case, gender, tense, etc. n See section 3.3 on why the second argument of the Permissive is rendered as a goal in (13). 12 It appears that in the grammar of modern speakers, the "causativized" form in the (b) examples is actually not registered as such: the complex predicates have been reanalyzed as simple transitives (see also Saksena, 1982b, on the lexicalization of such forms). 13 Note that the passive English gloss given for (20) is not entirely accurate: the ban in fact has no passive interpretation, as is illustrated by the German gloss. An English analog is a verb such as appear. 14 The examples presented here are all past tense, realized through "perfective" morphology. Urdu is a split-ergative language in which the ergative is roughly correlated with unergatives, transitivity, and perfective morphology. 15 For recent reformulations of the linking theory along the lines of Dowty's (1991) protoroles, see Alsina (1993, 1996) and Zaenen (1993). 16 But see Frank (1996) for an alternative view, where argument structure composition of syntactically independent items is carried out within a version of the lexicon that is taken to emcompass more than the encoding of lexical and sublexical properties. 17 A goal in Urdu as inherently classified as [ +0], unless it appears as the highest argument. In the latter case it is marked as [ —o]. This allows a mapping to subject in experiencer constructions, and an object argument in ditransitives. 18 Note that in (38) the object masaalaa 'spice' may alternate between nominative and accusative case. If it is marked with the accusative ko, the spice must be interpreted as "affected" and is incompatible with the durative adverbial. In (37) the object 'spice' may not be subject to such alternation since there is a restriction in Urdu/Hindi against more than one ko marked object per clause (T. Mohanan, 1994b). 19 Dowty is, of course, not the only one to express the general unhappiness with thematic roles. Other succint, recent discussions include (but are not limited to) Grimshaw (1990), Jackendoff (1990), and the collection of papers in Butt and Geuder (to appear). 20 Roles for external arguments are defined as well, but go beyond the scope of this paper.

Constraining Argument Merger

109

21

The examples are taken from Ramchand (1997:93-94). The abbreviation Dir stands for "direct case," OAgr for "object agreement," and Asp for "aspectual marker." The latter is my gloss. 22 The assumption of a two place causative is not necessary for an analysis of the Urdu causative paradigm presented in section 4. However, it may be needed for Romance examples such as les dieux peuvent faire pleuvoir 'The gods can make it rain' brought up by one of the anonymous reviewers. 23 Also compare van Geenhoven (1995, 1996) on an analysis of Noun Incorporation in Germanic and West Greenlandic, who proposes a semantic distinction between two differing kinds of objects which is very reminiscent of the distinction proposed by Ramchand, but addresses the issue of specificity in much greater detail than can be found in Ramchand (1997). 24 Note that no predictions are made here with regard to the interaction of passivization and causatives. Passivization of a causative is not possible in Urdu, but this may be due to independent morphosyntactic factors, as it does appear to be possible in languages like Japanese and Inuit as an anonymous reviewer pointed out.

REFERENCES Aissen, J., and D. Perlmutter (1983). Clause reduction in Spanish. In D. Perlmutter (ed.), Studies in Relational Grammar, Vol. 1, 360-403. The University of Chicago Press, Chicago. Alsina, A. (1992). On the argument structure of causatives. Linguistic Inquiry 23(4), 517-555. Alsina, A. (1993). Predicate Composition: A Theory of Syntactic Function Alternations. Ph.D. thesis, Stanford University. Alsina, A. (1995). The Fall of Function-Argument Biuniqueness. In G. V. Morrill and R. T. Oehrle (eds.), Proceedings of the Conference of the European Summer School in Logic, Language and Information, Barcelona. Alsina, A. (1996). The Role of Argument Structure in Grammar: Evidence from Romance. CSLI Publications, Stanford, CA. Alsina, A., and S. Joshi (1991). Parameters in causative constructions. In L. Dobrin, L. Nichols, and R. M. Rodriguez (eds.), Papers from the 27th Regional Meeting of the Chicago Linguistic Society, 1-16. Alsina, A., and S. Mchombo (1993). Object Asymmetries and Chichewa Applicative Construction. In S. Mchombo (ed.) Theoretical Aspects of Bantu Grammar. CSLI Publications, Stanford, CA. Baker, M. (1988). Incorporation: A Theory of Grammatical Function Changing. University of Chicago Press, Chicago. Baker, M. (1989). Object sharing and projection in serial verb constructions. Linguistic Inquiry 20, 513-553.

110

Miriam Butt

Bresnan, J. (ed.) (1982). The Mental Representation of Grammatical Relations. MIT Press, Cambridge, MA. Bresnan, J., and J. Kanerva (1989). Locative inversion in Chichewa: A case study of factorization in grammar. Linguistic Inquiry 20, 1-50. Bresnan, J., and L. Moshi (1990). Object asymmetries in comparative Bantu syntax. Linguistic Inquiry 21, 147-185. Bresnan, J., and A. Zaenen. (1990). Deep Unaccusativity in LFG. In K. Dziwirek, P. Farrell, and E. Mejias-Bikandi (eds.), Grammatical Relations: A CrossTheoretical Perspective. CSLI Publications, Stanford, CA. Butt, M. (1993a). A Reanalysis of long distance agreement in Urdu. In Proceedings of the Nineteenth Annual Meeting of the Berkeley Linguistics Society, Berkeley Linguistics Society, University of California, Berkeley. Butt, M. (1993b). Object specificity and agreement in Hindi/Urdu. In Papers from the 29th Regional Meeting of the Chicago Linguistic Society Vol. 1, 80-103. Butt, M. (1994). Complex predicate scrambling in Urdu. In M. Butt, T. King, and G. Ramchand (eds.), Theoretical Perspectives on Word Order in South Asian Languages. CSLI Publications, Stanford, CA. Butt, M. (1995). The Structure of Complex Predicates in Urdu. CSLI Publications, Stanford, CA. Butt, M., and W. Geuder (eds.) (to appear). Argument Projection: Lexical and Compositional Constraints. CSLI Publications, Stanford, CA. Dalrymple, M., J. Lamping, and V. Saraswat (1993). LFG semantics via constraints. In Proceedings of the Sixth Meeting of the European Chapter of the Association for Computational Linguistics. Davies, W., and C. Rosen (1988). Unions as multi-predicate clauses. Language 64, 52-CC 52-88.

de Hoop, H. (1992). Case Configuration and Noun Phrase Interpretation. Ph.D. thesis, University of Groningen, The Netherlands. Dowty, D. (1979). Word Meaning and Montague Grammar. Reidel, Dordrecht. Dowty, D. (1991). Thematic proto-roles and argument selection. Language 67(3), 547-619. Durie, M. (1993). Grammatical structures in verb serialization: Some preliminary proposals. In A. Alsina, J. Bresnan, and P. Sells (eds.), Complex Predicates. CSLI Publications, Stanford, CA. To appear. Evers, A. (1975). The Transformational Cycle in Dutch and German. Ph.D. thesis, University of Utrecht. (Distributed by the Indiana University Linguistics Club.) Frank, A. (1996). A note on complex predicate formation: Evidence from auxiliary selection, reflexivization, and past participle agreement in French and Italian. In M. Butt and T. King (eds.), Proceedings from the First LFG Conference, 174-189. RXRC, Grenoble, France. Gibson, J., and E. Raposo (1986). Clause union, the stratal uniqueness law and the Chomeur relation. Natural Language and Linguistic Theory 4, 295-331. Grimshaw, J. (1990). Argument Structure. MIT Press, Cambridge, MA.

Constraining Argument Merger

111

Grimshaw, J., and A. Mester (1988). Light verbs and 0-marking. Linguistic Inquiry 19(2), 205-232. Hale, K., and S. J. Keyser. (1993). On argument structure and the lexical expression of syntactic relations. In K. Hale and S. J. Keyser (eds.), The View from Building 20, 53-109. MIT Press, Cambridge, MA. Halvorsen, P.-K. (1983). Semantics for Lexical-Functional Grammar. Linguistic Inquiry 14(4), 567-615. Higginbotham, J. (1985). On semantics. Linguistic Inquiry 16, 547-593. Hinrichs, E., and T. Nakazawa (1990). Subcategorization and VP structure in German. In Huges, Schaun, and Salmons (eds.), Proceedings of the Third Symposium on Germanic Linguistics. Benjamins, John Amsterdam. Jackendoff, R. (1990). Semantic Structures. MIT Press, Cambridge, MA. Jespersen, O. (1954). A Modem English Grammar on Historical Principles. Allen and Unwin, London. Kachru, Y. (1980). Aspects of Hindi Grammar. Manohar Publications, New Delhi. Kaplan, R. (1995). The formal architecture of Lexical-Functional Grammar. In M. Dalrymple, R. Kaplan, J. Maxwell III, and A. Zaenen (eds.), Formal Issues in Lexical-Functional Grammar, 7-27. CSLI Publications, Stanford, CA. Kaplan, R., and J. Bresnan (1995). Lexical-Functional Grammar: A formal system for grammatical representation. In M. Dalrymple, R. Kaplan, J. Maxwell III, and A. Zaenen (eds.), Formal Issues in Lexical-Functional Grammar, 29-130. CSLI Publications, Stanford, CA. Kaplan, R., and J. Wedekind (1993). Restriction and correspondence-based translation. In Proceedings of the Sixth Meeting of the European Chapter of the Association of Computational Linguistics. University of Utrecht. Kiparsky, P. (1987). Morphology and grammatical relations. Unpublished manuscript, Stanford University. Krifka, M. (1992). Thematic relations as links between nominal reference and temporal constitution. In I. Sag and A. Szabolcsi (eds.), Lexical Matters, 29-53. CSLI Publications, Stanford, CA. Li, Y. (1995). Chinese resultatives. Natural Language and Linguistic Theory 13, 255-282. Manning, C. (1992). Romance is so complex. Technical Report CSLI-92-168. CSLI Publications, Stanford, CA. Manning, C. (1995). Valency versus binding: On the distinctness of argument structure. Unpublished manuscript, Carnegie Mellon University. Manning, C. D., and I. A. Sag. (1998). Dissociations between argument structure and grammatical relations. In G. Webelhuth, A. Kathol, and I-P. Koening (eds.), Studies in Constraint-Based Lexicalism. CSLI Publications, Stanford, CA. To appear. Marantz, A. (1984), On the Nature of Grammatical Relations. MIT Press, Cambridge, MA. Masica, C. (1991). The Indo-Aryan Languages. Cambridge University Press, Cambridge, UK.

112

Miriam Butt

Matsumoto, Y. (1996). A syntactic account of light verb phenomena in Japanese. Journal of East Asian Linguistics 5, 107-149. Mohanan, K. P., and T. Mohanan (1990). Dative subjects in Malayalam: Semantic information in the syntax. In M. K. Verma and K. P. Mohanan (eds.), Experiencer Subjects in South Asian Languages. CSLI Publications, Stanford, CA. Mohanan, T. (1988). Causativization in Malayalam. Unpublished manuscript, Stanford University. Mohanan, T. (1994a). Argument Structure in Hindi. CSLI Publications, Stanford, CA. Mohanan, T. (1994b). Case OCP: A constraint on word order in Hindi. In M. Butt, T. King, and G. Ramchand (eds.), Theoretical Perspectives on Word Order in South Asian Languages. CSLI Publications, Stanford, CA. Monachesi, P. (1995). A Grammar of Italian Clitics. Ph.D. thesis, Tilburg University. Neeleman, A. (1994). Complex Predicates. Ph.D. thesis, Utrecht University. Rambow, O. (1996). Word order, clause union, and the formal machinery of syntax. In M. Butt and T. King (eds.), Proceedings from the First LFG Conference, 366-374. RXRC, Grenoble, France. Ramchand, G. (1997). Aspect and Predication: The Semantics of Argument Structure. Oxford University Press, Oxford. Ritter, E., and S. Rosen (1993). Deriving causation. Natural Language and Linguistic Theory 11, 519-555. Rosen, S. (1989). Argument Structure and Complex Predicates. Ph.D. thesis, Brandeis University. Saksena, A. (1980). The affected agent. Language 56(4), 812-826. Saksena, A. (1982a). Case marking semantics. Lingua 56, 335-343. Saksena. A. (1982b). Topics in the analysis of causatives with an account of Hindi paradigms. Linguistics 98. University of California Press, Los Angeles. Sells, P. (1993). The Projection of Phrase Structure and Argument Structure in Japanese. Unpublished manuscript, Stanford University. Tenny, C. (1987). Grammaticalizing Aspect and Affectedness. Ph.D. thesis, Massachusetts Institute of Technology, Cambridge, MA. Tenny, C. (1994). Aspectual Roles and the Syntax-Semantics Interface. Kluwer Academic, Dordrecht. van Geenhoven, V. (1995). Semantic incorporation: A uniform semantics for West Greenlandic noun incorporation and West Germanic bare plural configurations. In Papers from the 31st Regional Meeting of the Chicago Linguistic Society. To appear. van Geenhoven, V. (1996). Semantic Incorporation and Indefinite Descriptions: Semantic and Syntactic Aspects of Noun Incorporation in West Greenlandic. Ph.D. thesis, Universitat Tubingen. (SfS-Report-03-96.)

Constraining Argument Merger

113

Vendler, Z. (1967). Linguistics in Philosophy. Cornell University Press, Ithaca, NY. Verkuyl, H. (1972). On the Compositional Nature of the Aspects. Reidel, Dordrecht. Verkuyl, H. (1993). A Theory of Aspectuality. Cambridge University Press, Cambridge, UK. Williams, E. (1980). Predication. Linguistic Inquiry 11, 203-238. Zaenen, A. (1993). Unaccusativity in Dutch: Integrating syntax and lexical semantics. In J. Pustejovsky (ed.), Semantics and the Lexicon, 129-161. Kluwer Academic, Dordrecht.

This page intentionally left blank

THIRD CONSTRUCTION AND VP EXTRAPOSITION IN GERMAN: AN HPSG ANALYSIS ERHARD W. HINRICHS* TSUNEKO NAKAZAWA *Seminarfur Sprachwissenschaft Universitdt Tubingen Tubingen, Germany Language and Information Sciences University of Tokyo Tokyo, Japan

1. INTRODUCTION

The term THIRD CONSTRUCTION has been coined by den Besten and Rutten (1989) to refer to a syntactic configuration in which subject-control verbs in Dutch split their VP complements as in (Ib): (1) a. ...dat zij

that '... that b. ...dat that '... that

waagden de vergadering te verstoren.

they dared the meaning to disturb they dared to disturb the meeting' zij de vergadering waagden te verstoren. they the meeting dared to disturb they dared to disturb the meeting'

Syntax and Semantics, Volume 30 Complex Predicates in Nonderiuational Syntax

115

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

116

E. W. Hinrichs and T. Nakazawa

Unlike ordinary VP extraposition, which extraposes a VP as a whole as in (la), in the third construction the subject-control verb waagden appears between the NP de vergadering and the extraposed verb te verstoren. As noted by a variety of authors, including Santorini and Kroch (1990), GeilfuB (1991), Haider (1993), Rambow (1994), and Wollstein-Leisten (1994), this third construction can also be found in German, at least in spoken language.1 (2) a. ... daft er versucht hat das Auto zu reparieren. that he tried has the car to repair '.. .that he tried to repair the car' b. ... daft er das Auto versucht hat zu reparieren. that he the car tried has to repair '... that he tried to repair the car' One reasonable hypothesis to pursue is to regard the third construction, as exemplified in (2b), as a special case of extraposition of full VPS. This is the position taken by den Besten and Rutten (1989), who argue for a derivational analysis of the third construction in Dutch. They suggest that this construction involves extraposition of a full VP from which an NP is scrambled to the left of the finite verb. Uszkoreit (1987) presents a similar analysis in a nonderivational framework where the scrambled NP is considered to be the extracted filler of the gap contained in an extraposed VP. This question whether the third construction is a special case of extraposition of full VPS can only be answered once a syntactic analysis of VP extraposition has been put in place that is justified on independent grounds. Therefore, we will first present an analysis of full VP extraposition before we present an analysis of the third construction itself. Full VP extraposition has been identified by Bech (1955) as one of the tests that are characteristic of what he calls INCOHERENT INFINITIVE CONSTRUCTIONS. In order to give an account of VP extraposition, we therefore need to embed this account in a general treatment of such incoherent constructions.

2. INCOHERENT INFINITIVE CONSTRUCTIONS IN GERMAN In an incoherent infinitive construction, the verb governed by a control verb forms a VP constituent with its objects. This constituenthood of the VP complement of such a control verb can be demonstrated by a variety of syntactic constructions.

Third Construction and VP Extraposition

117

Apart from VP extraposition, incoherent infinitive constructions allow VP relatives as in (3a), and topicalization of VP complements as in (3b): (3) a. Das ist das Auto, das zu kaufen er Peter uberreden wird. that is the car which to buy he Peter persuade will 'That is the car which he will persuade Peter to buy.' b. Das Auto zu kaufen wird Peter Maria Uberreden. the car to buy will Peter Maria persuade 'Peter will persuade Maria to buy the car.' In the relative clause in (3 a) the embedded verb zu kaufen is fronted together with its object NP that is realized as the relative pronoun das? In (3b) the embedded VP das Auto zu kaufen appears in clause-initial topicalized position. On the other hand, incoherent infinitive constructions do not allow auxiliary flip, as in (4a), and scrambling of complements out of the embedded VP, as in (5a): (4) a. * ... dafl er Peter das Auto wird zu kaufen uberreden konnen. that he Peter the car will to buy persuade can '(intended)... that he will be able to persuade Peter to buy the car' b. ... daft er das Auto wird kaufen konnen. that he the car will buy can '... that he will be able to buy the car' (5) a. * ... daft er das Auto Peter zu kaufen uberreden wird. that he the car Peter to buy persuade will '(intended)... that he will be persuade Peter to buy the car' b. ... dafl er das Auto Peter kaufen sehen wird. that he the car Peter buy see will ' ... that he will be see Peter buy the car' Auxiliary flip is triggered by a small set of infinitival forms, including konnen, and places its governing verb, wird in (4b), as the leftmost verbal element in the verbal cluster. However, in the case of an incoherent construction, as in (4a), the fronted auxiliary may not intervene between the elements of the embedded VP das Auto zu kaufen. Likewise, the object NP das Auto in (5a) cannot participate in the scrambling of arguments in the German MITTELFELD of the kind exemplified in (5b). If in incoherent infinitive constructions the verb governed by a control verb forms a VP constituent with its objects, the behavior with respect to auxiliary flip and scrambling should come as no surprise since these constructions cause the embedded VP to split up into discontinuous parts. Given this general behavior, the existence of the third construction among verbs that undergo full VP extraposition is indeed surprising, since

118

E. W. Hinrichs and T. Nakazawa

the third construction does precisely what scrambling and auxiliary flip are not allowed to do, namely, to split up the embedded VP. This anomalous behavior should cast some initial doubt on the hypothesis that the third construction is a special case of full VP extraposition.

3. VP EXTRAPOSITION

As we mentioned in the Introduction, the question whether the third construction can be regarded as a special case of extraposition of full VPS, can only be answered once a syntactic analysis of VP extraposition has been put in place. In this section we therefore present an analysis of full VP extraposition before we propose an analysis of the third construction itself. VP extraposition, VP relatives, and the lack of scrambling and auxiliary flip all attest to the constituenthood of the embedded VP in an incoherent infinitive construction. For simple cases such as (6) the constituent structure in (7), therefore, suggests itself:3 (6)

... dafi Peter Maria iiberredet, das Auto zu kaufen. that Peter Maria persuade the car to buy '... that Peter persuades Maria to buy the car'

(7)

The lexical entry for the control verb iiberreden which governs the embedded VP is shown in (8):4 (8) (Preliminary Version)

Third Construction and VP Extraposition

119

Notice that, in flat structures such as (7), VP extraposition can be treated in terms of ordering constraints vis-a-vis its sisters; i.e., no extraction seems to be required. The correct linearization will be effected by some binary feature, e.g., EXTRAP. A VP complement that carries the value [EXTRAP + ], such as das Auto zu kaufen in (7), will be extraposed. We will defer the presentation of the necessary LP rules until section 8. So far we have only considered a simple case of VP extraposition. Next, we consider more complicated examples and show that they can be treated without movement as well. The following example (9) shows a case of double VP extraposition. In the extraposed VP Karl zu zwingen, das Auto zu kaufen, the VP das Auto zu kaufen which is embedded under zwingen is itself extraposed. (9)

... daft Peter Maria iiberredet, Karl zu zwingen. that Peter Maria persuaded Karl to force das Auto zu kaufen. the car to buy '... that Peter persuaded Maria to force Karl to buy the car'

Notice that such cases of multiple VP extraposition can be handled purely in terms of ordering constraints, as the constituent structure in (10) for (9) shows:

(10)

In (10) the VP which dominates the entire extraposed material is an immediate daughter of the clause and can therefore be ordered vis-a-vis its clausal sisters. The extraposed VP itself has a ternary structure with another VP that can be ordered in the desired fashion with respect to its sisters.

120

E. W. Hinrichs and T. Nakazawa

Another class of examples to consider are sentences in which a control verb is embedded under one or more auxiliaries as shown in (11): (11)

... daft Peter Maria uberredet haben wird, das Auto zu kaufen. that Peter Maria persuade have will the car to buy '... that Peter will have persuaded Maria to buy the car'

At first glance examples such as (11) seem to strongly suggest that VP extraposition requires movement, or in HPSG terms employment of a SLASH mechanism. The control verb uberredet appears embedded under the auxiliaries haben and wird so that the VP embedded under uberredet seems to be the most deeply embedded constituent. The only way to extrapose this VP seems to be in terms of adjunction to the clause. Whether or not VP extraposition requires extraction when a control verb is embedded under one or more auxiliaries crucially depends on the kind of analysis that is given to auxiliaries. Following the work of Bech (1955), auxiliaries can be categorized as typical examples of what Bech calls COHERENT INFINITIVE CONSTRUCTIONS. Bech shows that coherent infinitive constructions differ systematically from incoherent constructions with respect to the syntactic constructions of auxiliary slip, scrambling, and VP relativization that we discussed in section 2. While incoherent infinitive constructions do not allow auxiliary flip and scrambling, coherent constructions do as examples (12) show: (12) a. .. .dafi Peter es wird lesen konnen. that Peter it will read be able to '... that Peter will be able to read it' b. ... daB es Peter lesen konnen wird. that it Peter read be able to will '... that Peter will be able to read it.' In (12a) the finite auxiliary wird is flipped in front of the main verb, while in (12b) the object NP es is scrambled relative to the subject NP Peter. On the other hand, coherent infinitive constructions do not allow VP relativization as the ungrammatical (13) shows: (13) *Das ist das Buch das lesen Peter konnen wird. that is the book which read Peter be able will '(intended) That is the book which Peter will be able to read.' What this syntactic behavior seems to suggest is that in coherent infinitive constructions in general and for auxiliaries in particular there is no embedded VP constituent.5 Instead, the object arguments of the embedded verb are raised to the clause level. In previous work (Hinrichs and

Third Construction and vp Extraposition

121

Nakazawa, 1994b and to appear) we have accounted for this "verb raising" phenomenon by assigning the syntactic structure in (14) to (12b):6

(14)

The raising of the NP complement es is accomplished by the operation of argument composition that is encoded in the lexical entry of auxiliaries as shown in (15) (see Hinrichs and Nakazawa, 1989, 1993, 1994a, for details):7 (15) (Preliminary Version)

Notice that as a result of argument composition, the NP complement appears as an immediate daughter of the clause, where it can then be scrambled by linearization rules. The notion of a verbal complex or verbal cluster for the analysis of coherent infinitive constructions in Dutch and German was first introduced into generative grammar by Evers (1975). However, apart from the fact that his analysis is couched in transformational terms, there are considerable differences between his account and ours. While we use the lexical notion of argument composition to raise nonverbal arguments into the Mittelfeld, Evers employs a movement transformation of "verb raising" to form a verbal cluster. Evers assumes an underlying structure for clauses such as daft es Peter lesen wird 'that Peter will read it' in which wird

122

E.W. Hinrichs and T.Nakazawa

subcategorizes for a sentential complement Peter es lesen. The embedded verb is then raised out of the embedded clause to form a verbal cluster with the auxiliary. It is also worth noting that Evers treats only coherent infinitive constructions and VP extraposition, but offers no account of the third construction. The alert reader may wonder whether the structure in (14) is the only one admitted. It appears that nothing in the lexical entry in (15) requires that argument raising is obligatory. In our analysis, the obligatory nature of argument raising can be derived from the feature NPCOMP which appears in the lexical entry for auxiliaries. We will explain these points more fully in section 6. Analogously to the raising of NP complements, auxiliaries can also raise VP complements so that we can derive the structure in (16) for cases of VP extraposition examples such as (11):

The VP complement in (16) appears as an immediate daughter of the clause, so that it can be extraposed by linearization rules.' The treatment of VP extraposition as reordering and not movement is a natural extension of the analysis of coherent infinitive constructions that we have independently motivated in previous work (Hinrichs and Nakazawa, 1989, 1993, and 1994a19 The linearization account of VP extraposition is also in keeping with the general strategy followed by nonderivational theories of grammar, namely, to employ the SLASH mechanism only when a nonmovement analysis is not possible.

4. THIRD CONSTRUCTION AS A SPECIAL CASE OF VP EXTRAPOSITION Let us return now to the question that we asked at the outset, namely, whether the third construction can be treated as a special case of VP extraposition. In light of our linearization analysis of VP extraposition, this

Third Construction and VP Extraposition

123

question boils down to whether we can generalize this linearization account to cover the third construction as well. We will argue that this is, in fact, possible and that the third construction requires no extraction. In addition, we will show that such a linearization account of the third construction can be based on the notion of argument composition, which is needed independently for coherent verbs. It is worth noting that those verbs that introduce the third construction appear in certain incoherent constructions as well. They admit VP relatives as in (17) and extraposition of full VPS as in (18): (17) Das ist das Buck das zu lesen Peter versuchen wird. that is the book which to read Peter try will 'That is the book which Peter will try to read.' (18)

.. .daft er versucht hat das Auto zu reparieren. that he tried has the car to repair '... that he tried to repair the car'

In order to account for these incoherence properties, we have to assume that verbs such as versuchen subcategorize for a full VP just like uberreden in (8). Next let us consider once more our earlier example of the third construction (2b), repeated here as (19): (19)

... daft er das Auto versucht hat zu reparieren. that he the car tried has to repair '... that he tried to repair the car'

Observationally speaking, the third construction splits the NP das Auto and the infinitive zu reparieren which are embedded under versuchen into discontinuous parts. We propose to analyze this by the Third Construction Lexical Rule shown in (20):

(20) Third Construction Lexical Rule (preliminary version)

124

E. W. Hinrichs and T. Nakazawa

The lexical rule substitutes the VP embedded under a control verb by a COMPS list that is composed of a verb and its subcategorized complements. The rule in a sense liberates the constituents of the embedded VP in the input structure and makes them immediate complements of the upstairs verb in the output structure. If the subject-control verb versuchen undergoes the lexical rule in (20), then the subordinate clause in (19) will be assigned the constituent structure in (21). For reasons to be explained shortly, the rule applies only to subject-control equi verbs without an overt object, i.e., to control verbs which have a VP as the only element on the COMPS list.

(21)

Since versuchen is embedded under the coherent verb hat, the complements of versuchen are raised to the clause level where they can be ordered by appropriate linearization rules. The lexical rule is formulated in such a way that it applies to all subject-control equi verbs without an overt object. The third construction, however, is known to be licensed by other kinds of verbs as well, including those listed in (22). The third construction seems to be productive for subject-control equi verbs without object as listed in (22a), for the subjectcontrol equi verbs in (22b) if their optional dative objects are not realized,

Third Construction and VP Extraposition

125

for phase verbs in (22c), and for object-control equi verbs with dative (rather than accusative) controller in (22d): (22) Third Construction Triggering Verbs a. subject-control equi verbs without object: uersuchen 'try'; glauben 'believe'; vergessen 'forget'; wagen 'dare'; vermogen 'be capable of; meinen 'think'; behaupten 'maintain'; vorhaben 'intend'; zugeben 'admit'; hoffen 'hope'; versaumen 'fail' b. subject-control equi verbs with optional dative object: versprechen 'promise'; drohen 'threaten'; zusichern 'assure'; schworen 'swear' c. phase verbs (subject-control raising verbs): anfangen 'start'; aufhoren 'stop'; beginnen 'begin' d. object-control equi verbs with dative controller: (Bayer and Kornfilt, 1990:38) empfehlen 'recommend'; befehlen 'order'; nahelegen 'suggest'; uerbieten 'forbid'; untersagen 'forbid'; gestatten 'allow'; bedeuten 'indicate'; freistellen 'leave'; zutrauen 'believe'; gebieten 'demand' For the group in (22b) the grammaticality judgments seem to vary from verb to verb, if the optional dative object is realized. For example, while versprechen seems to allow the third construction even with an overt dative object, drohen does not: (23) a.

... daft ich Peter den Brief versprochen habe abzuschicken. that I Peter the letter promised have to send '... that I have promised Peter to send the letter' b. ?* ... daft ich Peter den Brief gedroht habe abzuschicken. that I Peter the letter threatened have to send '(intended)... that I have threatened Peter to send the letter'

All object-control equi verbs listed in (22d) take a dative controller as pointed out by Bayer and Kornfilt (1990). For some reason unknown to us, object-control equi verbs do not trigger the third construction if they subcategorize for an accusative NP, as the contrast between (24a) and (24b) illustrates: (24) a.

.. .dafl sie dem Studenten das Buch geraten hat zu lesen. that she the-DAT student the book advised has to read '(intended) 'that she has advised the student to read the book' b. * .. .daft sie den Studenten das Buch gezwungen hat zu lesen. that she the-ACC student the book forced has to read '(intended)... that she has forced the student to read the book'

126

E. W. Hinrichs and T. Nakazawa

Furthermore, not all object-control equi verbs with a dative controller allow for the third construction as shown in the ungrammatical example in (25): (25) * .. .daB sie dem Studenten das Buck nachgesehen hat zu lesen. that she the-DAT student the book forgiven has to read '(intended)... that she has forgiven the student for reading the book' The somewhat idiosyncratic nature of the third construction raises the question whether one should propose a lexical rule for this construction or whether one should simply posit two lexical entries for third construction triggering verbs: i.e., one entry of the sort shown in (8) for the incoherent construction and a second one which looks like the output of the lexical rule (20). The reason we have chosen the lexical rule approach here is that it does seem to capture the productive nature of the third construction with respect to the verbs listed in (22a). However, even with the lexical rule in place, additional lexical entries have to be stipulated for subject-control equi verbs in (22b) which allow the third construction with an overt dative NP, phase verbs in (22c), and object-control equi verbs in (22d).10 The output of the lexical rule looks surprisingly similar to the lexical entries that we have already proposed for control verbs which license coherent structures, e.g., the lexical entry for konnen in (15). We noted that the lexical entries of those verbs encode argument composition with the embedded verb. In the output structure of the lexical rule (20) the same type of composition is encoded. There are some minor differences between the two types of lexical entries. In the case of coherent structures, the verbal complement may not be extraposed and is marked as [EXTRAP - ]. Also the governed verb can be a base infinitive or a zu -infinitive. Third construction triggering verbs, on the other hand, select only zu -infinitives.11 But apart from some minor differences we find it significant that the same type of operation can be used in both cases. A difference between the coherent constructions and the third construction which is not apparent from the output of the Third Construction Lexical Rule (20) is the positioning of the governed verb. In the case of coherent constructions, the governed verb forms part of the verbal complex and cannot be raised. In the case of the third construction, the governed verb can be and will, in fact, be forced to be raised and extraposed. The question of what makes the raising of extraposed verbs obligatory will be answered in some detail in section 6. We will now turn to more complex cases of the third construction where this construction interacts with one or more coherent or incoherent

Third Construction and vp Extraposition

127

constructions in the same sentence. We will show that our linearization analysis of the third construction can account for these more complex cases as well. In examples such as (26) the extraposed verbal constituent in the third construction is complex and contains one or more control verbs which license coherent constructions: (26) a. .. .daB that '... that b. .. .daft that '... that

er das Auto meint reparieren zu durfen. he the car thinks repair to be allowed he thinks that he is allowed to repair the car' er das Auto meint repariert haben zu durfen. he the car thinks repair have to be allowed he thinks that he is allowed to have repaired the car'

The constituent structure that our analysis admits for (26b) is shown in (27):

(27)

(27) shows that the extraposed verbal constituent forms a structured verbal complex which as a whole appears as a daughter of the clause and can therefore be ordered by an appropriate linearization statement. In general, what can be extraposed in the third construction is any lexical zw-infinitive or any coherent construction, i.e., a single verbal complex. However, incoherent constructions, i.e., complex verbal categories which contain an embedded VP, as in (28), are not permissible: (28)

* ... daft ich Maria vergessen habe zu versprechen that I Maria forgotten have to promise das Auto zu reparieren. the car to repair '(intended)... that I have forgotten to promise Maria to repair the car'

128

E. W. Hinrichs and T. Nakazawa

If (28) were grammatical, then uergessen would have to be in the third construction. The NP complement Maria of the embedded verb versprechen would, hence, be missing from the extraposed verbal category. The extraposed verbal category itself represents an incoherent construction since it contains the extraposed VP complement das Auto zu reparieren. Hence, (28) violates the empirical restriction observable about the third construction, namely, that the extraposed verbal category can only be complex if it constitutes a coherent construction. Next we will consider cases in which the third construction interacts with full VP extraposition in the same sentence. We will see that the third construction may be embedded under an extraposed full VP as in (29), and that the linearization account that we would like to propose for both constructions can still be upheld. (29)

... daft ich Maria uberreden mufi das Fahrrad that I Maria persuade must the bicycle zu versuchen zu reparieren. to try to repair '... that I must persuade Maria to try to repair the bicycle'

The constituent structure for (29) that the linearization account would predict is shown in (30): (30)

In (30), the VP das Fahrrad zu versuchen zu reparieren constitutes extraposition of a full VP, and within that VP the Third Construction Lexical Rule has applied to versuchen. When the third construction interacts with full VP extraposition, our analysis also correctly predicts that it is ungrammatical to place a constituent that forms part of the third construction outside the extraposed full VP. In the following two ungrammatical examples, the object NP das Fahrrad, which is involved in the third construction, remains nonextraposed while the rest of the complement of uberreden is extraposed in (3la),

Third Construction and VP Extraposition

129

and the rest of the complement of versuchen is extraposed in (31b): (31) a. * ... daft ich Maria das Fahrrad uberreden muB

that I

Maria the bicycle persuade must

zu versuchen zu reparieren.

to try to repair '(intended)... that I must persuade Maria to try to repair the bicycle' b. * ... daB ich Maria das Fahrrad zu versuchen that I Maria the bicycle to try uberreden muB zu reparieren.

persuade must to repair '(intended)... that I must persuade Maria to try to repair the bicycle' In both examples versuchen appears in the third construction and is supposed to be in the scope of the object-control verb uberreden. Uberreden subcategorizes for a saturated VP as its complement and licenses incoherent constructions. However, in both sentences there is no saturated VP constituent, i.e., das Fahrrad zu versuchen zu reparieren, that could satisfy this subcategorization requirement. Therefore, the sentences in (31) are not admitted. To conclude this section of the paper, we have shown that a linearization approach can account for simple cases of the third construction in German as well as for complex interactions of the third construction with VP extraposition. The linearization approach is not only observationally fully adequate, but it also extends quite naturally our analysis of coherent infinitives that we have proposed previously. Finally, it is important to note that the phenomena of VP extraposition and the third construction are both clause-bound; i.e., they are extrapositions whose domain is a single clause. Since the SLASH extraction mechanism is typically employed for dependencies of unbounded depth, the linearization analysis has to be considered the null hypothesis.

5. DUAL PROPERTIES OF THIRD CONSTRUCTION TRIGGERING VERBS

Additional evidence for treating third construction triggering verbs in terms of argument composition comes from the fact that these verbs exhibit syntactic properties characteristic of coherent infinitive constructions when the governed verb is not extraposed.

130

E. W. Hinrichs and T. Nakazawa

(32)

... daft es Peter zu reparieren versuchen wird. that it Peter to repair try will '... that Peter will try to repair it'

(33)

... dqfi es Peter wird zu reparieren versucht haben. that it Peter will to repair tried have '... that Peter will have tried to repair it'

(32) and (33) illustrate two of the constructions introduced by Bech as tests for coherent infinitive constructions: scrambling and auxiliary flip. In (32) the pronoun es which is the object of reparieren has been scrambled to the left of the subject NP.12 In (33) the finite auxiliary wird is fronted and appears to the left of the verbal complex. These coherence properties can be accounted for in our treatment of third construction triggering verbs since the output of the Third Construction Lexical Rule involves argument composition. What distinguishes the examples in (32) and (33) from the third construction itself is that the ZM-infinitive is not right-dislocated. In our analysis, therefore, such nondislocated infinitives are marked as [EXTRAP — ] and are admitted into the verbal complex in the same way as base infinitives are in the coherent constructions. (34) presents the tree that our analysis assigns to (32): (34)

In (34) es is raised by the control verb versuchen and then by the finite auxiliary wird. As a result, es and the subject NP Peter are sisters at the clause level and can therefore be scrambled. With respect to scrambling in the Mittelfeld, third construction triggering verbs follow the pattern characteristic of coherent infinitive constructions. However, as we have discussed in the previous sections, third construction triggering verbs also exhibit properties characteristic of incoherent infinitive constructions. For example, they allow full VP extraposition, which is Bech's primary test for incoherence. Moreover, third construction triggering verbs parallel the behavior of other control verbs that

Third Construction and VP Extraposition

131

appear in incoherent constructions in that the embedded VP can appear in the Mittelfeld instead of being extraposed.13 (35)

... daB ein solches Buck zu lesen Hans niemals versuchen wurde. that a such book to read Hans never try would '... that Hans would never try to read such a book'

It is important to note that sentences such as (35) cannot be regarded as examples of coherent infinitive constructions, since this would require zu lesen and versuchen to form a constituent in the same way zu reparieren and versuchen form a verbal complex in the coherent structure shown in (34). Rather (35) has to be assigned the structure shown in (36), in which ein solches Buck zu lesen forms a VP:

(36)

This dual behavior of third construction triggering verbs as a licensor of both coherent and incoherent constructions follows immediately from the two types of lexical entries that our analysis assigns to these verbs, i.e., a lexical entry which subcategorizes for a full VP and a second entry, derived by the lexical rule, which raises the arguments of its embedded verbs. On the other hand, verbs which do not trigger the third construction and full VP extraposition, e.g., auxiliaries, appear only in coherent structures. They will be assigned the latter type of entry. Finally, there is the class of verbs, e.g., iiberreden, which only appear in incoherent constructions and are assigned the first type of entry subcategorizing for a full VP.

6. ARGUMENT COMPOSITION, ID SCHEMAS, AND FORCING ARGUMENT RAISING IN THE THIRD CONSTRUCTION

Our linearization account of full VP extraposition and of the third construction requires that the verbal and nominal arguments are discharged at the right level of embedding. This can be accomplished if ID schemas and lexical rules are made sensitive to two features: the feature

132

E. W. Hinrichs and T. Nakazawa

EXTRAP, which marks whether a verbal category is to be extraposed or not, and the feature NPCOMP, which signifies whether or not that node directly or indirectly dominates any NFS. In (37) and (38) we list the two ID schemas that we have been tacitly assuming for German in the previous sections:

(37)

(38)

These ID schemas are similar to the schemas introduced in Hinrichs and Nakazawa (1994b and to appear). The schema in (37) licenses the flat structures of finite verbs and their complements as well as saturated VPS. Following Borsley's proposal for Welsh (Borsley, 1987), we are assuming here, as we did in Hinrichs and Nakazawa (1994b and to appear), that subjects of finite verbs in German appear on the COMPS list, rather than as the value of SUBJ. The Verbal-Complex ID Schema (38) licenses binary structures for the verbal complex, which always contains nonfinite verbs. It allows a coherent construction where a control verb as a head combines with a single verbal complement. The verbal complement is specified to be [EXTRAP — ] by the COMPS value of the control verb as shown in the lexical entry for konnen in (15). This feature specification is present to prohibit a zw-infinitive marked as [EXTRAP + ] from being discharged as the sister of a third construction triggering verb. This in turn means that if such a third construction triggering verb is embedded under a coherent construction, as in (26), the extraposed zw-infinitive is obligatorily raised. As discussed in section 4, this obligatory raising of extraposed zu-infinitives is crucial for our linearization analysis of the third construction.

Third Construction and vp Extraposition

133

The notion of verbal complex predates generative grammar and was described in such terms as KOHAERENZFELD 'coherence field' in Been (1955) and HAUPTVERBKOMPLEX 'main verb complex' in Bierwisch (1963). In our analysis, the notion is captured in terms of the [NPCOMP -] specification on verbal nodes. The Verbal-Complex ID Schema (38) recursively marks the mother and the verbal complement as [NPCOMP — ], and the feature signifies that the verbal constituent contains no NP complements. Note that this feature cannot be reduced to a more familiar feature, e.g., LEX, since the verbal complement marked as [NPCOMP — ] in a verbal complex may consist of only a lexical verb or multiple verbs, i.e., another verbal complex. Nor can it be identified with a nonempty COMPS list since lexical intransitive verbs may appear as a verbal complement in a verbal complex while their COMPS lists are empty. Recall that control verbs which only license coherent constructions subcategorize for verbal complements that are marked as [NPCOMP - ] as shown in lexical entry (15). Since the mother category in the Head-Complement ID Schema is marked as [NPCOMP + ], verbal constituents admitted by this schema can never appear as complements of those verbs which license coherent constructions. This then means that the raising of NP complements to the clause level is obligatory in coherent structures. Likewise the raising of full VP complements forming an incoherent structure is obligatory since such VPS are admitted by the Head-Complement ID Schema and are hence marked as [NPCOMP + ]. If the raising of VP and NP complements to the clause level was not enforced, then these complements could be discharged nondeterministically at any level of embedding within the verbal complex. Such nondeterminism would lead to a multiplicity of structures whose necessity cannot be motivated on independent grounds. The previous sections have shown that the linearization approach to the third construction and extraposition of full VPS has to interact with the argument composition mechanism for coherent constructions that we have discussed at length in previous work. But in Hinrichs and Nakazawa (1994b and to appear) we had to restrict argument raising to nonverbal categories in order to be able to block totally flat structures in which even elements of the verbal complex are direct daughters of the clause rather than forming a constituent. The assumption that argument raising can be restricted to nonverbal categories, is, of course, at odds with the assumption that we have made in this paper, namely, that verbal complements in incoherent infinitive constructions and the third construction can be raised as well. In order to accommodate the full range of categories that are allowed to be raised, we modify the lexical entry for coherent construction

134

E. W. Hinrichs and T. Nakazawa

verbs as shown in (39):

(39)

The list of raised arguments is specified disjunctively: Nonverbal categories are admitted, as before. In addition, verbal categories which carry the feature [EXTRAP + ] are allowed to account for cases of the third construction and full VP extraposition. Finally, we also admit verbal categories which carry the features [NPCOMP + ]. When a control verb licenses an incoherent construction, we assume that its VP complement is marked [NPCOMP + ], as shown in (40).

Thus, inclusion of the [NPCOMP + ] disjunct then also allows for the raising of nonextraposed VP complements in incoherent constructions. By contrast, the verbal complement of a control verb in a coherent construction is marked [EXTRAP -], as shown in (39). Due to the restriction which is

Third Construction and vp Extraposition

135

placed on the list of raised arguments in (39), such a verbal complement cannot be raised if the coherent construction is embedded under another coherent construction. Rather any verbal complement marked as [EXTRAP — ] is forced to be discharged as a sister of its governing verb inside the verbal complex. Now that we have modified the lexical entry of control verbs for coherent constructions, we also have to update the output specification of the Third Construction Lexical Rule since we treat verbs in the third construction as a special case of control verbs for coherent constructions. However, as we will discuss in the next section, we cannot simply duplicate the argument raising specification in (39) for verbs which license coherent constructions as the output of the lexical rule, since there exist some crucial differences between coherent infinitive constructions and the third construction.

7. THE THIRD CONSTRUCTION LEXICAL RULE

While third construction triggering verbs allow scrambling of NP complements in the same way as other control verbs do in coherent infinitive constructions, there is one systematic difference between third construction triggering verbs and other control verbs which license coherent constructions. While the control verbs can raise extraposed verbal complements, third construction triggering verbs may not. As the result, a third construction triggering verb is blocked from embedding another triggering verb whose verbal complement is extraposed, since it would mean that the higher triggering verb has to raise the extraposed verbal complement subcategorized by the lower triggering verb. This can be shown by the ungrammatical examples in (41a): (41) a. * ... daB ich das Fahrrad versprechen muft

that I

the bicycle promise

must

zu versuchen zu reparieren.

to try to repair '(intended)... that I must promise to try to repair the bicycle' b. * ... daB ich das Fahrrad zu versuchen that I the bicycle to try versprechen mufi zu reparieren.

promise must to repair '(intended)... that I must promise to try to repair the bicycle'

136

E. W. Hinrichs and T. Nakazawa

c. * ... daB ich zu versuchen versprechen muB that I to try promise must das Fahrrad zu reparieren. the bicycle to repair '(intended)... that I must promise to try to repair the bicycle' In (4la) versuchen is supposed to be in the scope of versprechen. In order to generate the sentence, both verbs would have to undergo the Third Construction Lexical Rule in (20) and versprechen would have to have the instantiated lexical category in (42). However, such an instantiation of the lexical category can be blocked if we do not allow third construction triggering verbs to raise verbal complements that are marked as [EXTRAP + ].

(42)

versprechen

Notice that the prohibition against raising extraposed verbal complements by third construction triggering verbs can also successfully block examples (41b) and (41c). The ungrammatical examples in (41b) and (41c) indicate that the third construction triggering verbs may not behave like ordinary control verbs for coherent constructions to form part of the verbal complex when they govern another control verb which subcategorizes for an extraposed verbal complement. In both examples, the third construction triggering verb versprechen appears in the verbal complex and governs the embedded verb versuchen which subcategorizes for the extraposed zuinfinitive verb in (41b) and for the extraposed VP in (41c). In order to generate example (41b), both versuchen and versprechen would have to undergo the Third Construction Lexical Rule (20). Furthermore, versprechen would have to raise the complement zu reparieren, which would be marked as [EXTRAP +]. In (41c) versprechen would have to undergo the Third Construction Lexical Rule (20) in order to raise the extraposed complement das Fahrrad zu reparieren. However, a prohibition against raising extraposed verbal complements in the lexical entry for third construction triggering verbs could block this.15 Hence, we propose the

Third Construction and VP Extraposition

137

modification to the Third Construction Lexical Rule shown in (43): (43) Third Construction Lexical Rule

The output of the lexical rule in (43) is different from control verbs such as konnen in (39) which appear only in coherent constructions in two aspects: first, in contrast to verbs for coherent constructions whose verbal complements are marked as [EXTRAP — ], the Third Construction Lexical Rule leaves the EXTRAP feature on the verbal complement unspecified. If the value of the EXTRAP feature is instantiated to be +, then the entry licenses the third construction. Such a feature specification is needed to license structures as in (21). If the EXTRAP value is instantiated as —, then the resulting entry licenses a coherent structure. Such a feature specification is needed for examples such as (32) and the corresponding structure in (34). Second, as shown in the lexical entry for konnen in (39), the list of arguments that can be raised by verbs which appear only in coherent constructions includes extraposed VPS. However, in the verbal complex governed by the output of the Third Construction Lexical Rule, only nonverbal categories and nonextraposed VP may be raised in order to rule out sentences such as (41). An example of the application of the lexical rule is given in (44) and (45). (44) shows the lexical entry for uersuchen, to which the Third Construction Lexical Rule applies to return the output

138

E. W. Hinrichs and T. Nakazawa

shown in (45): (44) versuchen

8. LP RULES

In this section we will discuss the LP rules that we have been tacitly assuming in the previous sections.16 LP rule (46) is taken from Hinrichs and Nakazawa (1993), where we assume that finite verbs in clause-initial or verb-second position carry the feature [INV + ]. All other verbal constituents are assumed to be marked as [INV - ]. (46) accounts for the placement of finite verbs in clause-initial position. A verb-second sentence can then be generated by a clause-initial topic constituent and a verb-initial clause (see Hinrichs and Nakazawa, 1994b, and to appear for details).

(46)

Third Construction and VP Extraposition

139

LP rule (47) orders noninverted verbal elements, regardless of whether they are heads or complements, to the right of any NFS: (47)

Since all verbal constituents that are headed by a nonfinite verb are marked as [INV - ], (47) places NFS to the left of the verbal complex and to the left of verbal complements of control verbs, regardless of whether they are extraposed or not. Since finite verbs in subordinate clauses are also marked as [INV — ], (47) also ensures the correct ordering between NFS and the finite verb in subordinate clauses. The remaining rules in (48)-(51) order verbal categories with respect to each other. Rules (48) and (49) order extraposed and nonextraposed VP complements of control verbs with respect to the verbal complex, which is always marked as [NPCOMP — ] by the Verbal-Complex ID schema given in (38): (48)

(49)

LP rules (50) and (51) order extraposed and nonextraposed verbal complements with respect to verbal heads:

(50)

(51)

140

E. W. Hinrichs and T. Nakazawa

LP rule (50) orders nonextraposed verbal categories to the left of noninverted verbal heads. Since the verbal category which dominates the verbal complex is marked as [EXTRAP — ], (50) ensures that the verbal complex precedes the finite verb in subordinate clauses. Since auxiliaries subcategorize for verbal complements that are marked as [EXTRAP - ], (50) also regulates the ordering of verbal heads and their complements within the verbal complex.17 Finally, (50) orders nonextraposed VP complements of control verbs to the left of noninverted finite heads. Here we have to distinguish between two cases: If the control verb is the finite verb of the subordinate clause, i.e., if the sentence contains no verbal complex, then (50) guarantees that its nonextraposed VP complement precedes it. If the control verb is nonfinite, i.e., if the sentence contains a verbal complex, then (50) orders the finite head to the right of the verbal complex, and LP rule (48) ensures that the nonextraposed VP complement of the control verb appears to the left of the verbal complex. (51) ensures that extraposed verbal complements in the third construction and VP complements of control verbs follow the finite head in subordinate clauses and hence appear in clause-final position. Note that the clause-final position of extraposed elements are ensured even when the finite head is inverted to the clause-initial position and hence does not immediately precede extraposed elements. Here we have to distinguish two cases: If the control verb is the inverted finite verb, then (51) together with LP rule (47) places the extraposed verbal complement in clause-final position. If the control verb is nonfinite and embedded within the verbal complex, then LP rule (49) makes sure that the extraposed verbal complement is ordered to the right of the verbal complex and hence appears in clause-final position.

9. LONG PASSIVES With our treatment of the third construction in place, we will consider one additional syntactic phenomenon that has been pointed out in the literature with regard to third construction triggering verbs, namely, the so-called LONG PASSIVE. Hohle (1978) was the first to note that, at least for some speakers, the object NP of the embedded verbal complement can be passivized when governed by third construction triggering verbs such as vergessen. (52) a.

... dafi der Hand uergessen wurde zu futtem. that the NOM dog forgotten was to feed '... that feeding the dog was forgotten'

Third Construction and VP Extraposition

141

b. * ... daB den Hund vergessen wurde zu futtern. that the ACC dog forgotten was to feed c. ... daB den Hund zu futtem vergessen wurde. that the ACC dog to feed forgotten was '... that feeding the god was forgotten' In (52a) the verb vergessen is passivized, and the logical object NP derHund of the embedded verb futtern appears in the nominative case while the verb futtern itself appears in the active mood. Thus is appears that the object NP of the downstairs verb has become the surface subject of the passivized upstairs verb: hence the term long passive.18 As (52b) shows, the nominative case marking is required; i.e., the object of the downstairs verb cannot retain accusative case. It is also possible to passivize vergessen without raising the embedded object to subject, as in (52c). In such cases the passivized verb vergessen licenses an incoherent construction by taking a full VP as a complement. The long passive of a third construction triggering verb in (52a) contrasts with regular "short-distance" personal and impersonal passives in (53)-(55), where a transitive verb exhibits passive morphology: (53) a. Peter sah den Mann. Peter saw the man 'Peter saw the man.' b. Der Mann wurde gesehen. the man was seen 'The man was seen.' (54) a. Peter half dem Mann. Peter helped the DAT man 'Peter helped the man.' b. Dem Mann wurde geholfen. the DAT man was helped 'The man was helped.' c. *Der Mann wurde geholfen. the NOM man was helped (55) a. Die Veteranen gedachten der Toten. the veterans commemorated the GEN dead 'The veterans commemorated the dead.' b. Der Toten wurde gedacht. the GEN dead were commemorated 'The dead were commemorated.'

142

E. W. Hinrichs and T. Nakazawa

c. *Die Toten wurde gedacht. the NOM dead were commemorated For personal passives, the accusative object NP of an active transitive verb, e.g., den Mann in (53a), corresponds to an NP with nominative case, e.g., der Mann in (53b). In impersonal passives, a dative or genitive NP complement of a transitive verb, e.g., the dative NP dem Mann in (54) and der Toten in (55), exhibits the same case assignment in the active and passive forms while promotion to nominative case is ungrammatical (cf. (54c) and (55c)). As in the case of ordinary passives, the long passive construction also allows both personal and impersonal passives. (56) a.

.. .daft dem Mann versucht wurde zu helfen. that the DAT man tried was to help '... that helping the man was tried' b. * ... daft der Mann versucht wurde zu helfen. that the NOM man tried was to help '(intended)... that helping the man was tried' c. ... daft dem Mann zu helfen versucht wurde. that the DAT man to help tried was '... that helping the man was tried'

The data in (56a) and (56c) parallel those in (52a) and (52c), except that (52a) shows an example of a long personal passive, while (56a) exemplifies the long impersonal passive. The ungrammaticality of (56b) shows that in long impersonal passives it is ungrammatical to promote a dative NP to nominative case. This parallels the ungrammaticality of ordinary impersonal passives in (54c). Additional data show that the distribution of long passives is not necessarily restricted to the third construction. Example (57) illustrates that third construction triggering verbs, i.e., subject-control equi verbs without an object NP, allow long passive even when the governed zw-infinitive is not extraposed. (57)

...daft der Hund zu futtem vergessen wurde. that the NOM dog to feed forgotten was '... that feeding the dog was forgotten'

This example exhibits an interesting contrast with (52c), where the entire VP constitues a complement of a passivized verb. The fact that the VP den Hund zu futtem is actually a constituent can be confirmed by (58a), where the entire VP is extraposed while the corresponding sequence der Hund zu futtem in (57) does not form a constituent and cannot be extraposed as

Third Construction and VP Extraposition

143

shown in (58b): (58) a.

...daft uergessen wurde den Hund zu futtem. that forgotten was the ACC dog to feed '... that feeding the dog was forgotten' b * ...daft uergessen wurde der Hund zu futtem. that forgotten was the NOM dog to feed '... that feeding the dog was forgotten'

The above data indicate that long personal passive can be characterized as a passivization of subject-control equi verbs with the passive subject raised from the COMPS lists of the governed verb. Thus it can be analyzed by the Long Passive Lexical Rule (59), which applies to the same set of lexical entries as the Third Construction Lexical Rule: (59) Long Passive Lexical Rule

The Long Passive Lexical Rule is not intended to replace, but rather to be introduced in addition to lexical rules that can license regular "short distance" personal and impersonal passives.19 It applies only to subjectcontrol equi verbs without an object NP and produces their passive forms. Unlike a regular passive lexical rule, the passive subject is specified as the object NP which the governed verb in the output subcategorizes for rather than the object NP of the input verbs. The governed zw-infinitive verb is

144

E. W. Hinrichs and T. Nakazawa

unspecified for the EXTRAP feature: if it is realized as [EXTRAP +], long passive appears as part of the third construction as in (52a), and if [EXTRAP - ], then long passive appears with a coherent construction as in (57). Note that other control verbs such as auxiliaries and Acl verbs also raise complements of the governed verb licensing coherent constructions, but they allow neither the third construction nor long passive as shown in (60): (60)

*Der Wagen wurde den Karl kaufen gesehen. the NOM car was the Karl buy seen '(intended) Karl was seen buying the car.'

This fact is accounted for the Long Passive Lexical Rule, which requires subject-control equi verbs as input. The object NP of the complement verb in the output of lexical rule (59) shares the index with the passive subject and is specified to be accusative. This specification guarantees that the nominative passive subject appears only when the complement verbs subcategorize for an accusative NP. In order to accommodate long impersonal passives, a slightly modified version of the Long Passive Lexical Rule would have to be postulated separately. That is, the COMPS list of the complement verb may not contain an accusative NP, and the SUBJ must take an empty list in the output. For further details of passivizability of different classes of verbs, see Pollard (1994). Kathol (1994) and Pollard (1994) have proposed to analyze German Passive without a lexical rule. The essence of their analysis is the view of passive as object-to-subject raising. In such a view, the passive auxiliary verb werden 'be' is the locus of passivization and its lexical entry specifies raising of an object NP of a complement verb, i.e., the main verb in the passive form, to the passive subject. Pollard (1994) analyzes long passive in terms of two steps: First, argument composition encoded in the lexical entry of a subject-control equi verb merges the COMPS list of the embedded verb, crucially including the downstairs object NP, into the COMPS list of itself. Then the auxiliary verb werden which governs the subject-control equi verb raises the downstairs object NP to the passive subject. Thus, in such an analysis of passivization, neither ordinary, "short-distance" passives, nor long passives require a lexical rule. Long passives simply follow from the interaction of argument composition and werden passives. Since it is beyond the scope of this paper to give a detailed comparison between the two approaches to passive, i.e., the lexical rule approach briefly sketched above and the approach taken by Kathol (1994) and Pollard (1994) to encode raising in the passive auxiliary werden, it suffices to point out that our analysis of the third construction is fully compatible with either approach and predicts the long passive data correctly.

Third Construction and VP Extraposition

145

10. A COMPARISON WITH PREVIOUS ANALYSES Den Besten and Rutten (1989) propose an analysis for Dutch that predicts that any third construction triggering verb also allows VP extraposition since the third construction is treated as a special case of VP extraposition: VP extraposition plus long-distance scrambling. Our analysis for German makes the same prediction since our lexical rule takes as input a certain subclass of control verbs, i.e., subject-equi verbs without overt object which license incoherent constructions. However, there are also some interesting differences between the scrambling account of den Besten and Rutten and our account of the third construction that is based on the notion of argument composition. These differences concern sentences in which a third construction triggering verb such as versuchen appears inside an extraposed VP, as in (61): (61) a.

... dqfl ich Maria uberreden muB zu versuchen that I Maria persuade must to try das Fahrrad zu reparieren. the bicycle to repair ' . . . that I must persuade Maria to try to repair the bicycle' b. ... daB ich Maria uberreden muB das Fahrrad that I Maria persuade must the bicycle zu versuchen zu reparieren. to try to repair '... that I must persuade Maria to try to repair the bicycle' c. * ... daB ich Maria das Fahrrad uberreden muB that I Maria the bicycle persuade must zu versuchen zu reparieren. to try to repair '(intended)... that I must persuade Maria to try to repair the bicycle'

Our analysis correctly predicts the grammaticality of both (61a) and (61b). (61a) is a case of double VP extraction. (61b) repeats our earlier example (29) and is an example of a third construction that appears inside an extraposed VP. Our analysis also straightforwardly accounts for the ungrammaticality of (61c) since in the third construction we treat das Fahrrad and zu reparieren as sisters of the triggering verb zu versuchen. Therefore, these complements can only occur inside the extraposed VP. Placement of das Fahrrad inside the Mittelfeld, as in (61c), is simply not possible because uberreden requires a saturated VP as a complement. Sentences such as (61c), however, appear to be a problem for den Besten and Rutten's scrambling account since nothing seems to block "long-distance scrambling" of das Fahrrad to the left of uberreden. Of

146

E. W. Hinrichs and T. Nakazawa

course, one could stipulate restrictions on long-distance scrambling that would rule out (61c). In our analysis no extra stipulations are necessary; rather the ungrammaticality of the sentence follows from independent features of the analysis. Moreover, at least in German there are verbs that allow VP extraposition, but not the third construction. The range of verbs that are in this class seem to differ from speaker to speaker. Therefore a lexical treatment, i.e., treating the third construction by a lexical rule, seems to be preferable to a syntactic account in terms of long distance scrambling since it seems much harder to constrain such a configurational account in such a way that long-distance scrambling is blocked for verbs that allow VP extraposition, but not the third construction.20 Kiss (1993) sketches an account of the third construction that differs significantly from the extraction analyses of Uszkoreit (1987) and of den Besten and Rutten (1989) and from the argument composition analysis proposed in this paper. Kiss proposes to analyze the third construction as a coherence phenomenon and wants to account for the word order facts in terms of reordering of verbs within the verbal complex. In his analysis, thus, the third construction in (62a) is generated in terms of reordering of verbal elements of the coherent construction in (62b). The proposed structure for (62a) is given in (63):21 (62) a. .. .daB er das Auto versucht zu reparieren.

that he the car tries to repair '... that he tries to repair the car' b. ... daB er das Auto zu reparieren versucht. that he the car to repair tries '... that he tries to repair the car'

(63)

Third Construction and VP Extraposition

147

While this reordering analysis can account for the word order facts of the third construction in simple examples, the analysis does not scale up to more complex cases such as (26a), repeated here as (64): (64)

.. .daB er das Auto meint reparieren zu dtirfen. that he the car thinks repair to be allowed '... that he thinks that he is allowed to repair the car'

The reason Kiss cannot account for examples such as (64) has to do with the type of branching that he assumes for the verbal complex. Kiss assumes the right-branching structure in (65) for the coherent structure which corresponds to the third construction in (64):

(65)

As shown in (65), Kiss assumes a right-branching structure for the verbal complex which is the inverse of the semantic scope relations among verbal elements. This branching structure is a consequence of Kiss' treatment of coherent structures. Kiss characterizes control verbs for coherent constructions as subcategorizing for lexical verbs, i.e., verbs that are marked as [LEX + ], and therefore does not need to employ a feature such as NPCOMP, which is needed in our analysis to force argument raising. However, as shown in (65), it is precisely Kiss' right-branching structure that prevents him from predicting the correct word order for complex cases of the third construction. Instead of the correct word order shown in (64), Kiss would predict the ungrammatical (66) generated as the result of two levels of reordering within the verbal complex: (66) * ... daft er das Auto meint zu durfen reparieren. that he the car think to be allowed repair

148

E. W. Hinrichs and T. Nakazawa

11. OPEN PROBLEMS Some authors have pointed out (e.g., Kvam, 1983; GeifuB, 1991; and Wollstein-Leisten, 1994) that the extraposed verbal constituent in the third construction may also contain a nominal complement if the extraposed verb is ditransitive, even though the acceptability of such a sentence is marginal at least for some speakers.22 (67) a. ... daB that '... that b. .. .daB that '... that

ich Maria versucht habe das Buck zu geben. I Maria tried have the book to give I have tried to give Maria the book' ich das Buck versucht habe Maria zu geben. I the book tried have Maria to give I have tried to give Maria the book'

Our analysis of the third construction which we have developed so far does not admit sentences such as (67) since the Third Construction Lexical Rule (43) requires the extraposed verbal constituent to be marked as [NPCOMP — ]. This means that only a verbal complex or a single zw-infinitive verb can appear in this position. In order to generate sentences such as (67) we would have to eliminate the [NPCOMP — ] specification on the extraposed verbal constituent. Moreover, we would have to ensure that unsaturated VPS, i.e., VPS which are still lacking a (nominal) complement, can be licensed by an ID schema. This could be accomplished by introducing an ID schema of the sort proposed by Pollard (1996):

(68)

The ID schema licenses an indeterminate number of complement sisters of a verbal head, as long as the mother still has a nonempty COMPS list. While the rule can generate partial VP such as das Buck zu geben as in (67a), the introduction of such a rule would lead to massive spurious ambiguity in a grammar of German, as has already been pointed out by Pollard (1996): apart from the totally flat structure of the finite verb and the binary hierarchical structure of the verbal complex, the grammar would also admit an indeterminate depth of embedding partial VPS for the same string of words. However, even if one were to tolerate spurious ambiguity, the analysis for sentences such as (67) which we have just sketched has the more

Third Construction and VP Extraposition

149

serious problem that it leads to overgeneration. As mentioned in section 4, an empirical generalization is that the extraposed verbal category that can appear in the third construction is either a lexical zw-infinitive verb or a single coherent construction, but crucially not an incoherent construction. However, with the Partial VP Expansion ID Schema (68) in place, an incoherent construction as in (28), repeated here as (69), could, in fact, be admitted in the third construction: (69)

* ... daB ich Maria vergessen habe zu versprechen das Auto that I Maria forgotten have to promise the car zu reparieren. to repair '(intended)... that I have forgotten to promise Maria to repair the car'

In (69), vergessen would be the third construction triggering verb. The extraposed VP in which versprechen is missing its dative complement Maria could be generated by the Partial VP Expansion ID Schema (68). We will have nothing further to say on the topic of partial VPS in the third construction in this paper and will have to leave this matter to future research. Another issue which we will not be able to solve here, but which constitutes a potential problem for our treatment of incoherent constructions concerns sentences such as (70a) and (70b): (70) a. * ...daft Peter Maria uberredet, [Karl [das Auto that Peter Maria persuades Karl the car zu kaufen] zu zwingen] to buy to force '(intended)... that Peter persuades Maria to force Karl to buy the car' b. * ... daB Peter Maria [Karl zu zwingen [das Auto zu kaufen]] uberredet. c. ...daft Peter Maria uberredet [Karl zu zwingen [das Auto zu kaufen]]. In (70a) the extraposed VP complement of uberredet, which licenses incoherent structures, contains a nonextraposed VP das Auto zu kaufen. In (70b) the VP embedded under iiberreden as a whole is not extraposed, but this VP contains the extraposed VP das Auto zu kaufen. Both sentences are ungrammatical, while (70c), where both VPS have been extraposed, is grammatical. Thus, when one VP is in the scope of the other, one cannot be extraposed without the other. Our analysis, as it stands, treats all three examples as grammatical. One way to force double extraposition would be

150

E. W. Hinrichs and T. Nakazawa

to mark the VP complement in the lexical entry of incoherent construction control verbs as [EXTRAP + ]. However, this would at the same time exclude sentences such as (71): (71)

. . . daB er Maria zu singen uberreden konnte. that he Maria to sing persuade could '... that he could persuade Maria to sing'

While extraposition seems to be strongly preferred for control verbs such as uberreden, nonextraposed VPS do seem to be possible.23 That is precisely the reason we left the VP complement of uberreden in (40), as well as other control verbs for incoherent constructions, unspecified for the EXTRAP feature. The factors which determine whether or not VP complements are extraposed may well include extragrammatical considerations such as processing constraints. However, since such issues are beyond the scope of the present paper, we will not be able to block examples such as (70a) and (70b) in our grammar.

12. CONCLUSION We have presented an analysis of the third construction as a special case of VP extraposition. Both phenomena are analyzed in light of a wider range of data, namely, infinitive constructions in general. Following Bech (1955), we classify German infinitive constructions into coherent and incoherent constructions, and have shown that not only extraposition but also other syntactic behaviors such as topicalization, auxiliary flip, and scrambling naturally follow from the different constituent structures we assign to the two classes of constructions. In coherent infinitive constructions, control verbs and embedded verbs form a constituent and allow both upstairs and downstairs NP complements to be sisters, correctly predicting the possibility of scrambling among the NP complements and the intervention by other elements between NP complements and the verbal complements. On the other hand, in incoherent infinitive constructions, embedded complements constitute a VP, and thus allow extraposition of the VP as a whole. We attribute the difference in the syntactic structures to the lexical specifications of control verbs. The lexical entries of the control verbs which license the coherent constructions encode argument composition in the COMPS value while the control verbs for the incoherent constructions subcategorize for a full VP complement. We have argued that the analysis of the third comstruction can be based on argument composition, which is needed independently for coherent

Third Construction and VP Extraposition

151

constructions, together with linearization among raised arguments. That is, we claim that the third construction requires no extraction and is a special case of extraposition. The fact that the third construction triggering verbs also appear in both coherent and incoherent infinitive constructions is accounted for in terms of the lexical rule which maps the two classes of lexical specifications for control verbs. The Third Construction Lexical Rule takes triggering verbs which are specified to subcategorize for a VP and license incoherent constructions. The output of the lexical rule specifies the COMPS value of such verbs in terms of argument composition. It licenses either a coherent construction or the third construction, depending on whether a zu -infinitive verb is raised and extraposed or stays in the verbal complex. Thus the third construction can be nicely characterized as a case of argument composition and requires little more than the lexical entries for control verbs which are proposed for the coherent and incoherent infinitive constructions.

ACKNOWLEDGMENTS We are grateful for invaluable comments from Thilo Gotz, Georgia Green, Tilman Hohle, Tibor Kiss, Detmar Meurers, Frank Morawietz, Jerry Morgan, John Nerbonne, Klaus Netter, Dick Oehrle, Carl Pollard, Frank Richter, and Ivan Sag. Our special thanks go to Andreas Kathol, the other editor of this volume, and two anonymous referees for detailed comments on earlier versions.

NOTES These authors also observe that grammaticality judgments for the third construction can differ considerably from speaker to speaker as to the range of verbs that can trigger the third construction. Grammaticality judgments about the third construction can also be affected by processing difficulties. For example, the fact that some speakers reject clauses where the third construction appears inside an extraposed VP, such as (i), may not be a matter of grammaticality as such, but due to the difficulty of processing a sequence of two zw-infinitives. (i)

. . . daB Peter versprochen hat es zu versuchen zu reparieren.

that Peter promised has it to try '... that Peter promised to try to repair it' 2

to repair

An anonymous referee has pointed out that the placement of das and zu kaufen in (3a) could arguably be analyzed as the result of scrambling each constituent

152

E. W. Hinrichs and T. Nakazawa

separately to the beginning of the clause, rather than relativization of a single VP constituent. Such a scrambling account has, in fact, been proposed by Haider (1985). However, a scrambling analysis of VP relatives cannot account for examples such as (i) where a relativized VP das zu lesen originates from the inside of the VP complement governed by the control verb sich weigern. (i) Das ist das Buck das zu lesen er sich weigerte seiner Frau zu versprechen. That is the book which to read he self refused his wife to promise. 'That is the book which he refused to promise his wife to read.' We refer readers interested in further details of the VP relativization construction in German to van Riemsdijk (1987), De Kuthy-GroBkopf (1996), and Hinrichs and Nakazawa (1997) for relevant discussion. For a survey of possible word order variations, marginal examples, and processing considerations, we refer readers to the excellent survey in Rambow (1994). See Kiss (1994) for a similar analysis of incoherent constructions though the constituent structures assigned to coherent constructions are drastically different from the ones we propose in the following sections. 4 We assume that the possible values for the feature VFORM for German include bse, inf, and psp. They are meant to differentiate among bare infinitives such as kaufen, infinitives marked by zu such as zu kaufen, and past participles such as gekauft, respectively. The three values also correspond to the three STATUS of verbs identified by Bech (1955). 5 One piece of evidence for assuming VP constituents in coherent infinitive constructions might at first glance be derived from the fact that a VP can be topicalized, as in (i) Das Buch lesen konnen wird Peter. the book read can will Peter 'Peter will be able to read the book.' For extensive discussion of topicalization in German, see Hinrichs and Nakazawa (1994b and to appear), where we claim, following Nerbonne (1994), that VPS can appear with control verbs which license coherent structures only in topicalized position, but never in the Mittelfeld. In example (14) and the following examples, © denotes the list append operation. Also, we are assuming here, as we did in Hinrichs and Nakazawa (1994b and to appear), that subjects of finite verbs in German appear on the COMPS list, rather than as the value of SUBJ. See Borsley (1987) for detailed justification. 7 The version of the lexical entry for coherent infinitives shown in (15) is slightly revised compared to the entry proposed in Hinrichs and Nakazawa (1994a, b, and to appear) which constrained the raised material to be nonverbal. 8 Notice that the finite auxiliary appears as an immediate constituent of the clause and is not part of the verbal complex. As discussed in more detail in Hinrichs and Nakazawa (1994b and to appear), this allows us to provide a nonmovement analysis of the finite verb in German of the sort first proposed by Pollard (1996). A number of researchers, including cf. Kiss and Wesche (1991),

Third Construction and w Extraposition

153

Netter (1992), Frank (1994), and Kathol (1995) have proposed an alternative analysis that places the finite verb as the highest verb in the verbal complex. Kiss and Wesche (1991), Netter (1992), and Frank (1994) employ a SLASH mechanism to account for verb-second placement of the finite verb, while Kathol shows that a nonmovement analysis of verb-second placement is possible, even if the finite verb is embedded in a verbal complex. Since the proper treatment of placement of the finite verb in different clause types of German is not the topic of this paper, we will adopt our earlier analysis (Hinrichs and Nakazawa, 1994a) without further discussion. For the issues at hand, it is important to note, however, that the alternative analyses cited above also employ argument composition for the internal structure of the verbal complex. 9 Incidentally, Pollard and Sag (1994:150) have sketched a nonmovement analysis of VP extraposition in English in terms of a lexical rule. 10 Kathol (1995), who proposes a similar lexical analysis of third construction triggering verbs, lists each lexical entry, rather than generating them through a lexical rule. u Among the control verbs which license coherent structures, most govern base infinitives. However, a small class of control verbs which license coherent structures, e.g., scheinen 'to appear' and pflegen 'to be in the habit of,' govern zw-infinitives. 12 See Hohle (1986) and Uszkoreit (1987) for a detailed analysis of the word order variation in the German Mittelfeld. We owe this example to Andreas Kathol (personal communication). 14 Due to space limitations, prefixes for some feature paths in (38), (39) (42-45) (48)-(50), and (59) are truncated, as long as the full feature paths can be recovered from the HPSG feature geometry assumed throughout the paper. 15 Andreas Kathol (personal communication) has pointed out an example with wagen as the third construction triggering verb that may be an exception to this generalization: (i)

... daft er zu versuchen wagte, den Wagen noch einmal zu reparieren. that he to try dared the car once again to repair '... that he dared to try to repair the car once more'

Starting with Reape (1993) there have been a number of proposals, including Richter and Sailer (1995) and Kathol (1995), for more powerful theories of word order for HPSG. In these theories word order generalizations can be stated not only for the domain of local trees, as in the ID/LP rules, but also for larger domains. While the power of the ID/LP rules seems to suffice for the treatment of linear order for the third construction and VP extraposition, our use of the ID/LP rules should not be taken as an implicit rejection of these more powerful theories. 17 Please note that the ID rule in (50), as it stands, fails to account for the auxiliary flip construction. We refer interested readers to Hinrichs and Nakazawa (1994a) for a detailed discussion of this construction and of how the LP rule (50) can be further specified to properly account for the construction in an ID/LP format.

154

E. W. Hinrichs and T. Nakazawa

18

See also "long-distance passive" in Bayer and Kornfilt (1990, 1991) and "remote passive" in Pollard (1994). 19 For examples of rules for ordinary passives, see the ones proposed by Pollard and Sag (1987) for English and by Kiss (1992) for German. 20 Uszkoreit's (1987) treatment of the third construction, which he calls "focus raising," suffers from the same problems as den Besten and Rutten's analysis since Uszkoreit treats placement of the focus-raised constituent (i.e., the NP complement) by the SLASH mechanism of GPSG. Uszkoreit's analysis differs from den Besten and Rutten's in that no link is established between extraposition and focus raising. This leads to additional difficulties since nothing seems to rule out focus raising of an NP from the incoherent construction while the zw-infinitive verb is not extraposed as in (ii): (i)

... daB der Mann [VPdas Buck zu lesen] sich weigerte. that the man the book to read self refused '... that the man refused to read the book' (ii) * ... daB das Buck der Mann zu lesen sich weigerte. that the book the man to read self refused 21 As can be gleaned from the tree structure in (63), Kiss assumes a hierarchical structure for NP complements in the Mittelfeld and treats the finite verb as part of the verbal complex. 22 A den Besten and Rutten-style analysis for the third construction in German could conceivably account for data such as (67) since their notion of "long-distance scrambling" seems to apply to any number of complements of a ditransitive verb. Also see Uszkoreit (1987) for examples in which extraposed zw-infinitives are accompanied by adjuncts. 23 This observation is also made by Bech (1955).

REFERENCES Bayer, J., and J. Kornfilt (1990). Restructuring effects in German. In E. Engdahl, M. Reape, M. Mellor, and R. Cooper (eds.), Proceedings of the Workshop on Parametric Variation in Germanic and Romance: Proceedings from a DYANA Workshop, September 1989. Edinburgh Working Papers in Cognitive Science, Vol. 6, 21-42. Bayer, J., and J. Kornfilt (1991). Against scrambling as move-alpha. In Proceedings 0f NELS21, 1-15. Bech, G. (1955). Studien uber das Deutsche Verbum Infinitum. Munksgaard, Copenhagen. Bierwisch, M. (1963). Grammatik des Deutschen Verbs. Stadia Grammatica II. Borsley, R. (1987). Subjects and Complements in HPSG. Technical Report CSLI117-87. CSLI, Stanford, CA.

Third Construction and VP Extraposition

155

De Kuthy-GroBkopf, K. (1996). Der Rattenfdngereffekt bei Relativ- und Interrogativsdtzen im Deutschen. Unpublished M.A. thesis, Seminar fur Sprachwissenschaft, Eberhard-Karls-Universitat Tubingen, den Besten, H., and J. Rutten (1989). On verb raising, extraposition and free word order in Dutch. In D. Jaspers et al. (eds.), Sentential Complementation and the Lexicon: Studies in Honour of Wim de Geest. Foris Publications, Dordrecht. 41-56. Evers, A. (1975). The Transformational Cycle in Dutch and German. Dissertation, University of Utrecht. (Reproduced by the Indiana University Linguistics Club.) Frank, A. (1994). Verb Second by Lexical Rule or by Underspecification. Arbeitsberichte des Sonderforschungsbereichs 340 Nr. 43-1994, University of Stuttgart. GeilfuB, J. (1991). Zur Syntax des Verbalkomplexes. In J. GeilfuB (ed.), Verb und Verbphrasensyntax, Bericht 11-91, 59-86, Arbeitspapiere des Sonderforschungsbereichs 340, Universitat Tubingen. Haider, H. (1985). Der Rattenfangerei muB ein Ende gemacht werden. Wiener Linguistische Gazette 35-36x, 37-50. Haider, H. (1993). Deutsche Syntax—Generativ: Vorstudien zur Theorie einerprojektiven Grammatik. Narr Verlag, Tubingen. Hinrichs, E., and T. Nakazawa (1989). Flipped out: AUX in German. Chicago Linguistic Society 25, 193-202. Hinrichs, E., and T. Nakazawa (1993). Aspects of German VP Structure: An HPSG Account. Technical Report SfS-Report-01-93, Seminar fur Sprachwissenschaft, Eberhard-Karls-Universitat Tubingen. Hinrichs, E., and T. Nakazawa (1994a). Linearizing AUXs in German verbal complexes. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 11-38. CSLI Publications, Stanford, CA. Hinrichs, E., and T. Nakazawa (1994b). Topicalization in German—An HPSG analysis. In E. Hinrichs, D. Meurers, and T. Nakazawa (eds.), Partial-VP and Split-NP Topicalization—An HPSG Analysis and Its Implementation, 1-46. Bericht 58-94, Arbeitsberichte des Sonderforschungsbereichs 340, Universitat Tubingen. Hinrichs, E., and T. Nakazawa (1997). VP Relatives in German—An HPSG Analysis. Paper presented at HPSG97, the International Conference on Head-Driven Phrase Structure Grammar, Ithaca. Hinrichs, E., and T. Nakazawa (to appear). Topicalization in German—An HPSG analysis. In R. Levine and G. Green (eds.), Readings in Modern Phrase Structure Grammar. Cambridge University Press, Cambridge, UK. Hohle, T. N. (1978). Lexikalistiche Syntax: Die Aktiv-Passiv Relation und andere Infinitkonstruktionen im Deutschen. Niemeyer, Tubingen. Hohle, T. N. (1986). Der Begriff 'Mittelfeld.' Anmerkungen iiber die Theorie der topologischen Felder. In A. Schone (ed.), Kontroversen alte und neue. Akten des VII. Intemationalen Germanistenkongresses, Vol. 3, 329-340. Gottingen.

156

E. W. Hinrichs and T. Nakazawa

Kathol, A. (1994). Passives without lexical rules. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 237-272. CSLI Publications, Stanford, CA. Kathol, A. (1995). Linearization-Based German Syntax. Dissertation, Ohio State University. Kiss, T. (1992). Variable Subkategorisierung—eine Theorie der unpersonlichen Konstruktionen im Deutschen. Linguistische Berichte 140, 256-293. Kiss, T. (1993). Infinite Komplementation. Neue Studien zum deutschen Verbum Infinitum. Theorie des Lexikons: Arbeiten des Sonderforschungsbereichs 282, Nr. 42, Heinrich Heine Universitat Diisseldorf. Kiss, T. (1994). Obligatory coherence: An investigation into the syntax of modal and semimodal verbs in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 71-107. CSLI Publications, Stanford, CA. Kiss, T. and B. Wesche (1991). Verb order and head movement in German. In O. Herzog and C.-R. Rollinger (eds.), Text Understanding in LILOG. Integrating Artificial Intelligence and Computational Linguistics, 216-240. Springer-Verlag, Berlin. Kvam, S. (1983). Linksverschachtelung im Deutschen und Norwegischen: Eine kontrastive Untersuchung zur Satzverschrdnkung und Infinitivverschrankung in der deutschen und norwegischen Gegenwartssprache. Niemeyer, Tubingen. Nerbonne, J. (1994). Partial Verb Phrases and Spurious Ambiguities. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar. CSLI Publications, Stanford, CA. Nerbonne, J., K. Netter, and C. Pollard (eds.) (1994). German in Head-Driven Phrase Structure Grammar. CSLI Publications, Stanford, CA. Netter, K. (1992). On non-head non-movement. An HPSG treatment of finite verb position in German. In G. Gorz (ed.), Proceedings of KONVENS 92, 218-227. Springer-Verlag, Berlin/New York. Pollard, C. (1994). Toward a unified account of passive in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, CSLI Publications, Stanford, CA. Pollard, C. (1996). On head non-movement. In H. Bunt, and A. van Horck (eds.), Discontinuous Constituency, 279-305. de Gruyter, Berlin/New York. Pollard, C., and I. Sag (1987). Information-Based Syntax and Semantics, Vol. 1. CSLI Lecture Notes No. 13, 273-296. CSLI Publications, Stanford, CA. Pollard, C., and I. Sag (1994). Head-Driven Phrase Structure Grammar. University of Chicago Press, Chicago, and CSLI Publications, Stanford, CA. Rambow, O. (1994). Formal and Computational Aspects of Natural Language Syntax. Ph.D. thesis, University of Pennsylvania. Reape, M. (1993). A Formal Theory of Word Order: A Case Study in West Germanic. Dissertation, University of Edinburgh. Reape, M. (1994). Domain union and word order variation in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 151-305. CSLI Publications, Stanford, CA.

Third Construction and VP Extraposition

157

Richter, F., and M. Sailer (1995). Remarks on Linearization: Reflections on the Treatment of LP-Rules in HPSG in a Typed Feature Logic. Master thesis, Universitat Tubingen. Santorini, B., and A. S. Kroch (1990). Remnant Extraposition in German. Unpublished manuscript. Uszkoreit, H. (1987). Word Order and Constituent Structure in German. CSLI Lecture Notes No. 8. CSLI Publications, Stanford, CA. van Riemsdijk, H. (1987). Der Rattenfanger Effekt bei Infinitiven in deutschen Relativsatzen. In W. Abraham (ed.), Erkldrende Syntax des Deutschen, 75-98. Tubingen. Visser, F. Th. (1963-1973). An Historical Syntax of the English Language, Part 111.2:2118. F. J. Brill, Leiden. Wollstein-Leisten, A. (1994). Zur Syntax der dritten Konstruktion. Unpublished manuscript, University of Stuttgart.

This page intentionally left blank

ARGUMENT COMPOSITION AND LONG-DISTANCE SCRAMBLING IN KOREAN: AN EXTENSION OF THE COMPLEX PREDICATE ANALYSIS CHAN CHUNG Department of English Language and Literature Dongseo University Churye-Dong, Sasang-ku Pusan, Korea

1. INTRODUCTION

Debates on configurationality have arisen from the observation that languages such as Warlpiri differ from languages such as English in that the former tend to have the following properties that tend to be lacking in the latter (Hale, 1982): (1) a. great freedom of surface word order b. frequent use of discontinuous constituents c. complex verb word system (verbal complexes) d. use of a rich case system e. frequent "pronoun drop" f. lack of pleonastic NFS (e.g., it and there in English) g. lack of NP movement According to these criteria, Korean must be considered a nonconfigurational language, since it has many of the properties in (1) such as (la), (Ic), (Id), (le), and (If). Syntax and Semantics, Volume 30 Complex Predicates in Nonderiuational Syntax

159

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

160

Chan Chung

The goal of this paper is to investigate the properties in (la) and (Ic) in Korean. It is a well-known fact that Korean allows long-distance scrambling (clause-external scrambling) as well as clause-internal scrambling. The main proposal of this paper is as follows: the coexistence of the verbal complex system and long-distance scrambling in Korean is naturally accounted for, since both cases are licensed by the same syntactic mechanism, called argument composition by Hinrichs and Nakazawa (1989, 1994). Roughly speaking, argument composition is a mechanism which "attracts" the arguments of the governed verb to the argument list of the governing verb. The organization of this paper is as follows. In section 2, we discuss the auxiliary verb construction (verbal complex construction) in Korean and its relevant properties. Here we propose that the mechanism of argument composition is crucial to account for certain facts about the auxiliary verb construction. In section 3, we explore how the argument composition mechanism can be extended to account for scrambling in various s- and vp-complement constructions (e.g., embedded-clause, control-verb, raising-verb, and causative constructions), proposing that this mechanism is also crucial to describe word order variation in those constructions. This analysis entails that both the verbal complex and long-distance scrambling constructions are licensed by the same syntactic mechanism. In section 4, we discuss some theoretical predictions made by our proposal. There we discuss long-distance passivization and weak crossover effects. Section 5 is the conclusion.

2. ARGUMENT COMPOSITION AND AUXILIARY VERB CONSTRUCTION

The organization of this section is as follows. In section 2.1, we discuss the properties of auxiliary verb constructions, focusing on the fact that auxiliary + verb combinations strongly resist separation or intrusion by other syntactic elements. In section 2.2, we discuss the problematic aspects of certain previous analyses which assume that an auxiliary takes a VP or an s as its complement. In section 2.3, we provide linguistic motivation for the argument composition analysis, showing how it can account for the problems raised in section 2.2.

161

Argument Composition, Long-Distance Scrambling

2.1. Properties of Korean Auxiliary Verbs The following are examples of Korean auxiliary verbs:

Auxiliary verb

(2) a. iss 'be in the process/ state of b. siph 'want'

Verb form of the governed verb ko

ko

c. toy 'be led to'

key

d. anh 'not'

ci

e. po 'try as a test'

a/e

f. chiwu 'do resolutely'

a/e

g. noh 'do in advance'

a/e

h. cwu 'do as a favor for'

a/e

i. nay 'do to the very end/ thoroughly' j. ha 'act like'

a/e

k. ci passive

a/e

a/e

Example cako issta sleep AUX 'is sleeping' cako siphta sleep AUX 'want to sleep' cwukkey toyta die AUX 'be led to die' cwukci anhta die AUX 'do not die' ilke pota read AUX 'try reading something' phala chiwuta sell AUX 'sell something resolutely' malyenhaye nohta prepare AUX 'prepare something in advance' malyenhaye cwuta prepare AUX 'prepare something as a favor for someone' kyentye nayta endure AUX 'endure to the last' siphe hata want AUX 'act like someone wants something' sewe cita build AUX 'be built'

162

Chan Chung

One special property of the auxiliary construction is that the auxiliary and its complement verb form a syntactic unit to the exclusion of other phrasal arguments (e.g., subject, objects, etc.) as illustrated in (3): (3) a. Mary-ka sakwa-lul mekko issta. M-Nom apple-Ace eat be in the process of 'Mary is eating an apple.' b. afterthoughts i. Sakwa-lul mekko issta, Mary-ka. apple-Ace eat be in the process of M-Nom ii. Mary-ka mekko issta, sakwa-lul. M-Nom eat be in the process of apple-Ace iii. *Mary-ka sakwa-lul issta, mekko. M-Nom apple-Ace be in the process of eat iv. * Mary-ka issta, sakwa-lul mekko. M-Nom be in the process of apple-Ace eat c. parentheticals i. Mary-ka hayekan sakwa-lul mekko issta. M-Nom anyway apple-Ace eat be in the process of 'Anyway, Mary is eating an apple.' ii. Mary-ka sakwa-lul hayekan mekko issta. M-Nom apple-Ace anyway eat be in the process of iii. Mary-ka sakwa-lul mekko hayekan issta. M-Nom apple-Ace eat anyway be in the process of The sentences in (3b) show that the governed verb mekko 'eat' cannot be separated from the auxiliary issta 'be in the process of as an afterthought expression. And the sentences in (3c) show that even a parenthetical expression cannot occur between an auxiliary and its complement verb. Most auxiliaries share properties with subject-control verbs such as sitoha 'try' in that the controller of the auxiliary is the subject of the main verb. However, auxiliaries differ from the ordinary control verbs in that the auxiliary and its governed verb cannot be separated by an independent word, as shown in (3). A genuine control verb such as sitoha 'try' and its governed verb do not constitute the same kind of syntactic unit, as is illustrated in (4): (4) a. Mary-ka ku mwuncey-lul phwulye-ko sitohayessta. M-Nom the problem-Ace solve tried 'Mary tried to solve the problem.' b. Mary-ka ku mwuncey-lul phwulye-ko kkuncilkikey sitohayessta. M-Nom the problem-Ace solve ceaselessly tried c. Mary-ka sitohayessta, ku mwuncey-lul phwulye-ko. M-Nom tried the problem-Ace solve

Argument Composition, Long-Distance Scrambling

163

In (4b), the adverbial expression kkuncilkikey 'ceaselessly' intervenes between the governed verb and the control verb; in (4c), the governed verb is part of an afterthought expression and is separated from the control verb. (See section 3 for detailed discussion on the control verb construction.) There are three possible structures for the auxiliary construction, as shown in (5): (5)

I will call (5a) the vp-complement analysis, (5b) the s-complement analysis, and (5c) the complex predicate analysis. In section 2.2, we discuss the VPand s-complement analyses and their problems. In section 2.3, we introduce the notion of argument composition and show how it works in the complex predicate analysis.

2.2. VP- and s-Complement Analyses and Their Problems

The vp-complement analysis in (5a), advocated by Yoon (1993) and No (1991) among others, states that the complement of an auxiliary is a VP. One of the difficulties with this analysis is that it allows no ready explanation for the tight bond between the head and its complement in the auxiliary case vs the control verb case, as was seen in the difference of behavior in (3) vs (4). To avoid this problem, we may assume that a mismatch between syntax and morphology occurs in the Korean auxiliary verb construction, as proposed by No (1991:169-171). According to No, an auxiliary takes a VP as its complement at the level of syntax, and hence the auxiliary and the governed verb do not form a unit at the syntactic level. At the level of morphology, however, the clitic auxiliary must attach to the head of the complement VP, and hence the auxiliary and the governed verb form a composite morphological word. A problem with this proposal is that no empirical syntactic evidence exists that the auxiliary takes a VP as its complement. Rather, some syntactic phenomena that are discussed below cause problems for the vp-complement analysis. Moreover, it would be theoretically more desirable to get rid of this kind of mismatch between

164

Chan Chung

two components in a grammar whenever possible, especially when the motivation for the mismatch is not obvious. The first empirical problem with the analysis in (5 a) is a fact about passivization. In Korean, two kinds of passivization are generally assumed. One is passivization through suffixes (-/, -hi, -li, and -ki), and the other is passivization using the auxiliary ci (henceforth a'-passivization). For the current discussion, only a'-passivization is relevant because auxiliaries can only be passivized using ci. The following provides an example: (6) a. Cengpwu-ka ku kenmwul-lul seywessta. government-Norn the building-Ace constructed The government constructed the building.' b. Ku kenmwul-i cengpwu-eyuyhay seywe ci-essta. the building-Norn government-by construct passive-past 'The building was constructed by the government.' Even though not all auxiliaries can be governed by the auxiliary ci, some auxiliaries can be governed and passivized, as shown in (7) and (8):1 (7) a. Ku nongpwu-ka malssengmanhun so-lul phala the farmer-Norn troublesome cow-Ace sell chiw-essta. do resolutely-past The farmer resolutely sold the troublesome cow.' b. Malssengmanhun so-ka (ku nongpwu-eyuyhay) troublesome cow-Norn the farmer-by phala chiwe ci-essta. sell do resolutely passive-past The troublesome cow was resolutely sold (by the farmer).' (8) a. Kwukhoyuywentul-i ku pepan-ul malyenhay nohu-nikka, Congressmen-Norn the bill-Ace plan do in advance-when taythonglyeng-i pantayhayssta. president-Norn raised an objection 'When congressmen planned the bill, the president raised an objection to it.' b. Ku pepan-i (kwukhoyuywentul-eyuyhay) malyenhay the bill-Norn Congressmen-by plan noha ci-nikka, taythonglyeng-i pantayhayssta. do in advance passive-when president-Norn raised an objection 'When the bill was planned by congressmen, the president raised an objection to it.'

Argument Composition, Long-Distance Scrambling

165

If we take auxiliary constructions to have the structure in (5a), the analysis of (7a,b) would be as in (9a,b), respectively: (9) a. [s Ku nongpwu-ka [VP [VP malssengmanhun so-lul the farmer-Norn troublesome cow-Ace phala] chiw-essta]] sell do resolutely-past 'The farmer resolutely sold the troublesome cow.' b. [s Malssengmanhun so-ka [VP [VP [VP (ku nongpwu-eyuyhay) troublesome cow-Norn the farmer-by phala] chiwe] ci-essta] sell resolutely passive-past 'The troublesome cow was resolutely sold (by the farmer).' Given standard lexicalist assumptions, an object can be passivized only when it is an argument of a verb on which a passive morpheme is realized. In (9a), the object NP malssengmanhun so-lul 'the troublesome cow' is an argument of the main verb phala 'sell,' but not of the auxiliary chiwe 'do resolutely' or the passive auxiliary ciessta. Then it is not clear how we can account for the fact that the auxiliary ci affects the typical passive valence change, i.e., the primary object changing to the subject, and the subject changing to a PP complement.2 Another problem for the vp-complement analysis arises from the fact that emotion auxiliaries such as siph 'want' and silh 'dislike' can trigger a case alternation of the NP object which is directly governed by the main verb, as shown in (10) and (11): (10) a. Nay-ka sakwa-lul mekessta. I-Nom apple-Ace ate 'I ate an apple.' b. *Nay-ka sakwa-ka mekessta. I-Nom apple-Nom ate (11) a. Nay-ka sakwa-lul mekko siphta. I-Nom apple-Ace eat want 'I want to eat an apple.' b. Nay-ka sakwa-ka mekko siphta. I-Nom apple-Nom eat want The sentences in (10) show that the verb mek 'eat' assigns only accusative case to the NP object. However, when the auxiliary siph 'want' governs the verb mek 'eat,' nominative case can also be assigned to the object, as shown in (lib). If we assume the vp-complement analysis, the structure of

166

Chan Chung

the sentences in (11) is as in (12): (12) a. [s Nay-ka [VP [VP I-Nom b. [s Nay-ka [VP [VP I-Nom

sakwa-lul mekko] siphta]]. apple-Ace eat want sakwa-ka mekko] siphta]]. apple-Norn eat want

Under standard lexicalist assumptions, case assignment is a local phenomenon requiring the case assignee to be head-governed by the assignor. Following this assumption, Pollard and Sag (1987:143-144) suggest that the case assignment is one of the phenomena which observe the Locality Principle:3 there is no verb in any language that assigns case to some NP properly contained within one of its complements. A problem for the structure in (12b) is that it violates this principle. That is, to account for the fact that the auxiliary can change the case of the object NP from accusative to nominative, we need to assume that the auxiliary can assign nominative case to the NP that is properly contained within the VP complement. Another case alternation problem similar to the above arises from the "deemotion auxiliary" ha 'act like.'4 The auxiliary ha always governs an emotion verb and "deemotionalizes" it. That is, when ha attaches to an emotion verb, the whole predicate no longer has typical properties of an emotion verb. One of the effects of deemotionalization is that the combination does not allow nominative case to be assigned to the primary object of the main verb, as shown in (13) (cf. (11)): (13) Mary-ka sakwa-lul/*-ka mekko siphe hanta. M-Nom apple-Ace/-Nom eat want act like 'Mary acts like she wants to eat an apple.' In the vp-complement analysis, the structure of (13) is as in (14): (14) [s Mary-ka [VP [VP [VP sakwa-lul/*-ka mekko] siphe] hanta]]. M-Nom apple-Ace/-Nom eat want act like Again, in this analysis, it is impossible, without violating the Locality Principle, to account for the fact that the auxiliary ha takes as its complement a VP, which—if headed by a transitive verb—always contains an NP object with accusative case. The s-complement analysis in (5b) has the same problems that occur in the vp-complement analysis. That is, it is difficult to account for (i) the strong bond between the auxiliary and the governed verb, (ii) the fact that an object NP of the main verb can be passivized while the passive mor-

Argument Composition, Long-Distance Scrambling

167

pheme is not realized on the verb, and (iii) the fact that an emotion or deemotion auxiliary can alter the case of the object NP of the main verb. In addition, the s-complement analysis has one more problem. In this analysis, the auxiliary has no direct access to the subject, so it is unclear how to treat honorific subject-verb agreement, as shown in (15): (15) a.

Halapeci-kkeyse ku chayk-ul ilke grandfather-Hon.Nom the book-Ace read po-si-essta. try as a test-Hon-past 'Grandfather tried reading the book.' b. #Halapeci-kkeyse ku chayk-ul ilke po-assta. grandfather-Hon.Nom the book-Ace read try as a test-past

In the s-complement analysis, the structure of (15a) is as in (16): (16) [s [s Halapeci-kkeyse ku chayk-ul ilke} grandfather-Hon.Nom the book-Ace read po-si-essta]. try as a test-Hon-past 'Grandfather tried reading the book.' The problem is that (16) also violates the Locality Principle in that it requires the postulation of a verb agreeing with an NP contained in its s-complement. In the following section, we discuss an alternative analysis that accounts for all the data given in this section without making them exceptions to the observed regularities. This alternative is the analysis in (5c). Here, the complement of the auxiliary is a simple verb rather than a VP or s. The auxiliary and the governed verb form a complex predicate, hence the term complex predicate analysis.

2.3. Complex Predicate Analysis Hinrichs and Nakazawa (1989, 1994) introduce the mechanism of argument composition into the HPSG framework to account for the linear order of clause-final verbal sequences in German. The idea has been adopted by other HPSG researchers such as Kiss (1994), Gerdemann (1994), and Ryu (1993). The effect of argument composition is to "attract" the arguments of the governed verb to the argument list of the governing verb. In HPSG, argument composition is represented by structure sharing between the valence value of the governed verb and that of the governing

168

Chan Chung

verb.5 A simplified example from Hinrichs and Nakazawa (1994) is as follows (here X 0 |<X 1 } ..., X B > is <X 0 , X l s . . . , X n ) ; that is, X^X^..., X B > means that X0 is the first element of some list of which < X 1 ? . . . , Xn> is the rest):

(17)

In (17), the SUBCAT value (the list of the subcategorized-for elements) of the transitive verb finden is and is tagged by DO. The verb itself is tagged by DG. The SUBCAT value of the auxiliary wird is DD|[2] (i.e., E3v|[II Subj, Comp*, Head b. V -> V, Head The LP constraint in (18) simply states that a head must be preceded by any category. (19a) is an abbreviation of the Head-Subject-Complement Schema, which is basically the same as the rule suggested in Hale (1982) which states that a sentence consists of an arbitrary number of nonhead daughters (a subject and complements) and a lexical head. Rule (19b) states that v consists of a complement daughter v and a lexical head. Here, we treat the verbal clusters (the combination of a main verb and auxiliaries) as a kind of compound word. This analysis is consistent with the

Argument Composition, Long-Distance Scrambling

169

compound analysis of auxiliaries in Cho (1988) and Sells (1991) even though they do not provide any specific mechanism for the compounding, especially with regard to the subcategorization frame of the compound verb. In our analysis, the analysis outlined in (20b) is given to the example in (20a): (20) a. Mary-ka ku chayk-ul ilke poko issta. Mary-Nom the book-Ace read try be in the process of 'Mary is giving the book a trial reading.'

No (1991) points out two facts suggesting that a main verb and an auxiliary must be considered to combine by syntactic principles rather than morphological word formation rules, and hence that the combination cannot be considered a syntactic word whose internal structure is invisible from the level of syntax. The first fact concerns bound morphemes called delimiters (e.g., -man 'only,' -to 'also,' -cocha 'even') which can intervene between an inflected main verb and an auxiliary (e.g., mekko-man issta (eat-only be-in-the-process-of) 'be only eating'). No (1991) points out that delimiters need to be viewed as a syntactic category since they occur not only after a verb but also after a noun or a postposition: i.e., their distribution is not confined to categories with verbal morphology. Then it is doubtful that syntactic material such as delimiters could intervene between a verb and the auxiliary if the verbal complex is a genuine syntactic word. No (1991) also points out that the combination of an auxiliary and its governed verb is wholly productive and semantically compositional. Productivity and compositionality are typical properties of syntactic rules, but not of word-formation rules.

170

Chan Chung

To accommodate the syntactic characteristics of the verbal complex, we need to assume the sort hierarchy in (21):

(21)

The hierarchy in (21) states that the sort sign has the two subsorts phrase and nonphrase. The sort nonphrase, in turn, has the two subsorts word and complex-word. The sort word corresponds to a syntactic unit whose internal structure is not visible from the level of syntax. It includes all kinds of inflected or derived simple words which belong to the subsort simple-word, and morphological compound words (e.g., blackboard) which belong to the subsort compound. The sort complex-word subsumes various kinds of so-called complex predicate constructions, e.g., the verbal noun construction discussed in Ryu (1993) and the auxiliary construction in question. It takes the attribute LEX-DTRS (lexical daughters) whose value is the new sort complex-word-structure with the features HEADWORD (the head word) and GOVERNEE (the governed word). The composition of objects of sort complex-word-struc is governed by syntactic principles and rules, as is the composition of subject and complement daughters with the head in a phrase. Complex words are licensed by the rule in (22), which states that a head word selects a nonphrase (a word or a complex word). Here the selection is made by the new valence feature "GOV," which takes as its value a singleton list containing the governed word.

(22)

Note that (22) must be regarded not as a word-formation rule but as a syntactic schema, in the same way that objects of sort phrase are constructed by a syntactic schema. For example, in this analysis, the valence features for the auxiliary iss 'be in the process of are as in (23), which embodies the assumption that the COMPS list of the verb selected by an auxiliary is inherited to the COMPS

Argument Composition, Long-Distance Scrambling

171

list of the auxiliary:

(23)

iss SUBJ CD COMPS \2\ GOV

(23) states the following: (i) the auxiliary iss selects a verb which has VFORM (verb form) -ko; (ii) the COMPS value of the auxiliary is structureshared with that of its governee verb, represented by [U; and (iii) the SUBJ value of the auxiliary is structure-shared with that of the governee verb, represented by CD. The valence features of most auxiliaries in Korean are the same as (23).7 On this approach, the sentence in (20a) is analyzed as follows:8

(24)

172

Chan Chung

The first advantage of the complex predicate approach is that the strong cohesion effects within verbal clusters is directly captured in terms of a structural unit, and hence there is no need to postulate the kind of syntax and morphology mismatch assumed in No (1991). The second advantage is that there is no need for a mechanism such as move-a or a VP-flattening metarule as proposed by Pullum (1982) to account for the scrambling phenomena illustrated in (25): (25) a. Mary-ka ku chayk-ul ilke poko issta. Mary-Nom the book-Ace read try be in the process of 'Mary is giving the book a trial reading.' b. Ku chayk-ul Mary-ka ilke poko issta. the book-Ace Mary-Nom read try be in the process of In our analysis, the arguments of the verbal complex (i.e., the subject Mary-ka and the object ku chayk-ul) are all sisters of each other as shown in (24), and scrambling results from the lack of LP constraints among the arguments. The third advantage is that the problems raised in the vp-complement analysis disappear. In the rest of this section, we will show how the problems of passivization (e.g., (7) and (8)) and case alternation (e.g., (11) and (13)) are solved in the complex predicate analysis. For d-passivization shown in (7) and (8), we assume the following lexical entry for the auxiliary ci (cf. Kathol, 1994) (here, "©" denotes the list append operation):

(26)

The lexical description in (26) states the following: (i) the auxiliary ci takes a verb as its governee; (ii) the subject of the governee (i.e., NP Q ) is coindexed with the pp[uyhay], which is a value of the COMPS list of the auxiliary; (iii) the first element of the COMPS list which is the primary object of the governee (i.e., ENP) is structure-shared with the subject of the auxiliary; (iv) the rest of the COMPS list of the governee (i.e., [I]) is attracted to the COMPS list of the auxiliary.

Argument Composition, Long-Distance Scrambling

173

On this approach, the analysis of the sentence in (7b) is as in (27): (27)

In (27), the primary object of the verb phala 'sell,' malssengmanhun so 'troublesome cow' (i.e., ONP), is attracted to the COMPS list of the auxiliary chiwu 'do resolutely,' percolates up to the mother node, and is attracted to the SUBJ list of the auxiliary ciessta (passive). The subject of the main verb phala, ku nongpwu 'the farmer' (i.e., HNP^), is attracted to the SUBJ list of the auxiliary chiwu, percolates up to the mother node, and is coindexed with the pp[uyhay] in the COMPS list of the auxiliary ciessta. Here the COMPS value of the auxiliary ciessta is just (pp[uyhay]^) since the rest of the COMPS list of the GOV minus the first element (the primary object) is the empty list. Before discussing case alternations triggered by emotion auxiliaries, we will briefly introduce the case assignment system utilized in this paper. Following Heinz and Matiasek (1994) and Pollard (1994), Yoo (1993)

174

Chan Chung

introduces the notion of "structural case" into HPSG to account for case assignment in Korean. The basic idea is that the case of some NFS is not lexically assigned, but is just specified as structural) in the lexicon and surfaces as either nominative or accusative depending on the syntactic context. The Case Principle for Korean proposed by Yoo is as follows: (28) Case Principle: A structural NP which is a daughter of a phrase a is [nom] if it is a SUBJ-DTR of a, and [ace] if it is a COMP-DTR of a. (Yoo, 1993:189) Case Principle (28) states that if structural case is assigned to an NP, it is realized as nominative case or accusative case when the NP is a subject daughter or a complement daughter, respectively. Yoo (1993) also assumes another type of case, namely, lexical case, to account for some other idiosyncratic case marking phenomena. In Korean, as mentioned already, certain verbs such as (de)emotion verbs directly assign nominative or accusative case to an NP. Such special cases are exempted from the principle stated in (28). Let us consider the following emotion verb construction as an example of lexical case: (29) a. Nay-ka sakwa-ka cohta. I-Nom apple-Norn like 'I like apples.' b. * Nay-ka sakwa-lul cohta. I-Nom apple-Ace like The valence structure for the emotion verb coh 'like' is as follows: (30)

In (30), nominative case is directly assigned to the complement NP sakwa 'apple' by the lexical entry, while structural case is assigned to the subject NP na 'I,' which is realized as bearing nominative case via the Case Principle. For case alternation triggered by emotion auxiliaries, we assume two kinds of lexical entries for emotion auxiliaries as in (31):9 (31)

Argument Composition, Long-Distance Scrambling

175

The feature structure in (3la) indicates that the COMPS value of the governed verb is directly attracted to the COMPS value of the emotion auxiliary. This lexical entry is responsible for sentences like (11 a). In (3la), the COMPS list of the governee is structure-shared as a whole with that of the auxiliary, and hence structural case assigned to the object of the governee is inherited to the COMPS list of the auxiliary without any change. Then by the Case Principle in (28), accusative case is assigned to the object NP. In contrast, the lexical entry in (31b) indicates (i) that the first NP of the COMPS list of the governee has structural case (i.e., NpUfr]^) and is coindexed with the first NP of the COMPS list of the auxiliary, (ii) that nominative case is assigned to the coindexed NP in the COMPS list of the auxiliary (i.e., ^[nom]^, and (iii) that the rest of the COMPS list of the governee, represented by GO, is attracted to the COMPS list of the auxiliary. This lexical entry is responsible for sentences such as (lib), which contain an object with nominative case.10 An example of another case alternation phenomenon is in (13), which is due to the deemotion auxiliary ha 'act like.' Here due to the deemotion auxiliary ha, the case of the object must be accusative even though it is governed by the emotion auxiliary siphe 'want.' To account for this phenomenon, we assume the following lexical entry for ha:

(32)

The feature structure in (32) indicates (i) that if the auxiliary ha takes as its governee a verb whose first COMPS value is an NP (the primary object represented by NP 0 ), the NP is coindexed with the first NP in the COMPS list, (ii) that accusative case is assigned to the coindexed NP in the COMPS list of the auxiliary (i.e., Np[acc]s), and (iii) that if such an NP (primary object) does not exist in the COMPS list of the governee, then the whole COMPS list, represented by CD, is attracted to the COMPS list of the auxiliary. In the complex predicate analysis, the COMPS value of the governed verb is directly accessible from the governing verb, and hence the governing verb can select a certain category as a governee and specify the case alternation of the governee's complement. Due to this selectability, we can exclude the ill-formed sentence in (13), which has an NP complement with nominative case, without violating the Locality Principle. As shown above, the emotion auxiliary can change the case of the primary object of the governed verb from accusative to nominative. However, when emotion auxiliaries such as siph 'want' govern the deemotion

176

Chan Chung

auxiliary ha 'act like,' the case of the primary object cannot be altered to nominative, as shown in (33): (33) a. Nay-ka sakwa-lul coha hako siphta. I-Nom apple-Ace like act like want 'I want to act like I like apples.' b. *Nay-ka sakwa-ka coha hako siphta. I-Nom apple-Norn like act like want From this fact, we may assume that the case alternation triggered by emotion auxiliaries is possible only when the case of the primary object of the governed verb is not lexically specified, i.e., only when its case is structurally assigned. Our approach correctly predicts this fact. The structure for (33a) is as follows: (34)

The lexical description indicated by node (a) above is an instance of (3la). The lexical entry in (31b) is not applicable to this node because (31b) is the lexical entry for the emotion auxiliary which takes as its governee a verb whose complement NP has structural case. That is, the lexical entry in node (a) has to take as its governee a verb whose complement NP is accusative,

Argument Composition, Long-Distance Scrambling

177

which is lexically assigned by the deemotion auxiliary ha 'act like' in (32). Therefore, the ill-formed sentence in (33b) is never licensed on our approach.

3. LONG-DISTANCE SCRAMBLING AND ARGUMENT COMPOSITION In the previous section, we discussed the auxiliary construction, proposing that the notion of argument composition is crucial to account for this construction. In this section, we extend the theory of the auxiliary construction, proposing that the mechanism of argument composition can be extended to treat word order variation facts in complex clauses. Instances of complex clauses include the VP complement constructions (e.g., the control, raising, and ha -causative constructions) and embedded clause constructions in which the matrix verb takes an s as its complement. In Korean, arguments of the embedded verb can be interleaved with arguments of the matrix verb, as shown in (35) and (36): (35)

VP Complement Construction a. Mary-ka John-hanthey [VP ku chayk-ul ilkulako] seltukhayssta. M-Nom J-Dat the book-Ace read persuaded 'Mary persuaded John to read the book.' b. Ku chayk-ul Mary-ka John-hanthey ilkulako seltukhayssta. the book-Ace M-Nom J-Dat read persuaded c. Mary-ka ku chayk-ul John-hanthey ilkulako seltukhayssta. M-Nom the book-Ace J-Dat read persuaded

(36)

Embedded Clause Construction a. Nay-ka John-hanthey [s Mary-ka ku chayk-ul ilkesstako] I-Nom J-Dat M-Nom the book-Ace read malhay cwu-essta. tell as a favor for-past 'I told John that Mary read the book as a favor for him.' b. Ku chayk-ul nay-ka John-hanthey Mary-ka ilkesstako the book-Ace I-Nom J-Dat M-Nom read malhay cwu-essta. tell as a favor for-past c. Na-nun ku chayk-ul John-hanthey Mary-ka ilkesstako I-Top the book-Ace J-Dat M-Nom read malhay cwu-essta. tell as a favor for-past

178

Chan Chung

The goal of this section is to explore the syntactic mechanism(s) with which such scrambling facts can be accounted for. What is proposed is that scrambling out of VP complements or embedded clauses is not an instance of an unbounded dependency construction commonly handled via the SLASH feature in GPSG and HPSG, or A-bar movement in GB, but rather a local phenomenon treated by argument composition. The organization of this section is as follows. In section 3.1, we discuss evidence for a VP constituent in Korean and various data showing different scrambling properties of the different vp-complement constructions. In section 3.2, we propose our own theory of word order variation in vp-complement constructions through the mechanisms of argument composition and lexical rules. We also extend our theory to account for scrambling out of a complement clause and adjunct scrambling. 3.1. VP-Complement Constructions in Korean

3.1.1. VP CONSTITUENTS IN KOREAN In this paper, we assume that Korean has a flat clause structure and that clause-internal scrambling results from the relative freedom of linear precedence constraints among the nonhead constituents at the sentence level. Then the question arises whether VP constituents exist at all in Korean. Even though we assume that Korean does not have a VP node that makes the clausal structure hierarchical, it does not necessarily follow that Korean does not have any VP constituents at all. A weaker assumption is to deny the existence of a subject-predicate schema, such as the one in (37): (37)

In fact, control and raising verbs can be shown to subcategorize for a VP. In such cases, a VP is one of the complements of a particular lexical head and thus is a sister to the head. Accordingly the sentence in (35a) is licensed by the flat schema in (38): (38) Head-Subiect-Complement Schema:

Schema (38) states that a head which is a nonphrase (i.e., a simple word or complex word) combines simultaneously with the subject and complements

Argument Composition, Long-Distance Scrambling

179

to form a complete sentence. Here, the head's GOV value is the empty list, and thus any GOV value will have to be discharged within a verbal complex. It is not a new idea to assume that only a certain type of VP constituent exists in a language. For instance, the normal order in a finite clause in Welsh is verb-subject, as shown in (39): (39) Gwelodd Emrys y ddraig. saw Emrys the dragon 'Emrys saw the dragon.' However, the normal order in a nonfinite clause is subject-verb, as illustrated in (40): (40) Disgwyliodd Emrys [ i Gwyn weld Megan]. expected Emrys to Gwyn see Megan 'Emrys expected Gwyn to see Megan.' To account for these facts about Welsh word order and related phenomena, Borsley (1989) proposes that a finite clause in this language is licensed by the schema in (41), which states that a sentence consists of a lexical head and an arbitrary number of complement daughters, while a nonfinite clause is licensed by the schema in (37).

(41) S -> V, C* Thus, on this account, the structure of (40) is as follows: (42) [s Disgwyliodd Emrys [s- i [s Gwyn [VP weld Megan]]]]. expected Emrys to Gwyn see Megan 'Emrys expected Gwyn to see Megan.' Kuh (1990) independently claims that Tamil has a flat finite clause structure but has a nonfinite VP constituent. In Tamil, a subject, an object, and a verb can occur in any order when the object is marked with accusative case. To account for this freedom of word order, Kuh assumes that Tamil has a flat clause structure and has no rule equivalent to (37). On the other hand, infinitival VPS do exist, for instance, as complements of control verbs such as anuppi 'send' as shown in (43): (43) a. Naan raaman-e [VP kkannan-e ppaakk-a] anuppineen. I-Nom Raman-Ace Kannan-Acc see-Inf sent-l.SG 'I sent Raman to see Kannan.' b. Naan [VP kkannan-e ppaakk-a] raaman-e anuppineen. I-Nom Kannan-Acc see-Inf Raman-Ace sent-l.SG c. *Kkannan-e naan ppaakk-a raaman-e anuppineen. Kannan-Acc I-Nom see-Inf Raman-Ace sent-l.SG

180

Chan Chung

In this construction, as shown in (43c), the argument within the VP complement cannot permute with any of the control verb's sisters, even though the subject of the control verb, the controller raaman-e, and the VP complement can occur in any order. Kuh states that this fact cannot be explained satisfactorily unless we assume that an infinitival VP category exists in this construction and that the sentences in (43a,b) are licensed by the rule in (41). Then what about Korean? Is there any evidence showing the existence of VP constituents? There seem to be at least two pieces of evidence that suggest the existence of VP constituents as complements of control and raising verbs. First, let us consider examples (44b) and (44c), where the verbal forms ku chayk-ul ilkulako 'to read the book' and ilkulako 'read' are afterthought expressions, respectively. (44) a. Mary-ka John-hanthey ku chayk-ul ilkulako seltukhayssta. M-Nom J-Dat the book-Ace read persuaded 'Mary persuaded John to read the book.' b. *Mary-ka John-hanthey ku chayk-ul seltukhayssta, ilkulako. M-Nom J-Dat the book-Ace persuaded read c. Mary-ka John-hanthey seltukhayssta, [VP ku chayk-ul ilkulako]. M-Nom J-Dat persuaded the book-Ace read In Korean, a lexical expression cannot be used as an afterthought expression, and thus the sentence in (44b) in which the embedded verb ilkulako 'read' alone is used as an afterthought expression is not acceptable. This strongly suggests that, in (44c), the embedded verb and its direct object from a phrasal constituent which is eligible to occur as an afterthought expression.11 Another piece of evidence for VP constituents is provided by the raising construction. Let us consider the examples in (45): (45) a.

Mary-ka John-ul New York-ey isstako mitnunta. M-Nom J-Acc New York-at stay believe 'Mary believes John to be/stay in New York.' b. ??/*New York-ey Mary-ka John-ul isstako mitnunta. New York-at M-Nom J-Acc stay believe c. ?? Mary-ka New York-ey John-ul isstako mitnunta. M-Nom New York-at J-Acc stay believe

In the raising construction, a constituent (New York-ey) directly governed by the embedded verb (isstako 'exist') cannot be interleaved with arguments (Mary-ka and John-ul) that are directly governed by the raising verb

Argument Composition, Long-Distance Scrambling

181

(mitnunta 'believe'), parallel to the control construction in Tamil (cf. (43)). As proposed in Kuh (1990), this fact can be straightforwardly accounted for if we assume that New York-ey isstako 'to be/stay in New York' forms a VP constituent. From the discussions above, we conclude that even though Korean does not possess the subject-head schema in (37), it nevertheless has nonfmite VP constituents. In the following sections, we will discuss various kinds of vp-complement constructions such as the control, raising, and ha -causative constructions, focusing on their scrambling properties. 3.1.2. CONTROL CONSTRUCTION An example of the control construction in Korean is given in (35), repeated here as (46): (46)

Mary-ka John-hanthey [VP ku chayk-ul ilkulako] seltukhayssta. M-Nom J-Dat the book-Ace read persuaded 'Mary persuaded John to read the book.'

Here the control verb seltukhayssta 'persuaded' subcategorizes for a VP complement and selects its verb form. In the HPSG framework, the lexical entry of the object-control verb is as in (47), where the SUBJ value of the VP complement (NPtsfr-]^) is coindexed with the Np[dat]^ complement:12 (47)

Here, it is important to realize that not the whole SYNSEM value of the subject of the VP complement is structure-shared with the controller in (47). Otherwise, the subject of the VP would also have to be a N?[dat]. In the context of the current discussion, scrambling facts are of particular interest. As shown in (48), the order among the subject, the controller, and the VP complement is free: (48) a. Mary-ka [VP ku chayk-ul ilkulako] John-hanthey seltukhayssta. M-Nom the book-Ace read J-Dat persuaded 'Mary persuaded John to read the book.' b. John-hanthey Mary-ka [VP ku chayk-ul ilkulako] seltukhayssta. J-Dat M-Nom the book-Ace read persuaded

182

Chan Chung

c. John-hanthey [VP ku chayk-ul ilkulako] Mary-ka seltukhayssta. J-Dat the book-Ace read M-Nom persuaded d. [VP Ku chayk-ul ilkulako} John-hanthey Mary-ka seltukhayssta. the book-Ace read J-Dat M-Nom persuaded More interestingly, an argument of a governed verb (e.g., ilkulako 'read' in (48a)) can also be scrambled with arguments of the control verb, as shown in (49): (49) a. Ku chayk-ul Mary-ka John-hanthey ilkulako seltukhayssta. the book-Ace M-Nom J-Dat read persuaded 'Mary persuaded John to read the book.' b. Mary-ka ku chayk-ul John-hanthey ilkulako seltukhayssta. M-Nom the book-Ace J-Dat read persuaded However, the sentence in (50), where the object (ku chayk-ul 'the book') of the governed verb (ilkulako 'read') linearly follows the verb, is ungrammatical:13 (50)

*Mary-ka John-hanthey ilkulako ku chayk-ul seltukhayssta. M-Nom J-Dat read the book-Ace persuaded 'Mary persuaded John to read the book.'

Also note that the control verb construction allows the following scrambling possibilities: (51) a. Ku chayk-ul John-hanthey ilkulako Mary-ka seltukhayssta. the book-Ace J-Dat read M-Nom persuaded 'Mary persuaded John to read the book.' b. Ku chayk-ul Mary-ka ilkulako John-hanthey seltukhayssta. the book-Ace M-Nom read J-Dat persuaded In (51), the string ku chayk-ul ilkulako 'to read the book' does not constitute a VP complement, nor do seltukhayssta 'persuaded' and ilkulako 'to read' form a verbal complex. This construction will be analyzed in section 3.2.1.1.

3.1.3. RAISING-TO-OBJECT CONSTRUCTION14 In section 3.1.2, we showed that an argument of a governed verb can be scrambled with arguments of the governing verb when the governing verb is a control verb. However, when the governing verb is a raising verb, scrambling seems to be more restricted. The following setence is a typical

Argument Composition, Long-Distance Scrambling

183

example of the raising-to-object construction: (52) Mary-ka John-ul [VP cengcikhatako] mitnunta. M-Nom J-Acc be honest believe 'Mary believes John to be honest.' In HPSG, a raising verb takes as its complements a controller NP and a VP whose subject is identical to the controller NP (cf. Pollard and Sag, 1987, 1994, for English, and Yoo, 1993, for Korean, among others). It differs from a control verb in that the latter assigns a semantic role (e.g., the role of PERSUADEE in (47)) to the controller, whereas the former does not assign any semantic role to the controller. The other important difference is that in the control construction, the controller and the understood subject of the VP complement are coindexed (roughly, they are only semantically identical), whereas in the raising construction, the whole SYNSEM values of the controller and the understood subject of the VP are structure-shared (i.e., they are identical syntactically as well as semantically). On this approach, the lexical entry of the raising verb mit 'believe' is as follows:

(53)

As for scrambling in the raising construction, let us consider the following examples: (54) a.

b. c. d. e.

Mary-ka John-ul [vp New York-ey isstako] mitnunta. M-Nom J-Acc New York-at stay believe 'Mary believes John to be/stay in New York.' John-ul Mary-ka [yp New York-ey isstako} mitnunta. J-Acc M-Nom New York-at stay believe John-ul [VP New York-ey isstako] Mary-ka mitnunta. J-Acc New York-at stay M-Nom believe ??/*Mary-ka [VP New York-ey isstako] John-ul mitnunta. M-Nom New York-at stay J-Acc believe ??/*[Vp New York-ey isstako} John-ul Mary-ka mitnunta. New York-at stay J-Acc M-Nom believe

The above examples show that the subject, the controller and the VP complement can be scrambled as long as the controller precedes the VP

184

Chan Chung

complement. These scrambling possibilities in the raising construction show a sharp contrast with those in the control construction, since such a restriction does not exist in the latter (cf. (48a,d)). Also, an argument of the embedded verb is only marginally allowed to occur among the argument(s) of the raising verb, as shown in the examples in (55), which are variations of (54a). Such a restriction does not exist in the control construction (cf. (49)). (55) a. I!/*New York-ey Mary-ka John-ul isstako mitnunta. New York-at M-Nom J-Acc stay believe 'Mary believes John to be/stay in New York.' b. ?? Mary-ka New York-ey John-ul isstako mitnunta. M-Nom New York-at J-Acc stay believe As with control constructions, the embedded verb must follow its argument, as shown in (56) (cf. (50)): (56) * Mary-ka John-ul isstako New York-ey mitnunta. M-Nom J-Acc stay New York-at believe 'Mary believes John to be/stay in New York.' To summarize, scrambling possibilities in the raising construction are more restricted than those in the control construction. It seems that a constituent within the VP complement of the raising verb can be scrambled with the constituents outside the VP only with great difficulty, if at all. Also, the controller in the raising construction must precede the complement VP. This construction will be analyzed in section 3.2.1.2. In the following section, we will disucss another kind of VP complement construction, involving the ha -causative construction, which differs from either of the preceding ones in its scrambling properties. 3.1.4. //a-CAUSATivE CONSTRUCTION In Kang (1988), No (1991), and Chung (1993) among others, the causative verb ha 'cause' is considered as an auxiliary verb. However, this does not seem to be correct because there is an obvious difference between the causative verb ha and ordinary auxiliary verbs such as siph 'want,' as shown in (57) and (58): (57) a.

Mary-ka John-hanthey/-ul ku chayk-ul ilkkey hayssta. M-Nom J-Dat/-Acc the book-Ace read caused 'Mary caused John to read the book.' b. (?) John-hanthey/-ul ku-chayk-ul ilkkey Mary-ka hayssta. J-Dat/-Acc the book-Ace read M-Nom caused

Argument Composition, Long-Distance Scrambling

185

(58) a. Nay-ka ku chayk-ul ilkko siphta. I-Nom the book-Ace read want 'I want to read the book.' b. *Ku chayk-kul ilkko nay-ka siphta. the book-Ace read I-Nom want The sentence in (57b) shows that the causative verb hayssta 'caused' can be separated from the governed verb ilkkey 'read' by the main verb subject Mary-ka, especially when the subject has focal stress. However, the auxiliary verb siphta 'want' in (58) can never be separated from the governed verb ilkko 'read' as discussed in section 2. This difference suggests that causative ha has to be considered distinct from the category of auxiliaries. Syntactically, there are three different instances of ha according to the case borne by the causee, as shown in (59): (59) a. Mary-ka John-i ku chayk-ul ilkkey hayssta. M-Nom J-Nom the book-Ace read caused 'Mary caused John to read the book.' b. Mary-ka John-ul ku chayk-ul ilkkey hayssta. M-Nom J-Acc the book-Ace read caused c. Mary-ka John-hanthey ku chayk-ul ilkkey hayssta. M-Nom J-Dat the book-Ace read caused The causee NP is marked nominative in (59a), accusative in (59b), and dative in (59c). Semantically, these fall into two classes, (59a) and (59b) on the one hand, and (59c) on the other. Kang (1988) observes that ha in (59c) requires that Mary did something directly to John so that John should read the book, e.g., by intentionally forcing him. That is, (59c) cannot be appropriately uttered in a situation in which Mary unintentionally recommended a book to John, e.g., she recommended a book to the others without noticing John's presence, but somehow John was motivated to read the book. In contrast, ha in (59a,b) does not necessarily require Mary's action to be directed to John and can be appropriately uttered in the situation mentioned above. This semantic distinction seems to correspond to the distinction between the control and raising constructions discussed in sections 3.1.2 and 3.1.3. That is, in (59c), ha directly assigns a causee role to John, as a control verb assigns a semantic role to the controller, whereas in (59a,b), ha does not assign such a role to John, as a raising verb does not assign a semantic role to the controller. This classification for ha as a control vs a raising verb is straightforwardly

186

Chan Chung

expressed in the following lexical descriptions:

(60) a.

The sentence in (59a) is licensed by (60a), which indicates that ha takes a sentence as its complement. Here John has nominative case because it is the subject of the s complement.15 Also note that no role is assigned to the subject of the s complement, John, and thus there is no requirement to construe John as causally affected by the causer's action. The sentence in (59b) is licensed by (60b), which indicates that ha takes as its complements an NP with structural case and a VP whose subject's SYNSEM value, represented by DO, is structure-shared with that of the complement NP. As in (60a), this type of ha assigns no semantic role to the complement NP, and again nothing requires John to be construed as causally affected. The sentence in (59c) is licensed by (60c), which states that this instance of ha takes a dative NP and a VP as its complements. Unlike the other types of ha, here a causee role is assigned to the NP complement. Since we are now discussing vp-complement constructions, let us focus on (59b,c) and (60b,c). As already suggested, the lexical entry in (60b) looks the same as that of a raising verb (cf. (53)), and the lexical entry in (60c) resembles that of a control verb (cf. (47)). In other words, in our analysis, the ha -causative construction with an accusative controller (accusative

Argument Composition, Long-Distance Scrambling

187

/ia-causative henceforth) forms a natural class with the raising construction, while the ha -causative construction with the dative controller (dative /za-causative henceforth) forms a natural class with the control construction. This classification is also supported by certain scrambling data. As illustrated in (61), the dative /ifl-causative allows the causee to occur between the VP complement and ha, as the control construction does (cf. (48a)): (61) (l)Mary-ka [VP ku chayk-ul ilkkey] John-hanthey hayssta. M-Nom the book-Ace read J-Dat caused 'Mary caused John to read the book.' Even though (61) is slightly marginal compared with the counterpart in the control construction, it is acceptable to most speakers, especially when focal stress is given to the causee. The dative /w-causative and control constructions also exhibit similarities of behavior in other types of scrambling, as illustrated in (62): (62) a. Ku chayk-ul Mary-ka John-hanthey ilkkey hayssta. the book-Ace M-Nom J-Dat read caused 'Mary caused John to read the book.' b. Mary-ka ku chayk-ul John-hanthey ilkkey hayssta. M-Nom the book-Ace J-Dat read caused That is, in both constructions, an argument of the embedded verb is allowed to intermingle with arguments of the matrix verb (cf. (49)). In contrast, the accusative ha -causative and raising constructions show the same pattern of scrambling in that neither allows the controller to occur between the VP complement and the matrix verb, as shown in (63) (cf. (54d)): (63)

*Mary-ka [VP ku chayk-ul ilkkey] John-ul hayssta. M-Nom the book-Ace read J-Acc caused 'Mary caused John to read the book.'

However, the accusative ha -causative construction differs from the raising construction in that it also has some of the scrambling properties of the control construction. As is shown in (64), an argument of the embedded verb (cip-ulo) occurs among the arguments of the ha verb (cf. (49)): (64) a. Mary-ka cip-ulo John-ul tolakakey hayssta. M-Nom home-to J-Acc go back caused 'Mary caused John to go back home.' b. Cip-ulo Mary-ka John-ul tolakakey hayssta. home-to M-Nom J-Acc go back caused

188

Chan Chung

The important aspects of scrambling discussed in this section are summarized in (65): (65)

Placement of the matrix subject between the embedded verb and the matrix verb Placement of a complement of the embedded verb between the embedded verb and the matrix verb Scrambling between the VP complement and the controller Scrambling of an argument of the embedded verb with the argument(s) of the matrix verb

Accusative Control Raising ha -causative Yes Yes Yes

Dative ha -causative Yes

No

No

No

No

Yes

No

No

Yes

Yes

No

Yes

Yes

In the next section, we will propose that scrambling out of VP complements should not be considered an unbounded dependency construction, but instead as a local phenomenon to be treated via argument composition and lexical rules. 3.2. Argument Composition, Lexical Rules, and Word Order Variation in Complex Clauses The organization of this section is as follows. In section 3.2.1, we suggest how scrambling in various vp-complement constructions is appropriately handled via argument composition and lexical rules. Here we also propose some LP constraints to account for restrictions on scrambling. In sections 3.2.2 and 3.2.3, we extend this analysis to account for scrambling in the s-complement construction and adjunct scrambling, respectively. 3.2.1. ANALYSES OF VP-COMPLEMENT CONSTRUCTIONS 3.2.1.1. Control and dative ha-causative constructions. Following Pollard and Sag (1994), Yoo (1993) assumes that vp-complement constructions in Korean have a flat structure. For example, (66) is the structure of the

Argument Composition, Long-Distance Scrambling

189

control construction (e.g., (46)) which is licensed by the lexical entry in (47) for the control verb seltukhayssta 'persuaded':

(66)

The flat structure in (66) can account for scrambling among the NP 0 , NP E , and VP (e.g., (48)). However, it cannot account for the fact that the direct object ku chayk-ul 'the book,' which is governed by the embedded verb ilkulako 'read,' can be scrambled with arguments directly governed by the main verb seltukhayessta 'persuaded' (e.g., (49) and (51)). To account for this, we propose that a control verb like seltukhayssta 'persuaded' can take as its complement not only the VP but also the embedded verb ilkulako 'read.' When the control verb takes a verb as its complement, the complement(s) of the complement verb is (are) attracted to the COMPS list of the control verb, as in the auxiliary constructions. To this end, we propose that a control verb that takes a VP as its complement is related to the other control verb that takes a v as its complement by the lexical rule in (67).16

(67)

The lexical rule in (67) takes as input a lexical entry selecting a VP complement and a controller NP^, and gives as output a lexical entry that requires the controller, a verb, and the verb's complements). Thus, in the output lexical entry, the COMPS list of the complement verb represented by ID is attracted to the COMPS list of the control verb.

190

Chan Chung

On our approach, the sentence in (46) also can be analyzed as in (68a), which is licensed by the lexical entry of the control verb in (68b). (68b) is an instance of the output entry of the lexical rule in (67).

(68) a.

Then, the fact that an argument of the embedded verb occurs among arguments of the main verb in (49) and (51) is accounted for by the flat analysis illustrated in structure (68a). Note that the lexical rule in (67) is schematized to account for scrambling in the subject-control construction shown in (69) and (70): (69) a. Mary-ka John-hanthey [VP ku chayk-ul pilyecwu-kesstako] M-Nom J-Dat the book-Ace lend-will yaksokhayssta. promised 'Mary promised John to lend the book.' b. Mary-ka ku chayk-ul John-hanthey pilyecwu-kesstako M-Nom the book-Ace J-Dat lend-will yaksokhayssta. promised c. Ku chayk-ul John-hanthey pilyecwu-kesstako the book-Ace J-Dat lend-will Mary-ka yaksokhayssta. M-Nom promised (70) a. Mary-ka [w ku chayk-ul ilkulyeko] sitohayssta. M-Nom the book-Ace read tried 'Mary tried to read the book.' b. Ku chayk-ul Mary-ka ilkulyeko sitohayssta. the book-Ace M-Nom read tried c. Ku chayk-ul ilkulyeko Mary-kay sitohayssta. the book-Ace read M-Nom tried

Argument Composition, Long-Distance Scrambling

191

That is, rule (67) states (i) that the verb optionally selects the NP complement, which is indicated by the parentheses (NPS), and (ii) that the understood subject of the VP complement is coindexed with either the subject or the complement of the control verb. Then the lexical entry yaksokhayssta 'promised' in (71), for instance, feeds the lexical rule in (67), producing an entry which licenses the flat structure:

(71)

The control construction differs from the auxiliary constructions discussed in section 2, in that the control construction allows syntactic material to intervene between the governed verb and the control verb, while such an intervention is not allowed in the auxiliary construction. This difference is reflected in the constituent structure. In the auxiliary construction, the governed verb and the auxiliary verb form a complex word. In the control construction, in contrast, the governed verb and the control verb do not form a verbal complex, as shown in (68a), and thus syntactic material can intervene between the control verb and its complement verb as shown in (51), (69c), and (70c). As shown in (50), however, an embedded verb in the control construction must follow its complement. This generalization is captured by the following LP constraint which instantiates the clause-final occurrence of heads in Korean: (72)

SYNSEMDG
]

As shown in table (65), the dative ha -causative construction has the same scrambling possibilities as the control construction. This fact is captured because in our analysis, the dative ha -causative verb is essentially treated as a control verb. It can feed the lexical rule in (67), producing an entry which licenses flat structures such as (68a). 3.2.1.2. Raising-to-object construction. As discussed in section 3.1.3, the raising construction has more restricted scrambling possibilities than the

192

Chan Chung

control construction. In the former, an argument within the VP complement cannot scramble with arguments outside the VP, and the controller cannot occur between governed verb and the raising verb (e.g., (54d,e) and (55), respectively). The absence of scrambling out of the VP complement in this construction follows from the fact that a raising verb does not feed the lexical rule in (67). In the input entry of the lexical rule, the controller has its own semantic role. As shown in (53), however, the controller in the raising verb does not have its own semantic role (no role of BELIEVEE in (53)). Thus (53) is not a licit lexical entry for the input, and the lexical rule in (67) does not allow it to have the output counterpart. Under this assumption, the structure of the raising sentence in (54a) is analyzed only as in (73):

(73)

Here New York-ey cannot scramble out of the VP complement since it is confined within the VP. To account for the fact that the controller cannot occur after the complement VP in the raising construction, we assume the following LP constraint, which is suggested by Yoo (1993):

(74)

[SYNSEMH] < [SUBJ ]

The LP constraint in (74) states that a subject must precede the category which selects it as a subject. Like the LP constraint in (72), the LP constraint in (74) is another instantiation of the head-final property of Korean. In (54d,e) John-ul is the subject of the VP New York-ey isstako 'to stay in New York' but does not precede the VP. Thus, these sentences violate (74). It is important to note that the LP constraint in (74) does not apply in the case of the control construction. The constraint in (74) applies only when the whole SYNSEM value of the controller and the subject of the VP

Argument Composition, Long-Distance Scrambling

193

complement is the same. As shown in (68b), however, the controller and the understood subject of the verbal complement of the control verb are just coindexed, and the whole SYNSEM value cannot be the same because of the case mismatch. Hence, in the control construction, (74) does not apply, and the controller is allowed to linearly follow the verbal complement, Finally, the subject of the raising verb can occur between the complement VP and the raising verb, as shown in (54c). This is a consequence of the flat analysis shown in (73).17 3.2.1.3. Accusative ha-causative construction. As discussed in section 3.1.4, the scrambling possibilities of the accusative ha -causative construction are not the same as those of the control or raising construction. On the one hand, that variant of ha shares properties of a raising verb in that the controller cannot occur after the governed verb (e.g., (63)), but on the other hand, it also behaves like a control verb in that an argument within the VP complement can scramble with arguments outside the VP (e.g., (64)). On our approach, all the properties of the accusative ha -causative fall out naturally if we assume that this type of ha has the lexical description in (75), which is a revision of (60b):

(75)

This descrition resembles the output of the lexical rule in (67). In our account, however, there is no lexical entry corresponding to an input for (67), and therefore (75) must simply be listed separately. The lexical entry in (75) licenses the structure in (76b) for the sentence in (76a) (= (59b)): (76) a. Mary-ka John-ul ku chayk-ul ilkkey hayssta. M-Nom J-Acc the book-Ace read caused 'Mary caused John to read the book.'

194

Chan Chung

In (76b), John-ul and the subject of the verb ilkkey 'read' have the same SYNSEM value, represented as SNP. Thus, by the LP constraint in (74), John-ul must precede ilkkey. The sentence in (63) is unacceptable since it violates this LP constraint. Also in (76b), ku chayk-ul 'the book' must precede ilkkey because of the LP constraint in (72), which states that a complement must precede the category which subcategorizes for it. Our analysis correctly predicts that all permutations of three NPS and a complement verb are possible as long as these two LP constraints are observed.

3.2.2. ANALYSES OF S-COMPLEMENT CONSTRUCTIONS In this section, we will extend the idea of scrambling via argument composition further, proposing that scrambling out of an s-complement also can be accounted for by the same mechanism. In Korean, arguments of an embedded clause also can scramble with arguments of the main verb, as shown in (77) and (78):18, 19 (77) a. Nay-ka [s Mary-ka ku chayk-ul ilkesstako] sayngkakhayssta. I-Nom M-Nom the book-Ace read thought 'I thought Mary read the book.' b. Ku chayk-ul nay-ka Mary-ka ilkesstako sayngkakhayssta. the book-Ace I-Nom M-Nom read thought (78) a.

Mary-ka motun salam-hanthey [s John-i ku mwuncey-lul M-Nom all people-Dat J-Nomthe problem-Ace haykyelhaysstako} malhayssta. resolved told 'Mary told all the people that John resolved the problem.' b. (?)Ku mwuncey-lul, Mary-ka, motwun salam-hanthey, John-i the problem-Ace M-Nom all people-Dat J-Nom haykyelhaysstako malhayssta. resolved told

Argument Composition, Long-Distance Scrambling

195

c. ? Mary-ka, ku mwuncey-lul, motwun salam-hanthey, John-i M-Nom the problem-Ace all people-Dat J-Nom haykyelhaysstako malhayssta. resolved told In our analysis, arguments of the head of the embedded clause can be attracted to the COMPS lists of the matrix verb. To this end, we posit the following lexical rule which is similar to that in (67):

(79)

The input description in (79) takes a sentence as its complement. When the element represented by GD is the empty list, the input matches a lexical entry that takes only a sentence as its complement (e.g., sayngkakhayssta 'thought' in (77)). When H] is realized as {NP[dat]}, the input entry matches a lexical entry that takes both an Np[dat] and an s as its complement (e.g., malhayssta 'told' in (78)). In contrast, the output description of (79) takes a verb as its complement in addition to S, and the subject and complement of the complement verb (represented by NP[nom] 0 and GO) are attracted and appended to the COMPS lists of the matrix verb. See section 4.1 for motivation of the idea to let the embedded subject be attracted to the matrix verbs' COMPS list, rather than to the SUBJ list. Instances of scrambling as in (77b) and (78b,c) are licensed by output entries of the lexical rule. For example, (80a) is the structure of (77b)

196

Chan Chung

licensed by (80b):20 (80) a.

Note that the rule in (79) allows arbitrarily many arguments of the embedded clause to scramble out. This is in fact possible in Korean, as shown below:21 (81) a. Nay-ka [s Mary-ka ku chayk-ul John-hanthey pilyecwuesstako] I-Nom M-Nom the book-Ace J-Dat lent sayngkakhayssta. thought 'I thought that Mary lent the book to John.' b. Ku chayk-ul John-hanthey nay-ka Mary-ka pilyecwuesstako the book-Ace J-Dat I-Nom M-Nom lent sayngkakhayssta. thought c. John-hanthey ku chayk-ul nay-ka Mary-ka pilyecwuesstako J-Dat the book-Ace I-Nom M-Nom lent sayngkakhayssta. thought

3.23. ADJUNCT SCRAMBLING In previous sections, we focused only on scrambled arguments. In Korean, however, an adjunct that modifies a verbal expression can also scramble with arguments, as shown in (82b,c).

Argument Composition, Long-Distance Scrambling

197

(82) a. Ku kongwen-eyse Mary-ka John-ul mannassta. the park-at M-Nom J-Acc met 'Mary met John at the park.' b. Mary-ka ku kongwen-eyse John-ul mannassta. M-Nom the park-at J-Acc met c. Mary-ka John-ul ku kongwen-eyse mannassta. M-Nom J-Acc the park-at met In this section, we will discuss how to handle scrambled adjuncts. We have already assumed a flat structure for a sentence. Hence it is natural to posit that an adjunct is also licensed by the flat structure. In order to account for scrambling among complements and adjuncts in German, Kasper (1994) proposes to analyze adjuncts as sisters to complements, and hence to the head verb: i.e., an adjunct is contained in a flat structure. In this flat structure analysis, we need the following HeadSubject-Complement-Adjunct schema for Korean:

(83)

The schema in (83) states that zero or more adjuncts can appear in a clause and that if at least one adjunct is present, it is a sister to the arguments (subject and complements) and to the verb which subcategorizes for the arguments.22 A special property of (83) is that the value of an adjunct modifying a predicate is cospecified in the valence value of the predicate, i.e., as generally assumed, an adjunct still selects what it modifies, but the list of the modifiers is also specified as the value of the new valence feature ADJT on the modified predicate.23 The proposal to make adjuncts part of the valence specification is justified in that there seems to be no syntactic difference between arguments and adjuncts in Korean. Both can scramble within a clause, scramble out of an s- or vp-complement, be afterthought expressions in a simplex clause, and be extracted out of an s- or vp-complement to be afterhtought expressions. However, an adjunct differs from an argument in its semantics: i.e., an adjunct does not have any specific semantic role in a predicate which is modified by the adjunct, and it is a

198

Chan Chung

semantic head in that the CONTENT value of an adjunct is identical to that of the mother node. These facts will be discussed one by one shortly. In this section, we focus on the case where an adjunct scrambles out of a VP or s complement. As pointed out in Saito (1985), an adjunct generally can scramble out of a complement clause or a VP complement, as shown in (84) and (85):

(84) a.

Mary-ka [s John-i nayil kkaci ku swukcey-lul M-Nom J-Nom tomorrow by the homework-Ace machyeya-hantako] malhayssta. finish-must said/told 'Mary said that John had to finish the homework by tomorrow.' b. (?)Nayil kkaci Mary-ka John-i ku swukcey-lul tomorrow by M-Nom J-Nom the homework-Ace machyeya-hantako malhayssta. finish-must said/told

(85) a.

Mary-ka John-hanthey [VP nayil kkaci ku swukcey-lul M-Nom J-Dat tomorrow by the homework-Ace machilako] malhayssta. finish said/told 'Mary told John to finish the homework by tomorrow.' b. (?)Nayil kkaci Mary-ka John-hanthey tomorrow by M-Nom J-Dat ku swukcey-lul machilako malhayssta. the homework-Ace finish said/told c. (?)Mary-ka nayil kkaci John-hanthey M-Nom tomorrow by J-Dat ku swukcey-lul machilako malhayssta. the homework-Ace finish said/told

Here, (84b) and (85b,c) are slightly marginal compared with the (a) counterparts, but they are quite acceptable. On our approach, this fact is accounted for by the flat structure analyses, under the assumption that an adjunct of a complement verb can also be attracted to the ADJT list of a matrix verb when arguments are attracted. To this end, we need to revise the argument composition lexical rules in (67) and (79). For example, the

Argument Composition, Long-Distance Scrambling

199

rule in (67) is revised into (86):

(86)

(Here a = 1 or 3.) The rule in (86) differs from (67) only in that, in the former, the adjunct (ADJT) list of a complement verb is attracted to the ADJT list of a matrix verb as well. On this approach, (84b) has the structure in (87a):

(87)

a.

200

Chan Chung

In (87), the adjunct PP nayil kkaci 'by tomorrow' takes the complement QDv as its MOD value. Note that the schema in (83) is nonspecific as to whether the modified category is the head verb or a complement verb. The structure in (87) instantiates the latter case, where the modified category is a complement verb. In this case, if we directly apply the semantic principle of Pollard and Sag (1994) given in (88), we do not get the correct interpretation. (88) In a headed phrase, the CONTENT value is token-identical to that of the adjunct daughter if the DTRS value is of sort head-adjunctstruc, and to that of the head daughter otherwise. (Pollard and Sag, 1994:56) The daughters of the s category in (87) constitute the head-subj-compadjunct-structure which contains an adjunct daughter. Then, by (88), the semantic content of the whole sentence in (87) must be identical to the semantic content of the widest-scoping adjunct daughter, which can be roughly represented as follows: by-tomorrow' (necessarily' (finish' (Mary', the-homework'))). However, this is obviously not the correct interpretation of the whole sentence. This problem can be solved by a slight revision of (88) to (89): (89) In a headed phrase, the CONTENT value is token-identical to that of an adjunct daughter if the head daughter is the value of MOD of the adjunct daughter, and to that of the head daughter otherwise. In (87), the adjunct daughter does not take the head daughter as its MOD value, and hence the CONTENT of the whole sentence is the same as that of the head daughter, which is roughly represented as follows: said' (Mary', by-tomorrow' (necessarily' (finish' (Mary', the-homework')))). On our approach, the argument composition lexical rules play a crucial role in the determination of the correct meaning mentioned above. Here the important point is that the value of the CONTENT of the input entry is token-identical to that of the output entry: i.e., only the valence value of the input entry differs from that of the output value. (90) shows the structure of (84a) which is licensed by the lexical entry to which the argument composition lexical rule does not apply. Here the structures of [Tjv:[3] and 0V: GO are shown in (91) and (92), respectively.24

Argument Composition, Long-Distance Scrambling

(90)

(91)

(92)

201

202

Chan Chung

If we apply the argument composition lexical rule to (92), we obtain the following output entry:

The lexical description in (93) is the same as the head verb mv in (87b) and licenses the flat structure in (87a), whereas the one in (92) is the same as the matrix head verb EV in (90) and licenses the hierarchical structure. However, the CONTENT value of both is identical so that the structures in (87a) and (90) have exactly the same interpretation, which can roughly be represented by the CONTENT value QD in (92). A potential problem for our analysis is asymmetrical distribution of adjuncts in s- and vp-complement structures. The flat structure in (87) predicts that an adjunct which modifies the matrix head verb may occur at any place in a sentence. This prediction is borne out in the case of the vp-complement construction. The sentences in (94) show that the PP ilcwuil ceney 'one week ago' which modifies the matrix verb malhayssta 'told' can be placed rather freely: (94) a.

Mary-ka ilcwuil ceney John-hanthey ku project-lul M-Nom one week ago J-Dat the project-Ace kkuthnaylako malhayssta. finish said/told 'One week ago Mary told John to finish the project.' b. Mary-ka John-hanthey ilcwuil ceney ku project-lul M-Nom J-Dat one week ago the project-Ace kkuthnaylako malhayssta. finish said/told c. (?)Mary-ka John-hanthey ku project-lul ilcwuil ceney M-Nom J-Dat the project-Ace one week ago kkuthnaylako malhayssta. finish said/told d. Mary-ka John-hanthey ku project-lul kkuthnaylako M-Nom J-Dat the project-Ace finish ilcwuil ceney malhayssta. one week ago said/told

Argument Composition, Long-Distance Scrambling

203

Even though (94c) is slightly marginal compared with the others, it is quite acceptable when a short pause is given after the adjunct, or stress is given to the adjunct.25 As shown in (95), however, in the s-complement construction, an adjunct modifying a matrix verb never occurs between the embedded subject and its predicate verb: (95) a. Mary-ka ilcwuil ceney John-i ku project-lul M-Nom one week ago J-Nom the project-Ace kkuthnayya-hantako malhayssta. finish-must said/told 'One week ago Mary said that John had to finish the project.' b. *Mary-ka John-i ilcwuil ceney ku project-lul M-Nom J-Nom one week ago the project-Ace kkuthnayya-hantako malhayssta. finish-must said/told c. *Mary-ka John-i ku project-lul ilcwuil ceney M-Nom J-Nom the project-Ace one week ago kkuthnayya-hantako malhayssta. finish-must said/told d. Mary-ka John-i ku project-lul kkuthnayya-hantako M-Nom J-Non the project-Ace finish-must ilcwuil ceney malhayssta. one week ago said/told The sentences in (95b,c) cannot be ameliorated by any pause or stress and present a sharp contrast with (94c). Therefore, the prediction seems not to be borne out that an adjunct which modifies the matrix head verb may occur at any place in a sentence. In our analysis, however, sentences such as (95b,c) can be ruled out by the interpretive principle mentioned in note 21. The principle is repeated here as (96): (96) Interpretive Principle Suppose (i) that Y is an NP[NOM], (ii) that X is the first verb following Y, and (iii) that Z is any constituent which occurs between Y and X. Then Z cannot be a "semantic dependent (semantic argument or functor)" of a verb superordinate to X. For example, the structure of semantic dependency for (95b) is as follows:

(97)

204

Chan Chung

Thus, (95b,c) are ruled out by (96) and do not present counterexamples to the flat structure analysis. Now, let us consider some matters that are relevant to LP constraints on adjuncts. First, an adjunct cannot occur after a category that is modified by the adjunct. The relevant examples are in (98): (98)

*Mary-ka [s John-i ku swukcey-lul M-Nom J-Nom the homework-Ace machyeya-hantako nayil kkaci] malhayssta. finish-must tomorrow by said/told 'Mary said that John had to finish the homework by tomorrow.'

To account for this fact, we need to assume another LP constraint in (99) which is another instance of head-finality:26 (99) SYNSEMIH < [ADJT < . . . CD ... >] The constraint in (99) simply states that an adjunct precedes a category which is modified by the adjunct. Another matter that deserves discussion is scrambling of an adverb phrase (ADVP) which is derived from a stative verb ("true adjunct" in terms of Saito, 1985).27 This kind of adjunct scrambles out of the complement clause or VP only with great difficulty, if at all, as shown in (100): (100) a.

Mary-nun [s John-i yelsimhi enehak-ul M-Top J-Nom hard linguistics-Ace kongpwuhaysstako] sayngkakhanta. studied think 'Mary thinks John studied linguistics hard.' b. ??/*Yelsimhi Mary-nun John-i enehak-ul kongpwuhaysstako hard M-Top J-Nom linguistics-Ace studied sayngkakhanta. think

(101) a.

Mary-ka John-hanthey [VP yelsimhi enehak-ul M-Nom J-Dat hard linguistics-Ace kongpwuhalako] malhayssta. study said/told 'Mary told John to study linguistics hard.' b. llYelsimhi Mary-ka John-hanthey enehak-ul kongpwuhalako hard M-Nom J-Dat linguistics-Ace study malhayssta. said/told

Argument Composition, Long-Distance Scrambling

205

However, considering the examples in (102), we cannot simply assume that an ADVP does not scramble out of an s or VP at all: (102) a. (?)Enehak-ul yelsimhi Mary-nun John-i kongpwuhaysstako linguistics-Ace hard M-Top J-Nom studied sayngkakhanta. think 'Mary thinks John studied linguistics hard.' b. Enehak-ul yelsimhi Mary-ka John-hanthey kongpwuhalako linguistics-Ace hard M-Nom J-to study malhayssta. said/told 'Mary told John to study linguistics hard.' In (102), the ADVP scrambles out of the s- or vp-complement together with the object NP of the embedded verb. Such examples are notably more acceptable than (l00b) or (l0lb). Thus, in our analysis, it seems that the unacceptable (b) sentences in (100) and (101) are licensed by syntax but ruled out by certain processing factors whose exact nature is presently unknown to us. That is, ADVPS are considered the same as other kinds of adjuncts and there is no specific syntactic constraint on ADVP scrambling. In this section, we showed how the account of verbal complexes can be extended to the account of so-called long-distance scrambling. The crucial mechanisms in our analysis of long-distance scrambling are argument composition and lexical rules. Kiss (1994) extends the argument attraction mechanism in Hinrichs and Nakazawa (1989, 1994) to account for German modal, control, and raising verb constructions. The suggestions in this section share some of the same basic ideas. However, our proposals differ from his in several respects. The main differences are as follows: (i) arguments of the complement verb which are attracted to the nonauxiliary matrix verb are all sisters to the matrix and complement verbs (cf. Bouma and van Noord, 1998, for the same suggestion on the Germanic verb clusters); (ii) different scrambling possibilities among different vp-complement constructions are accounted for by the applicability of an argument composition lexical rule, i.e., whether the matrix verb has only the input entry of the lexical rule or only the output entry, or both of them; (iii) scrambling out of a complement clause is also accounted for via argument compostion; and (iv) the argument composition analysis extends to scrambling of adjuncts. In the next section, we discuss what theoretical predictions can be made with our approach to long-distance scrambling.

206

Chan Chung

4. THEORETICAL PREDICTIONS The goal of this section is to show what theoretical predictions our theory can make. To this end, we discuss some phenomena that have been raised in previous studies of scrambling: i.e., long-distance passivization and weak crossover effects. 4.1. Long-Distance Passivization Korean exhibits so-called long-distance passivization phenomena, where an object of an embedded verb is passivized and appears as the matrix subject, while only the matrix verb bears passive morphology. Examples are given in (103) and (104) (see Kiss, 1992, Nagai, 1991, and Yatabe, 1993, for discussions on long-distance passivization in German and Japanese): (103) a. Nay-ka Mary-hanthey ku cengchayk-ul sihaynghalako cisihayssta. I-Nom M-Dat the policy-Ace carry out ordered 'I ordered Mary to carry out the policy.' b. Ku cengchayk-i naey-uyhayse Mary-hanthey sihaynghalako the policy-Nom I-by M-Dat carry out cisi-toy-essta. order-Pass-Past (Literally) 'The policy was ordered by me for Mary to carry out.' ('It was ordered by me for Mary to carry out the policy.') (104) a. Motun salam-i hyun cengpwu-ka ku cengchayk-ul all people-NoM current government-Nom the policy-Ace sihaynghalilako sayngkakhayssta. will carry out thought 'All people thought the government would carry out the policy.' b. Ku cengchayk-i motun salamey-uyhayse hyun cengpwu-ka the policy-Nom all people-by current government-Nom sihaynghalilako sayngkak-toy-essta. will carry out think-Pass-Past (Literally) 'The policy was believed by all people that the government would carry out.' ('It was believed by all people that the government would carry out the policy.') In (103b), the object NP ku cengchayk-ul 'the policy' within the VP complement is passivized, while the passive morpheme -toy is realized on the matrix verb, e.g., cisi-toy-essta 'was ordered.' The same long-distance passive can also affect arguments of the s-complement as shown in (104b). This kind of passivization is problematic for Phrase Structure Grammar in general, because under standard assumptions an object can be passivized only when it is an argument of a verb on which the passive morpheme is realized.

Argument Composition, Long-Distance Scrambling

207

Note that the subject of the embedded clause also can be the subject of the long-distance passive, as shown in (105): (105) Hyun cengpwu-ka motun salam-eyuyhayse current government-NoM all people-by ku cengchayk-ul sihaynghalilako sayngkak-toy-essta. the policy-Ace will carry out think-Pass-Past (Literally) The current government was thought by all people to carry out the policy.' Long-distance passivization phenomena are naturally accounted for by our theory of long-distance scrambling. According to that proposal, the object NP ku cengchayk-ul 'the policy' of the embedded verb in (103b) and (104b) can be attracted to the COMPS list of the matrix verb. Then the attracted NP can be promoted to subject by the passive lexical rule in (106), while the passivization morpheme is realized on the matrix verb: (106)

The passive lexical rule in (106) takes as input a base form verb (e.g., cisiha 'order' in (103a)) and gives as output a passive form verb (e.g., cisitoy 'be ordered' in (103b)). The subject of the input is removed from the SUBJ list, and its index is reassigned to the PP complement of the output. In addition, the nondative (i.e., nominative or structural) NP complement of the input is placed in the SUBJ list of the output entry. For instance, the structure of (103a) can be analyzed as a flat structure as in (107): (107)

208

Chan Chung

The application of the passive lexical rule in (106) to the head verb, [5]v in (107), yields (108) as output: (108)

fsuBJ I COMPS (pp[dat]fl,

1 HV, pp[uyhayse]m} I

Then the lexical entry in (108) licenses the passivized sentence in (103b) as follows:

(109)

In our theory, examples such as (105) can also be accounted for in terms of the passive rule in (106). The subject of the embedded verb is attracted to the COMPS list of the matrix verb: i.e., the embedded subject NP is a complement NP in the matrix verb in the output entry of the argument composition lexical rule, and thus the matrix verb feeds the passive rule. For example, (110) is the lexical entry for the matrix verb sayngkakha 'think,' which corresponds to output of the argument composition lexical rule:

(110)

The description in (110) feeds the passive rule in (106), and its output car be either (Ilia) or (lllb), since either Np[nom]^ or NPt^r]^ can be th< subject of the passive verb:

(111) a.

Argument Composition, Long-Distance Scrambling

209

The description in (Ilia) licenses the sentence in (104b), while (lllb) licenses the sentence in (105). 4.2. Amelioration of Weak Crossover Effects Our unified account of long-distance scrambling and clause-internal scrambling in terms of the flat analysis also predicts some facts about the weak crossover (WCO) effect. In Korean, clause-internal scrambling ameliorates WCO effects as shown in (112): (112) a. *[Kui-uy/proi emma-ka] nwukwUj-lul salangha-ni! he-Gen/pro mother-Nom who-Acc love-Q 'WhOj does his; mother love?' b. NwukwUf-ul [ku^uy/^prOf emma-ka] salangha-nil who-Acc he-Gen/pro mother-Nom love-Q In (112a), the w/z-operator nwukwu-lul 'who' neither o-commands nor precedes the NP dominating a pronoun coindexed with the operator, which induces WCO effects.28 In contrast, in (112b), the operator precedes the NP dominating a pronoun, resulting in improved acceptability. Thus scrambling in Korean can ameliorate WCO effects. In Korean, the amelioration of the WCO effect also occurs with longdistance scrambling (Cho, 1994), as shown in (113) and (114):29 (113) a. ^KUf-uy/prOf emma-ka] [s Mary-ka nwukwurlul he-Gen/pro mother-Nom M-Nom who-Acc ttaylyesstako] sayngkakha-ni! hit think-Q (Literally) 'Whoi did hiSj mother think Mary hit?' b. Nwukwiij-ul [kUj-uy/^prot emma-ka] Mary-ka who-Acc he-Gen/pro mother-Nom M-Nom ttaylyesstako sayngkakha-ni! hit think-Q (114) a. *[KUj-uy/proi emma-ka] [s Mary-ka nwukwunkttj-lul he-Gen/pro mother-Nom M-Nom someone-Acc ttaylyesstako] sayngkakhanta. hit think-Q (Literally) 'HiSj mother thinks Mary hit someonej.'

210

Chan Chung

b.

Nwukwunka^-lul [kui-uy/(^proi emma-ka] Mary-ka someone-Acc he-Gen/pro mother-Nom M-Nom ttaylyesstako sayngkakhanta. hit think

In the (a) sentences, nwukwu-lul 'who' and nwukwunka-lul 'someone' neither c-command nor precede the NP dominating the pronoun, which induces a WCO effect. In contrast, in the (b) sentences, the operator precedes the NP dominating the pronoun, and thus the WCO effect is ameliorated. This observation shows that, in Korean, the WCO effect is ameliorated by scrambling regardless of whether the type of scrambling involved is clause-internal or clause-external.30 This kind of data shows that long-distance scrambling in Korean cannot be treated the same as long-distance scrambling in Hindi, because long-distance scrambling in that language does not ameliorate the WCO effect (e.g., (115b)), while clause-internal scrambling does (e.g., (115a)) (Mahajan, 1990).31 (115) a. kis-kOi uskiit bahin pyaar kartii hEt who his sister loves 'WhOj, his; sister loves tj?' b. *kis-koi uskiii bahin-ne socaa [ki raam-ne ti dekhaa thaa]. who his sister thought that Ram seen be-past 'WhOj, his; sister thought that Ram had seen tj?' If the behavior with respect to WCO effects is considered to be diagnostic of different kinds of dislocation, then the above evidence suggests that long-distance scrambling and clause-internal scrambling in Korean are not qualitatively distinct phenomena. Then there is no reason to assume another mechanism such as A-bar movement in GB or SLASH feature percolation in GPSG/HPSG for long-distance scrambling. In our theory, this lack of distinction between long-distance and clause-internal scrambling is captured since we analyze both as licensed in terms of argument composition, resulting in flat constituent structures. In this section, we discussed what theoretical predictions are made by our theory of long-distance scrambling. The phenomena discussed above suggest that, in Korean, clause-internal scrambling and long-distance scrambling are syntactically indistinguishable. Our theory of long-distance scrambling can naturally account for this fact because both types of scrambling are licensed through flat structures. This unified account is possible due to argument composition. The examples of long-distance passivization in section 4.1 suggest not only that both types of scrambling have the same properties, but also that the arguments of the embedded verb are attracted (or raised) to the COMPS list of the matrix verb.

Argument Composition, Long-Distance Scrambling

211

5. SUMMARY AND CONCLUSIONS The main proposals of this paper are summarized as follows: (i) clauseinternal scrambling and long-distance scrambling are syntactically indistinguishable, and thus they must be licensed in a uniform way; (ii) both types of scrambling result from the relative lack of linear precedence constraints among the nonhead constituents at flat clausal structures; (iii) the crucial mechanism involved here is argument composition, which is utilized to account for the verbal complex system; (iv) the coexistence of verbal complexes and long-distance scrambling in languages like Korean is accounted for in a principled way because both are licensed by the same mechanism. Our analysis is similar in spirit to the A-movement approach in GB (Yoshimura, 1989), since argument composition can be regarded as an extension of raising. However, the A-movement approach seems to be problematic under the GB framework. Under the standard assumptions, A-movement (e.g., raising and passivization in English) is triggered by certain syntactic factors (Bayer and Kornfilt, 1991). The problem is that there do not seem to be any syntactic factors triggering scrambling in Korean. Based on facts about the verbal noun construction, however, Lee (1990) proposes that case alternation and word order variation are derived from the interaction between (i) the scope of an aspect morpheme and (ii) different licensing conditions for nominative, accusative, and genitive case. (See Miyagawa, 1991, for similar claims in Japanese.) That study suggests that scrambling is allowed as long as a constituent is within the scope of an appropriate Case assignor, and thus that scrambling is a Case-driven phenomenon. However, it is still an open issue how long-distance scrambling and its effects on the WCO effect and passivization can be handled in the Case-driven theory. Yatabe (1993) proposes an analysis of Japanese long-distance scrambling which is similar to our theory. That is, scrambling out of an embedded s or VP is accounted for by either "pseudo-raising" (A-movement in terms of GB) which ameliorates the WCO effect or extraposition (A-bar movement in terms of GB) which does not ameliorate the WCO effect. In the analysis of Korean, however, there is no motivation for the distinction because the amelioration of the WCO effect occurs in the Korean counterpart of Yatabe's extraposition example, as shown in (116): (116) a. *Kui-uy/proi emma-ka [s John-i etten chinkwUj-hanthey he-Gen/pro mother-Nom J-Nom which friend-Dat yok-ul haysstako] sayngkakhayss-ni*! curse-Ace did thought-Q (Literally) 'Which friendj did hisj mother think John cursed?'

212

Chan Chung

b. Etten chinkwu-hanthey kui-uy/^prolemma-ka John-i which friend-Dat he-Gen/pro mother-Norn J-Nom yok-ul haysstako] sayngkakhayss-nil curse-Ace did thought-Q Note that in Yatabe's theory, pseudo-raising allows only the NP with accusative case to be raised and thus the examples in (116) cannot be instances of pseudo-raising. However, the WCO effect is ameliorated even in this case in Korean, which suggests that the distinction between pseudo-raising and extraposition cannot be motivated by this kind of data in that language. Finally, let us compare our argument composition analysis with the verb projection raising (VPR) analysis proposed by Haegeman and van Riemsdijk (1986). They propose the reanalysis rule in (117) to account for verbal complexes in the Germanic languages: (117)

Reanalysis If the representation of a sentence contains the line XV q V r Y, where 0 < / < 2 and Vr is a VR (verb raising) verb, then add the line XV X Y to that representation.

Here, the superscript i in Vq allows the verbal category Vq to be instantiated as V°, V, or VP, depending on the language. The rule is intended for stating that when there is a sequence of verbs, Vq and Vr, and when Vr is a VR verb, Vq and Vr are reanalyzed as Vx, as in (118):

(118)

In structure (118), where Vr is the head, the internal 0-role of Vq is assumed to percolate up to Vx, while both the internal and external 0-roles of Vr are assumed to percolate up to Vx. Then the percolated internal 0-role of Vq is assumed to be "associated with" that of Vr. Here, the notion of "associated with" is not clearly defined, but we can regard it as having the same effect as argument composition or structure-sharing. In our account, (i) the COMPS list of Vq is attracted to the COMPS list of Vr by the lexical structure of Vr, and (ii) the SUBJ list and COMPS list with the attracted elements percolate up by the Valence Principle. We may say that argument composition and the Valence Principle correspond to the internal 0-role association and 0-role percolation, respectively. Despite the apparent similarity between the VPR and our analyses, there is at least one crucial difference, which bears on the applicability range of argument composition and 0-role association. In Haegeman and

Argument Composition, Long-Distance Scrambling

213

van Riemsdijk's (1986) framework, the exhaustive association of internal 0-roles between the governed verb and the governing verb occurs only when two verbs are reanalyzed as a verbal complex. As discussed in sections 3 and 4, however, such internal 0-role association also seems to be required to account for long-distance scrambling in Korean where the governed verb and the governing verb do not form a verbal complex. That is, the notion of 0-role association cannot be extended or generalized into the account of this phenomenon because syntactic material can intervene between the governed and governing verbs in the long-distance scrambling case (e.g., (51), (69c), and (70c)). In contrast, we can use the same argument composition mechanism to account for the same phenomenon by simply changing the selection mechanism, i.e., by lexically specifying that the governed verb is selected by the governing verb's COMPS list, rather than by the GOV list of an auxiliary verb. The flexibility of our system is due to the fact that argument composition is based not on configuration but on structure sharing within a lexical entry.

ACKNOWLEDGMENTS Many parts of this work are based on my dissertation at Ohio State University. I thank Carl Pollard, Peter Culicover, Robert Levine, Robert Kasper, Craige Roberts, and Andreas Kathol for their valuable comments and suggestions. This work has also benefited from the comments of two anonymous reviewers and from discussions with the audience at the HPSG workshop held at the University of Tubingen in 1995. All remaining errors are, of course, mine.

NOTES :

The distinction between passivizable and nonpassivizable auxiliaries seems to result from semantic constraints on cz-passivization. According to Li and Thompson (1975), the passive construction in topic-prominent languages such as Japanese and Korean carries a special meaning (e.g., the "adversity" passive in Japanese). The cz-passive in Korean also seems to have a semantic constraint: i.e., the verb or verbal cluster to which the passive ci attaches must not be a verb denoting a certain state, in terms of Dowty (1979). For example, the following sentences show sharp constrasts when compared with the sentences in (7b) and (8b): (i) a. *Ku namwu-ka Mary-eyuyhayse caluko isse

the tree-Norn M-by

cut

ci-essta.

be in the process of passive-past

214

Chan Chung

'The tree was in the process of being cut.' b. *Ku namwu-ka na-eyuyhayse caluko siphe ci-essta. the tree-Norn I-by cut want passive-past 'The tree was wanted to be cut by me.' c. *Ku namwu-ka Mary-eyuyhayse calukey toye ci-essta. the tree-Nom M-by cut be led to passive-past 'The tree was led to be cut by Mary.' Here the verbal clusters to which the passive ci attaches represent states, while those in (7) and (8) represent activities. 2 We may assume that the passive morpheme ci subcategorizes for the TVP (VP without the primary object) rather than the whole VP or v, and that the primary object is directly subcategorized for by ci. An example of the analysis is illustrated in (i): (i) [s Malssengmanhun so-ka [vp (ku nongpwu-eyehay) troublesome cow-Nom the farmer-by [VP [-rvp John-hanthey phala chiwe] ci-essta]]] J-Dat sell resolutely passive-past 'The troublesome cow was resolutely sold to John (by the farmer).' A problem with this analysis is that it is hard to find any syntactic evidence for the constituency of the TVP. As discussed in section 3.1.1, any phrasal category can occur after the head verb as an afterthought expression. However, the TVP in (i) cannot be an afterthought expression as shown in (ii): (ii) *Malssengmanhun so-ka ku nongpwu-eyehay ci-essta, troublesome cow-Nom the farmer-by passive-past [TVP John-hanthey phala chiwe]. J-Dat sell resolutely Also note that no parenthetical can intervene between ci and its governed verb, as shown in (iii), again exhibiting a behavior different from that of other VP complement constructions (cf. (3ciii) and (4)): (iii)

*Malssengmanhun troublesome phala chiwe] sell resolutely

so-ka ku nongpwu-eyuyhay [yyp John-hanthey cow-Nom the farmer-by J-Dat hayekan ci-essta. anyway passive-past

3 Pollard and Sag (1987) define the Locality Principle as follows: the SUBCAT elements of lexical sign (word in Pollard and Sag, 1994) specifies values for SYNTAX and SEMANTICS (SYNSEM in Pollard and Sag, 1994) but crucially not the attribute DAUGHTERS. However, Pollard and Sag (1994) assume that elements of SUBJ or COMPS lists are not of sort sign but sort synsem, which does not include the attribute DAUGHTERS. In this framework, the Locality Principle is incorporated into the definition of the features, and thus it does not need to be stated separately. 4 This terminology is taken from No (1991). 5 Hinrichs and Nakazawa (1989, 1994) use the version of HPSG of Pollard and Sag (1987) in which the valence is not divided into subject and complement lists, but instead encoded in a single list (SUBCAT).

Argument Composition, Long-Distance Scrambling

215

6 This is essentially the same as the CONS function in LISP which takes an element and a list to return a list: element + list => list. 7 An exception is the passive auxiliary ci, which will be discussed shortly. 8 Rule (22) excludes the possibility that an auxiliary first combines with the subject and/or complements, not with the governee verb, and licenses an ill-formed sentence in (ia): that is, in the structure in (ib), rule (22) cannot apply to the top node to discharge the GOV value because the sort of the head is not a simple word but a phrase:

(i) a. *Mekko Mary-ka sakwa-lul siphessta. eat M-Nom apple-Ace wanted 'Mary wanted to eat an apple.' b

' *S[SUBJ< > , C O M P S < >,GOV]

9

A similar approach is taken in Yoo (1993:196). Yoo adopts the GOV feature in Chung (1993) for lexical entries of emotion auxiliaries. The lexical entries proposed by her have a minor technical problem as they are stated, since they lead to saying that an emotion verb takes only a transitive verb as its governee. 10 We admit that the assumption of two separate lexical entries for siph weakens the argument. Presently, we have no good explanation why siph causes this case alternation. 11 In Chung (1995), the afterthought expressions are treated differently from scrambling. See Chung (1995: Chap. 4) for detailed discussion on this matter. 12 Some more object-control verbs in Korean are kangyoha 'compel,' malha 'tell,' canglyeha 'encourage,' caychokha 'urge,' pwuthakha 'ask,' helakha 'allow,' and sikhi 'make.' Subject-control verbs are mayngseyha 'vow,' sitoha 'try,' and tonguyha 'agree.' The control verb construction in Korean discussed above is similar to the so-called "optionally coherent" construction in German that has been discussed since Bech (1955). 14 In Korean, the raising-to-subject construction is an instance of the verbal complex construction, which is headed by auxiliaries such as po 'seem,' tusha 'seem, look like,' etc. In this construction, NP arguments can scramble with each other (e.g., (ia,b)). However, the governed verb and the auxiliary cannot be separated by inserting syntactic material (e.g., (ic,d)) or by making the governed verb an afterthought expression (e.g., (ie)). As discussed in section 2, these are the typical properties of verbal complexes. (i) a. Mary-ka ku chayk-ul ilk-ess-na pota. M-Nom the book-Ace read-Past-vform.na seem 'Mary seemed to read the book.'

216

Chan Chung

b. Ku chayk-ul Mary-ka ilk-ess-na pota. the book-Ace M-Nom read-past-vform.na seem c. *Ku chayk-ul ilk-ess-na Mary-ka pota. the book-Ace read-past-vform.na M-Nom seem d. *Mary-ka ku chayk-ul ilk-ess-na hayekanpota. M-Nom the book-Ace read-Past-vform.na anyway seem 'Anyway Mary seemed to read the book.' e. *Mary-ka pota ku chayk-ul ilk-ess-na. M-Nom seem the book-Ace read-Past-vform.na On our approach, the valence value of the raising-to-subject auxiliary po is the same as (23). See Sells (1990) for a similar analysis. 15 In Korean, nominative case can generally be assigned to the subject of a nonfinite clause, as shown in (i): (i) John-un [s [atul-i tolawa-se] kippessta. J-Top son-Norn return (base-form)-because was happy 'John was happy because his son returned.' 16

The lexical rules for long-distance scrambling in (67) may be considered lexical versions of the liberation metarule in Pullum (1982) or the liberation principle in Zwicky (1986). 17 Even though our theory can capture the different scrambling possibilities between control and raising verb constructions by restricting the application of the composition rule to the control construction, it does not explain why such a difference exists. If the difference is caused by an independent reason, we may generalize the lexical rule so that it can apply to all lexical entries which take any kind of verbal expression as a complement. For example, an alternative analysis for the raising-to-object construction in Korean may be possible if we consider Hong's (1990) analysis. According to that proposal, the accusative controller is an embedded topic. This analysis naturally eliminates the unacceptable sentences in (54d,e) and (55), i.e., the embedded topic should precede all of its clause-mate elements. However, we still need the LP constraint in (74) to account for the accusative ha -causative construction, where the causee NP cannot be assumed to be an embedded topic since it allows the complement of the embedded verb to precede the causee (e.g., (64)). A more detailed analysis of this construction is provided in the next section. 18 Another way to analyze (77b) and (78b,c) may be as instances of English-style topicalization (Bratt, 1993). This view is suspicious because no obvious evidence exists which shows that extraction out of an embedded clause differs from the other extractions. See section 4 for detailed discussion on this. 19 We have no explanation at this point for the marginal status of (78b,c). The acceptability improves with short pauses after the NP arguments, which is indicated by commas. 20 In (80b), HP'S case is not specified when assigned case is structural) just for expository convenience. 21 Case marking on arguments can affect scrambling possibilities: i.e., an argument of an embedded verb cannot scramble with an argument of a matrix verb

Argument Composition, Long-Distance Scrambling

217

when they bear the same case marker, and a nominative argument of an embedded verb cannot scramble with an argument of a matrix verb. To account for the former case, Chung (1995) assumes an LP constraint into which the notion of obliqueness is incorporated: coarguments with the same case must be linearized in order of obliqueness: i.e., the less oblique one must precede the more oblique one. To account for the latter case, Chung (1995) assumes an interpretive principle which roughly states the following: Suppose (i) that Y is an Np[nom], (ii) that X is the first verb following Y, and (iii) that Z is any constituent which occurs between Y and X. Then Z cannot be a "semantic dependent (semantic argument or functor)" of a verb superordinate to X. Here semantic dependency involves structure sharing between the functor and its argument in their CONTENT values. The interpretive principle is not affected by argument composition because argument composition changes only the structure of the VALENCE value. See Chung (1995: Chap. 4) for more detailed discussion on this matter, and also see section 3.2.3 in this paper for another example where this principle is utilized. 22 A potential problem for (83) may be the semantic scope relations among more than one adjunct. In Kasper (1994), the scope of multiple adjuncts is captured in the flat structure analysis (i) by the order of adjunct daughters in the adjunct daughter list, which is ordered not by the surface order of the adjuncts but by their semantic scope, and (ii) by a split of the MOD value into SYN and SEM, reflecting the dual nature of adjuncts as semantic heads (functors), but syntactic nonheads. While a more sophisticated version of (83) will have to take these complications into account, the present version will suffice for our purposes. 23 The idea of cospecification of the modifier and modified can be considered as a hybrid of Pollard and Sag (1987) and Pollard and Sag (1994). In the former, a modifier is selected by the modified category, while in the latter, the modified category is selected by its modifier. 24 Here, in order to simplify the CONTENT value, we do not provide the semantics of the modal auxiliary verb, hanta 'must,' which is not directly relevant to current discussion. Also we simplify the CONTENT value of the adjunct nayil kkaci 'by tomorrow' by using the notation of intensional logic: by-tomorrow' (3). 25 Presently, it is not clear why (94c) is awkward. Our conjecture is that the awkwardness may be due to some processing factor. In Korean, when an adjunct immediately precedes a category which is modifiable by the adjunct, speakers tend to interpret the adjunct as modifying the immediately following category. In (94c), the adjunct ilcwuil ceney 'one week ago' immediately precedes the verb kkuthnaylako 'finish,' which is a category modifiable by the PP. Thus, the PP tends to be interpreted as modifying kkuthnaylako rather than malhayssta 'told.' A pause or stress may block this kind of processing interference. This conjecture is supported by (i): (i) Mary-ka John-hanthey ku project-lul ilcwuil ceney kuphi M-Nom J-Dat the project-Ace one week ago in haste kkuthnaylako malhayssta. finish said/told 'One week ago Mary told John to finish the project in haste.'

218

Chan Chung

Here, the adjunct ilcwuil ceney 'one week ago' and the embedded verb kkuthnaylako 'finish' are separated by another adjunct which modifies the embedded verb; consequently, (i) is much better than (94c). 26 The LP constraints encoding head-finality (e.g., (72), (74), and (99)) can be schematized as in (i), where F ranges over all valence features: (i)

[SYNSEM[1] < [VAL|F < . . . , m , . . . >] 27

These adverbs in Korean are usually derived by suffixing -(h)i or -key to a root of a stative verb. 28 See Chung (1995) for more detailed discussion in WCO effects in Korean and Pollard and Sag (1994) for the definition of o-command. 29 Saito (1992) and Yoshimura (1989) make the same observation about Japanese. 30 See (116) in section 5 for the case where an indirect object affects the WCO effect. 31 In our theory, long-distance scrambling in Hindi may be treated as a "genuine" unbounded dependency construction, on a par with English-style topicalization.

REFERENCES Bayer, J., and J. Kornfilt (1991). Against scrambling as move-alpha. NELS 21, 1-15. Bech, G. (1955). Studien fiber das deutsche Verbum Infinitum, Vol. 1 KDVS35, 2. Copenhagen. Borsley, R. (1989). An HPSG approach to Welsh. Journal of Linguistics 25, 333-354. Bouma, G., and G. van Noord (1998). Word order constraints on verb clusters in German and Dutch. In E. Hinrichs, A. Kathol, and T. Nakazawa (eds.), Complex. Predicates in Nonderivational Syntax. Syntax and Semantics, Vol. 30, 43-72. Academic Press, San Diego. Bratt, E. (1993). Clause structure and case marking in Korean causatives. In S. Kuno et al. (eds.), Harvard Studies in Korean Linguistics V, 241-251. Hanshin, Seoul. Cho, J-H. (1994). On scrambling: Reconstruction, crossover, and anaphor binding. In Y-K. Kim-Renaud (ed.), Theoretical Issues in Korean Linguistics, 255-274. CSLI Publications, Stanford, CA. Cho, J. O. (1988). Suffixed verb forms and compound verb constructions. In E. Back (ed.), Papers from the Sixth International Conference on Korean Linguistics, 77-106. Hanshin, Seoul. Chung, C. (1993). Korean auxiliary verb constructions without VP nodes. In S. Kuno et al. (eds.), Harvard Studies in Korean Linguistics V, 274-286. Hanshin, Seoul. Chung, C. (1995). A Lexical Approach to Word Order Variation in Korean. Ph.D. dissertation, Ohio State University.

Argument Composition, Long-Distance Scrambling

219

Dowty, D. (1979). Word Meaning and Montague Grammar. Reidel, Dordrecht. Dowty, D. (1990). Toward a minimalist theory of syntactic structure. In Tilburg Workshop on Discontinuous Constituency, 34-72. Gerdemann, D. (1994). Complement inheritance as subcategorization inheritance. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 341-363. CSLI Publications, Stanford, CA. Haegeman, L., and H. van Riemsdijk (1986). Verb projection raising, scope, and the typology of rules affecting verbs. Linguistic Inquiry 17, 417-466. Hale, K. (1982). Preliminary remarks on configurationality. NELS 12. (J. Pustejovsky and P. Sells (eds.).) Heinz, W., and J. Matiasek (1994). Argument structure and case assignment in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in HeadDriven Phrase Structure Grammar, 199-236. CSLI Publications, Stanford, CA. Hinrichs, E., and T. Nakazawa (1989). Flipped out: AUX in German. Chicago Linguistic Society 25. Hinrichs, E., and T. Nakazawa (1994). Linearizing AUXs in German verbal complexes. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 11-37. CSLI Publications, Stanford, CA. Hong, K-S. (1990). Subject-to-object raising in Korean. In K. Dziwirek, P. Parrel, and E. Mejias-Bikandi (eds.), Grammatical Relations: A Cross-Theoretical Perspective, 215-226. CSLI Publications, Stanford, CA. Kang, B. M. (1988). Functional Inheritance, Anaphora, and Semantic Interpretation in a Generalized Categorial Grammar. Ph.D. dissertation, Brown University. Kasper, R. (1994). Adjuncts in the Mittelfeld. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 39-69. CSLI Publications, Stanford, CA. Kathol, A. (1994). Passive without lexical rules. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 237-272. CSLI Publications, Stanford, CA. Kiss, T. (1992). Infinite Komplementation. Neue Studien zum deutschen Verbum Infinitum. Ph.D. dissertation, Bergische Universitat-Gesamthochschule Wuppertal. Kiss, T. (1994). Obligatory coherence: The structure of German modal verb constructions. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 71-107. CSLI Publications, Stanford, CA. Kuh, H. (1990). Correlation between Inflection and Word Order. Ph.D. dissertation, Ohio State University. Lee, Y-S. (1990). Case alternation and word order variation in nominal clauses. In H. Hoji (ed.), Japanese /Korean Linguistics. CSLI Publications, Stanford, CA. Li, C., and S. Thompson (1975). Subject and topic: A new typology of language. In C. Li (ed.), Subject and Topic, 457-489. Academic Press, New York. Mahajan, A. (1990). The A/A-Bar Distinction and Movement Theory. Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge, MA.

220

Chan Chung

Miyagawa, S. (1991). Case realization and scrambling. Unpublished manuscript, Ohio State University. Nagai, N. (1991). Complex passive and major subjects in Japanese. Linguistics 29, 1053-1092. No, Y. (1991). Case Alternations on Verb-Phrase Internal Arguments. Ph.D. dissertation, Ohio State University. Pollard, C. (1994). Toward a unified account of passive in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 273-296. CSLI Publications, Stanford, CA. Pollard, C., and I. Sag (1987). Information-Based Syntax and Semantics. CSLI Lecture Notes, Vol. 13. CSLI Publications, Stanford, CA. Pollard, C., and I. Sag (1994). Head-Driven Phrase Structure Grammar. University of Chicago Press, Chicago, and CSLI Publications, Stanford, CA. Pullum, G. (1982). Free word order and phrase structure rules. NELS 12, 209-220. Ryu, B-R. (1993). Structure Sharing and Argument Transfer: An HPSG Approach to Verbal Noun Constructions. Sfs-Report-04-93, University of Tubingen. Saito, M. (1985). Some Asymmetries in Japanese and Their Theoretical Implications. Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge, MA. Saito, M. (1992). Long distance scrambling in Japanese. Journal of East Asian Linguistics 1, 69-118. Sells, P. (1990). VP in Japanese: Evidence from -te complements. In H. Hoji (ed.), Japanese /Korean Linguistics, 319-334. CSLI Publications, Stanford, CA. Sells, P. (1991). Complex verbs and argument structures in Korean. In S. Kuno et al. (eds.), Harvard Studies in Korean Linguistics IV, 395-406. Hanshin, Seoul. Yatabe, S. (1993). Scrambling and Japanese Phrase Structure. Ph.D. dissertation, Stanford University. Yoo, E. J. (1993). Subcategorization and case marking in Korean. In A. Kathol and C. Pollard (eds.), Papers in Syntax. Ohio State University Working Papers in Linguistics Vol. 42, 178-198. Yoon, J. H. S. (1993). Tense, coordination, and the clausal structures of English and Korean. In S. Kuno et al. (eds.), Harvard Studies in Korean Linguistics V, 436-446. Hanshin, Seoul. Yoshimura, N. (1989). Parasitic pronouns. Paper presented to the Japanese/Korean Linguistic Conference at UCLA. Zwicky, A. (1986) Concatenation and liberation. CLS 22. 65-74.

CONSTITUENCY AND LINEARIZATION OF VERBAL COMPLEXES ANDREAS KATHOL Department of Linguistics University of California at Berkeley Berkeley, California 94720

1. INTRODUCTION

In the descriptive toolbox of modern feature-based theories such as Lexical-Functional Grammar (LFG) and Head-Driven Phrase Structure Grammar (HPSG), one idea that has led to important new ways to think about the nature of linguistic knowledge is the concept of informational identity, that is in the HPSG case token-identity of information from possibly rather disparate sources. Analyses based on this concept have, for instance, yielded significant descriptive and explanatory advances in the area of nonfinite complementation. On the common transformational approach, any dissociation between syntactic and semantic "subjecthood," such as in subject-to-subject raising constructions, is achieved by means of transformational relationships among syntactic tree representations. By contrast, it has become part of most nonderivational accounts of English to think of English raising constructions in terms of a (partial) identity between the raising verb's subject and the understood subject of the VP complement. In Pollard and Sag's (1994) rendition of raising as structure sharing, the subcategorization information of the subject-raising verb consists of an NP subject and a VP complement. That VP is an unsaturated verbal projection, which means that it still requires a subject to yield a full Syntax and Semantics, Volume 30 Complex Predicates in Nonderivational Synta.

221

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

222

Andreas Kathol

sentence. This subject is not provided by means of some inaudible element, but rather, the coreference tag " Q] " in (1) indicates token-identity with the subject of the embedding raising verb. In other words, the "understood subject" of the VP complement is realized syntactically as the subject of the higher predicate. (1)

[ . . . | SUBCAT ( CDNP, VP [ . . . | SUBCAT < CD >] ) ]

In the same lexicalist spirit, extending the domain of application for structure sharing further, Hinrichs and Nakazawa's concept of argument composition (first introduced in Hinrichs and Nakazawa, 1990 has had a great impact on the analysis of nonfmite complementation in head-final languages. The basic idea is to extend the linkage between the valence of the embedding and embedded predicates to all arguments, not just the subject. As the description in (2) illustrates, the whole valence value is shared among the raising verb itself and the verbal constituent subcategorized for by the latter. The entire list of selected arguments then consists of the singleton list containing the verbal complement appended (notated as " °") to the list of arguments "attracted" from the lower predicate. (2)

[. . . | SUBCAT H o ]

Ignoring the verbal complement, this means that the raising verb becomes an intransitive predicate if and only if the selected predicate is intransitive and so on for predicates of all arities, as exemplified for transitives in (3a). Moreover, in a language such as German which possesses impersonal predicates—that is, without a syntactic subject even in the finite forms—structure sharing of the valence information entails that the raising verb too will become impersonal if and only if the embedded verb is. This situation simply corresponds to the case in which the list of attracted arguments is empty, as shown in (3b): (3)

a. [ ... | SUBCAT El ° ]

b. [... I SUBCAT Q] < > ° ] As Kiss (1992) shows, there are essentially no syntactic differences in their behavior under argument composition between different classes of predicates in German that have not necessarily been treated alike in the transformational literature.1 Thus, not only classical raising verbs such as schein(en), but also tense auxiliaries and a subclass of modals pattern exactly alike in this characteristic realization of valence requirements of some predicate by some other verb. For instance, all three are eligible

Verbal Complexes

223

governors of impersonal passive constructions: (4) a. daB gearbeitet zu werden scheint. that worked to be seems 'because there seems to be working going on' b. daft gearbeitet worden ist. that worked been is 'because there has been working going on' c. daft gearbeitet werden muB. that worked be must 'because there ought to be working going on' As Kiss further shows, it is in fact possible to correlate the idea of argument composition with the notion of coherent construction, which was introduced into the syntactic description of German by Bech (1955). To this end, it is necessary to allow for a variant of argument composition in which the embedding predicate inherits all the arguments of the lower verb, but at the same time requires there to be a thematic subject. This situation arises in the case of certain EQUi-control verbs such as versuchen ('try'), which allow for a coherent complementation pattern indicative of argument composition as indicated for instance by the "scrambled" argument order in (5a), but at the same time obligatorily take a subject (5b). (5) a. daB ihm der Mann zu helfen versucht. that him-DAT the man-NOM to help tries 'that the man tries to help him.' b. *dafi gearbeitet zu werden versucht. that worked to be tries In such cases, the pattern of shared information is as outlined in (6):2 (6)

[. . . | SUBCAT Q] o ]

2. ARGUMENT COMPOSITION AND BRANCHING STRUCTURE In their original proposal, Hinrichs and Nakazawa assume that a cluster of verbal elements3 related by argument composition is the result of a sequence of binary combinations of an embedding verb with its verbal complement to the exclusion of any inherited ("proper") arguments. Canonical sequences of governors following governed verbs will then arise

224

Andreas Kathol

from left-branching structures, as the following example illustrates: (7) a. daft Peter das Buch finden konnen wird. that Peter the book find can will 'that Peter will be able to find the book' b.

However, the lexical specification of verbs involved in argument composition is actually silent with respect to the way that lexical information is projected onto syntactic structure. This means that alternative branchings are possible—and, in fact, have been proposed in the literature. For instance, Kiss (1992) adopts a right-branching analysis of the sort illustrated in (8): (8)

Apart from the directionality of branching, the analyses in (7) and (8) also differ in another respect, namely, whether the source of raised arguments can itself get raised by a higher predicate. Thus, in (7), only phrasal arguments get successively inherited from finden by konnen and wird. By contrast, in (8) the argument source finden is among the list of arguments that wird inherits from konnen. As a consequence, the highest governing predicate (here, wird) may accumulate an arbitrary number of verbal elements on its SUBCAT list whereas in Hinrichs and Nakazawa's approach,

Verbal Complexes

225

each governing verb will have at most one verbal complement in addition to the list of inherited phrasal arguments. The possibility of allowing sources for raised arguments to be inherited themselves by higher predicates does not automatically lead to a rightbranching structure.4 Rather, the syntactic structure of the verb cluster—if it is to be treated as a syntactic unit at all—crucially depends on the type of syntactic licensing mechanism adopted. While Kiss assumes binary branching generally, a popular alternative has been to assign a flat structure to the German Mittelfeld. On this approach, favored for example by Pollard et al. (1993), Baker (1994), van Noord and Bouma (1995), and Bouma and van Noord (1998), verb clusters simply constitute a subset of sisters within the flat tree making up a clause in a single step:5

(9)

This display of possibilities for the analysis of complex verbal structures in languages such as German raises the immediate question whether there is any empirical ground for a choice among the options. This study is an attempt to provide a positive answer to that question based on considerations of constituency and order variation. The analysis argued for here takes the left-branching analysis of Hinrichs and Nakazawa as its point of departure, contrasting it with the right-branching alternative which is found wanting in serious respects.6 However, closer inspection reveals that Hinrichs and Nakazawa's approach leads to a number of serious shortcomings which can be avoided if the licensing mechanism involved in the construction of verbal complexes is dissociated from that of phrasal arguments. This then also forms the basis for a novel approach toward the linearization of verbal complexes which is able to accommodate certain kinds of discontinuities seen in various German dialects and Dutch.

3. CONSTITUENCY 3.1. The Case for Left Branching One of the crucial arguments adduced by Kiss for his choice of a right-branching structure over a left-branching alternative is that it ex-

226

Andreas Kathol

ploits the division into word vs phrase-level constituents that is already provided by the grammar. That is, the valence characterization of an argument raiser such as wird in (8) makes reference to the fact that the element providing the inherited arguments is a word. In HPSG, the distinction between lexical and phrasal constituents is often made in terms of different values of the feature LEX; accordingly the lexical description of an argument-composing predicate would require a positive specification of this property:

(10) Hence, argument composition involves what Kiss calls lexeme selection. Together with ordinary, phrasal selection, this then sets up a typology of complementation types which is in parallel to the word-phrase distinction already made in the grammar. Contrast this situation now with the kind of constituents involved in left-branching structures. In (8), we need to identify sequences such as finden konnen as syntactic units that can be selected by the embedding verb. As things stand, the only distinguishing property between a "pure" verbal complex like finden konnen and one that has combined with phrasal complements such as Peter das Buch finden konnen lies in their valence—yet the central characteristic of argument raisers like wird is that they are indiscriminate with respect to the length and content of the inherited argument list. Thus in order to force the construction of verbal complexes to the exclusion of phrasal arguments, Hinrichs and Nakazawa need to have a way of distinguishing purely verbal combinations from those containing phrasal constituents and make the selectional properties of argument inheritors sensitive to that distinction. For this purpose, they postulate the feature NPCOMP, which bears a negative value so long as no phrasal constituent has combined with the amalgam of verbal elements:

(11)

Verbal Complexes

227

At first blush this seems like a crude way of achieving the required task, as it does not appear to pick out a natural class of objects or line up with any other distinction made in the grammar. But following a proposal first made by Rentier (1994), a somewhat different conclusion suggests itself in which the combination of verbal elements fits more straightforwardly into the typology of syntactic categories than Kiss suggests. Rentier adopts the characterization in (10), but gives a somewhat different interpretation to the feature LEX. Instead of lining up [LEX ± ] with HPSG's sortal distinction of signs into word and phrase, Rentier also allows verbal combinations such as finden konnen to bear the specification [LEX + ]. Being of sort word is then a sufficient, but no longer necessary, condition of a positive value of the LEX attribute. We can therefore regard elements with such a positive specification as possessing "(quasi-)lexical" status. This means that, even though syntactically the status of (partial) verbal complexes is phrasal, the amalgamation of verbal elements nevertheless behaves as a single element with regard to selecting phrasal arguments.7 Further crosslinguistic support for this conception comes from considering the status of corresponding structures in other head-final languages such as Korean and Japanese. While, in German, verbal complexes clearly constitute combinations of separate lexical elements—this conclusion is inevitable8 given the kind of order variation to be discussed in later sections—a clear separation from morphological amalgams, particularly those involving derivation or compounding, becomes much less clear (cf. for instance causatives (lida et al., 1994:7)): (12) a. Ken-ga Naomi-ni hon-o yom-ase-ta. Ken-NOM Naomi-DAT book-ACC read-CAUS-PAST 'Ken made Naomi read the book.' b. daft Hans Maria das Buch lesen lassen hat. that Hans Maria the book read made has 'that Hans made Maria read the book' From this perspective, then, both German and Japanese verbal complexes consist of combinations of verbal material with an exactly parallel combinatorial structure resulting in a construction that behaves like a single lexical unit ([LEX + ]) with respect to valence properties. The only difference is whether the combination takes place exclusively in the syntax or may also involve other components of the grammar, that is, morphology. By contrast, no comparable parallelism would exist if a right-branching analysis were adopted. This is because, in analogy with (8), one would be required to make the implausible assumption that strings of derivational affixes always combine into some morphological unit before attaching to the main verb. Of course this argument only carries force if one is willing

228

Andreas Kathol

to allow for the possibility of certain linguistic phenomena exhibiting parallel properties even though both may belong to clearly distinct grammatical subsystems, but this seems hardly a contentious issue. Having shown that Kiss' typology-based argument for right-branching structures can easily be countered by an appeal to cross-linguistic parallelism of structure, we now move on to factual empirical evidence for left branching. This involves two constructions in which the subconstituents postulated within the verbal complex actually exhibit properties of a syntactic unit. In the first of these constructions (sequences of) nonfinite verbs occur fronted in prefinite position in V2 clauses. As the examples in (13) show, not only ordinary phrasal constituents such as the direct object in (13a), but verbal constituents as well may be fronted, either alone, as in (13b), or as a sequence, as in (13c):9 (13) a. Das Buch wird Peter finden konnen. the book-ACC will-FIN Peter-NOM find can 'Peter will be able to find the book.' b. Finden wird Peter das Buch konnen. find will-FiN Peter-NOM the book-ACC can c. [Finden konnen] wird Peter das Buch. find can will-FiN Peter-NOM the book-ACC Even though, as we will see in due course, the exact formulation of this pattern presents certain difficulties under standard HPSG assumptions about complementation, the basic explanation is quite straightforward. Any governed subconstituent within a verbal complex can be fronted, that is, any verb or combination of verbs that is selected by a higher predicate (inheriting the former's arguments). By contrast, a right-branching structure not only has to accommodate the frontability of elements that do not form a syntactic unit,10 it also provides syntactic groupings, such as konnen wird, that do not behave as such under any diagnostic.11 Furthermore, it is far from clear how to reconstruct a relevant notion of government under the premise that branching is to the right. One may, for instance, consider contiguous sublists of verbal elements on the SUBCAT list as candidates for preposing. But not only would this solution be ad hoc, more significantly this move would essentially amount to reconstructing a notion of syntactic unit that is already given for free in the left-branching analysis. Another powerful argument against the right-branching analysis comes from order variation within the verbal complex in German. In cases where the verbal cluster contains a contiguous sequence of two or more morphological bare infinitives,12 the governing verb may (as in the case of the future auxiliary werden) or must (as with the perfect auxiliary haben)

Verbal Complexes

229

precede, rather than follow, the sequence of governed verbs. This type of construction is known among German grammarians as Oberfeldumstellung; an alternative label Hinrichs and Nakazawa use is Aux Flip. (14) a. daft Peter das Buch wird finden konnen. that Peter the book will find can 'that Peter will be able to find the book' b. daft Peter das Buch hat finden konnen. that Peter the book has find can 'that Peter has been able to find the book' The exact mechanisms employed by Hinrichs and Nakazawa to account for this construction will be subject to more detailed scrutiny later. At this point, however, it is important to note that their proposal lends itself rather naturally to cover these cases in terms of a simple reordering of the governing verb with respect to the governed subcomplex:13

(15)

By contrast, in a right-branching analysis, Oberfeldumstellung cannot be accounted for in terms of a simple reordering between governing verb and governed subcomplex. In fact, it is not clear if any mechanism could be found that yields an analysis of Oberfeldumstellung which remains faithful to the constituent structure in (8)—to the best of our knowledge, no such proposal has ever been advanced.14 The two phenomena considered here—partial verb cluster fronting and Oberfeldumstellung—thus provide strong evidence for the kind of constituency structure that arises in connection with assigning a binary leftbranching structure to the verb cluster. Whatever the initial plausibility of right-branching analysis, it has been shown that it cannot make up for the complete inability to provide any insight into these constructions. On the other hand, a reinterpretation of the function of the LEX feature allows us

230

Andreas Kathol

to view verbal complexes in German as the syntactic realization of complex predicates which in other languages exhibit more distinctly morphological properties. Despite its success, however, a closer look at partial verb cluster fronting cases will necessitate a reconsideration of Hinrichs and Nakazawa's idea that verbal complexes are built up via the same valence attribute as is involved in phrasal complementation, viz. SUBCAT.15 3.2. The Case for VCOMPL When we consider what parts of a verbal complex can undergo fronting, it turns out that there are severe limits which have long been known in the literature. Thus, while a main verb can be fronted alone, as was shown in (13b), repeated in (16a), this is not the case for a predicate that triggers argument composition. Thus, the example in (16b) is strongly ungrammatical: (16) a. Finden wird Peter das Buch find will-FiN Peter-NOM the book-Ace b. *Konnen wird Peter das Buch can will-FiN Peter-NOM the book-Acc

konnen. can finden. find

On the standard assumption that all types of selectional dependencies arise from regular valence features—that is, either the unitary SUBCAT feature or combination of SUBJ and COMPS features—this intuition turns out to be rather difficult to state succinctly. For instance Nerbonne (1994:139) imposes on his lexical rule for complement extraction the condition that if (the projection of) a modal verb is fronted via SLASH, there must not be any verbal element on that modal verb's SUBCAT list. This condition hence involves negated existential quantification over set elements. In addition, it is not clear if one can find a natural motivation for such a constraint. In contrast to modal verbs, a predicate such as versuchen may take a phrasal complement with zu -infinitival morphology. In such cases, the verbal complement seems to be able to "stay behind" when the governing predicate is fronted—examples such as in (17), while not entirely perfect, nevertheless are markedly better than the one in (16b): (17) ?Versprochen wird er ihr wohl [den Wagen zu waschen] promised will he her probably the car to wash nicht haben. not have 'He will probably not have promised her to wash the car.' If this is correct, Nerbonne's condition against inheriting verbal elements from modals (or tense auxiliaries) seems arbitrary.16

Verbal Complexes

231

While there seems to be a general consensus to assimilate the construction of verbal complexes to complementation in general by regarding the selection to be in terms of SUBCAT/COMPS, the particular assumptions about branching structure nevertheless force Hinrichs and Nakazawa to adopt a separate ID schema dubbed "Verbal Complex ID Schema" (Hinrichs and Nakazawa, 1994b). But from this, it is only a small step toward projecting (partial) verbal complexes by an altogether different valence attribute. This is indeed what Chung (1993, 1998) proposes with the new valence feature GOV for Korean, which has since been adopted for Dutch by Rentier (1994). We will do the same here, albeit using the name VCOMPL for "verbal complement/complex,"17 which is arguably more adequate as it stresses the special involvement in the formation of complex predicates. Given this reformulation, the lexical description of a predicate controlling argument composition is as along the lines given in (18):

(18)

The removal of the verbal complement from the SUBCAT list gives us, in fact, a straightforward way of characterizing frontable constituents as the natural class of elements that bear the specification [VCOMPL < )]. This description covers ordinary phrasal arguments at the same time that it excludes predicates such as konnen or other verbal sequences bearing a nonempty VCOMPL value. In addition, there are at least three more arguments for separating the dependencies that play a role in complex predicate construction from ordinary complementation via SUBCAT. Kiss (1992) presents an attempt to relate extraposability to the incoherent vs coherent construction type in German, which is correlated with the selection of VP complements vs argument composition. From that perspective, it is tempting to identify extraposable verbal constituents in quite general terms by virtue of their valence properties, i.e., to allow for VP complements to be extraposed while the nonphrasal projections selected in argument composition are barred from extraposition. The problem with such a proposal, however, is the existence of VPS headed by lexically intransitive verbs. If a verb of this kind, for instance schlafen ('sleep'), combines with an argument-raising governor, it will bear the same valence properties as a "genuine," phrasal VP, namely, [SUBCAT {synsem}]. Yet, as the example in (19b) illustrates, despite the identity in valence and morphological properties (zw-infinitive), the

232

Andreas Kathol

complement of scheinen may not extrapose: (19) a. weil Lisa versucht [zu schlafen] because Lisa tries to sleep 'because Lisa tries to sleep' b. *weil Lisa scheint [zu schlafen] because Lisa seems to sleep One possibility is to assign a difference in terms of the word-phrase status for the complement of versuchen vs scheinen. This could be accomplished by means of Pollard's (1996) Schema B', which allows the creation of a phrasal verbal projection even when there are no complements to discharge. But then this schema gives rise to a nonbranching structure for which one can arguably provide no theory-independent motivation. On the other hand, once the distinction is made between selection via VCOMPL vs SUBCAT, it becomes straightforward to treat complements differently in each case. Only verbal complements18 selected via SUBCAT are eligible for extraposed occurrence, but never those selected by VCOMPL—regardless of their morphological or valence properties. There is a second aspect in which the formation of verbal complexes via VCOMPL seems advantageous over an approach that does not distinguish between phrasal and quasi-lexical arguments. While separable prefixes in German are mainly prepositional in origin, there are examples of noun + verb combinations such as Autofahren ('drive (a) car'), in which the nominal part arguably has the status of a separable prefix. In some cases, such "incorporated" nouns occur in the plural, as for instance in Kartenspielen ('play cards'). If the object Karten is selected via VCOMPL, then we have the prediction that even though it has the appearance of an accusative object, it should not be eligible to undergo passivization—the latter being a valence alternation only for subcategorized, phrasal arguments selected via SUBCAT. As a result, the whole predicate should count as intransitive, allowing for impersonal passive. As Kroch and Santorini (1991:295) observe, this is precisely what we find, as shown by the number inflection on the passive auxiliary in (20): (20) Damals wurde/* warden haufig Karten gespielt. then was/were often cards played 'At that time, there was frequent playing of cards.' Finally, as has been pointed out for instance by Stiebels and Wunderlich (1994:927), separable prefixes behave differently from regular phrasal arguments in progressive constructions formed with the amalgamated preposition am ('at the') in certain German dialects. As is illustrated in (21), am necessarily precedes the elements of the verb cluster, including

Verbal Complexes

233

separable prefixes such as auf: (21) a. Er ist sein Zimmer am auf he is his room at-the up He is cleaning his room.' b. *Er ist am sein Zimmer auf he is at-the his room up c. *Er ist sein Zimmer auf am he is his room up at-the

raumen. clean-iNF raumen. clean-iNF raumen. clean-iNF

The very same behavior can be seen with noun + verb combinations lending support to the claimed syntactic parallelism of these bare nouns with other elements of the verb cluster. (22) a. Er ist am Auto fahren. he is at-the car drive-iNF 'He is driving a car.' b. *Er ist Auto am fahren. he is at-the car drive-iNF The pattern in (22) is in marked contrast to the behavior of phrasal arguments such as the full NP seinen VW ('his VW). As is shown in (23), it is impossible to place the progressive marker am before the full NP:19 (23) a. ?.£> ist seinen VW am fahren. he is his VW at-the drive-iNF 'He is driving his VW.' b. *Er ist am seinen VW fahren. he is at-the his VW drive-iNF The distinction between arguments internal to the verb cluster and phrasal arguments outside of it is straightforwardly captured in our analysis in terms of the distinction in selectional attributes.20 Having established empirical motivation for the kind of branching structure proposed by Hinrichs and Nakazawa, we can now address the question of what constraints govern the linearization of verbal complexes.

4. LINEARIZATION

It has been a common assumption that despite their close syntactic similarity in other respects, German and Dutch differ in the organization of the elements in the verb cluster. In particular, verb clusters in German are typically head-final, whereas those in Dutch are head-initial. This

234

Andreas Kathol

observation has prompted Rentier (1994) to use his complex-building valence attribute to set up the distinction between orderings in German vs Dutch in terms of the linear precedence (LP) constraints sketched in (24):21 (24) a. German [SYNSEM D] ]

- < [ . . . | VCOMPL < [H >]

b. Dutch [... I VCOMPL < CD >] -< [SYNSEM 0] Thus, while the configuration of clusters in German and Dutch is the same, the two languages differ in that the first places the governor after the governed subcomplex, whereas the opposite order is observed in Dutch. However, this formulation is obviously too simplistic. As the examples of Oberfeldumstellung in (14) demonstrate, there are indeed instances in which the governor precedes, rather than follows, its verbal complement in German.22 Moreover, the treatment of Dutch in terms of (24b) is equally simplistic as it totally ignores the phenomenon, sometimes referred to as inversion (cf. van Noord and Bouma, 1995), in which the order between governor and governee exhibits left-headedness. This is a frequently attested possibility for tense auxiliaries such as the perfectivizer hebben (25), and to a lesser extent also seen with modals (26), such as willen ('want'): (25) a. dat Joop de krant heeft gelezen. that Joop the newspaper has read 'that Joop has read the newspaper' b. dat Joop de krant gelezen heeft. that Joop the newspaper read has (26) a. dat Joop de krant wil lezen. that Joop the newspaper wants read 'that Joop wants to read the newspaper' b. dat Joop de krant lezen wil. that Joop the newspaper read wants It is therefore necessary to make reference to more fine-grained distinctions than the government relationships alone. As a second approximation toward an adequate account of ordering relations within the verb cluster, we adopt a solution which is essentially isomorphic to the one Hinrichs and Nakazawa propose to account for the Oberfeldumstellung cases in (14). Their central idea is to record on nonfinite verbs the possible relative position of any embedding governor. This is

Verbal Complexes

235

achieved via the binary-valued head feature FLIP. Negative specification indicates canonical government to the left, whereas a positive value requires government by a "flipped," i.e., preceding, governor. To avoid unwanted connotations regarding markedness or exceptionality, we replace FLIP with the head attribute G(O)V(ERN)OR whose value directly indicates the required placement of a governor.23 Thus, [GVOR -> ] records on a verb that, in a verb cluster, its governor has to appear to the right, whereas [GVOR ]

b. [... I VCOMPL < E > ] -< [SYNSEM 03 [... I GVOR «- ]] Apart from being able to make the requisite fine-grained distinctions regarding order, this formulation has the obvious advantage over Rentier's system that German and Dutch can be treated on a par with respect to the involved LP constraints, while differing in the possible values that (classes of) lexical items may bear. For instance, let us assume that nonfinite main verbs in Dutch are underspecified in terms of their value for GVOR (hence as the GVOR value we have dir, which is the supersort of the atomic sorts -> and ]

b. Dutch [. . . | HEAD | GVOR dir]

We then have an immediate account of why inversion with main verbs is never possible in Standard German; cf. (29): (29) a. *dqfi that b. *daB that

Lisa Lisa Lisa Lisa

die Zeitung the newspaper die Zeitung the newspaper

hat gelesen. has read will lesen. wants read

Turning to nonmain verbs in German, we note that their nonfinite forms display different behavior with respect to the positioning of a higher governor. In particular infinitival modals and a few other predicates which themselves take a bare-infinitival complement are underspecified regarding their GVOR value. (30)

236

Andreas Kathol

This then gives rise to an analysis of Oberfeldumstellung that is isomorphic to Hinrichs and Nakazawa's initial proposal. Note in particular that since GVOR is a head feature, the value chosen on the infinitival auxiliary will also be the same borne by the subcomplex headed by a verb such as konnen. The resulting distribution of values is given for the canonical order in (3la) and the Oberfeldumstellung case in (31b):

(31) a.

b.

As Hinrichs and Nakazawa further observe, the account given so far ignores the fact that if both haben and werden occur in a sentence, their properties with respect to Oberfeldumstellung interact. Hence, our previous discussion could be taken to suggest that cluster-final placement is always possible, whereas Oberfeldumstellung is rather restricted, as is evidenced for instance by the ungrammaticality of (29). In particular then, if the verbal complement of werden contains an element such as haben which also displays variation in the linear relations with the complement, the positioning of werden should nevertheless be independent. However, this is not the case, as the examples in (32) show. Only the variant (32a) is grammatical, with its structure shown in (33): (32) a. dafi er die Lieder wird [haben singen konnen]. that he the songs will have sing can 'that he will have been able to sing the songs' b. *daB er die Lieder [haben singen konnen] wird. that he the songs have sing can will

Verbal Complexes

237

Hinrichs and Nakazawa (1994a) propose to account for this fact by assuming that haben takes on the trigger properties of the complement it embeds; hence the infinitival form of haben must satisfy the following description:24 (34)

This means that whenever haben is involved in an Oberfeldumstellung construction, the whole constituent will itself require a higher governor to precede. Otherwise, it will occasion the governor verb to occur finally, as in (35):25 (35)

Hinrichs and Nakazawa's (1994a) account can be shown to correctly capture all the grammatical permutations of verbs and their verbal complements listed in Bech (1955:63). Using his convention of indicating depth

238

Andreas Kathol

of embedding by increasing subscripts, these can be listed as in (36):26

(36)

For comparison, this table also lists two ungrammatical cases in which the highest verb "ignores" the reversal of the governee-governor order occurring in its complement (cf. (32b)). One important aspect of Hinrich and Nakazawa's system, which can be seen rather clearly in Bech's tabular representation (36), is the fact that the sequence of verbs can be divided into at most one subsequence of monotone increasing levels of embedding and exactly one monotone decreasing subsequence of levels of embedding—separated in (36) by means of "|". The first of these subsequences corresponds to what Bech calls Oberfeld ('upper field'), while the second is his Unterfeld ('lower field'), hence the terminology Oberfeldumstellung ('displacement into the upper field'): (37)

In the case of i = n — 1, the Oberfeld is empty; that is, no governor occurs in reversed position. In the next section we will encounter evidence that challenges the idea that verbal complexes can always be correlated with a linear partitioning of the set of elements that form part of the verb cluster. 4.1. Discontinuous Verbal Complexes Bech's theory of order within the verb cluster and its implementation in HPSG terms by Hinrichs and Nakazawa discussed above is to some extent a sanitized version of the facts. For one thing, it ignores the possibility of nominal elements occurring among cluster elements in what are commonly referred to as V-Projection Raising constructions such as in (38):27, 28 (38) daft Lisa dem Jungen wird das Buch geben wollen. that Lisa the boy-DAT will the book give want 'that Lisa will want to give the book to the boy' In this study, however, we focus on another set of apparent counterexamples to the implicit claim that governors may never occur inside the

Verbal Complexes

239

sequence of cluster elements corresponding to the governed verbal complex. There exists substantial evidence that this view is too strict, as such interspersals do indeed occur in a number of dialects.29 Meurers (1994) gives a number of grammatical examples in which the preposed auxiliary's placement is within the Unterfeld. Examples of this construction, referred to by Meurers as Zwischenstellung ('intermediate placement') or Verbal Complex Split, are given in (39) (Meurers, 1994): (39) a. zu dem Zeitpunkt an dem ich mich entscheiden3 hdttel mussen2. at the point at which I me decide had must 'at the point at which I should have made a decision' b. daft er das Examen bestehen3 wird/hat1 konnen2. that he the exam pass will/has can 'that he will be/has been able to pass the exam' c. daft ihm die Entscheidung nicht durchgehen3 hdtte2 durfen2. that him the decision not slip had should 'that the decision should not have passed him by' d. daft es so kommen3 hdttel mussen2. that it so come had must 'that it had to come to this' According to Meurers, the presence of two flipped auxiliaries leads to the following pattern of grammaticality:30 (40) daft er das Buch . . . that he the book a. wird1 haben2 kopieren5 Iassen4 mussen^. will have copy let must b. wirdl kopierens haben2 Iassen4 mussen3. c. ?wjrrf1 kopieren5 Iassen4 haben2 mussen^. d. *wird1 kopieren5 Iassen4 mussen3 haben2. e. kopieren5 wird1 haben2 lassen4 mussen3. f. Ikopieren5 wird1 Iassen4 haben2 mussen3. g. *kopieren5 wird1 lassen4 mussen3 haben2. h. kopieren5 lassen4 wirdl haben2 mussen3. i. *kopieren5 lassen4 wird1 mussen3 haben2. ]. *kopieren5 lassen4 mussen2 wirdl haben2. Furthermore, as we noted earlier, Standard Dutch displays the inversion order with tense auxiliaries, that is, the reversal of the typical governor-governee serialization. As has been pointed out, for instance, by van Noord and Bouma (1995), inversion is also possible with nonfinite forms of the tense auxiliary hebben. This means that in addition to the

240

Andreas Kathol

canonical order in (41a), the examples in (41b,c) are possible as well: (41) a. dat Jan dit boek moetl hebben2 gelezen3. that Jan this book must-FiN have-iNF read-iNF 'that Jan must have read book' b. dat Jan dit boek moetl gelezen3 hebben2. that Jan this book must-FiN read-psp have-iNF c. dat Jan dit boek gelezen3 moet1 hebben2. that Jan this book read-PSP must-FiN have-iNF Under the assumption that the same government and constituency relations are involved in these examples as are in the German cases, the cluster in (41b) is structurally isomorphic to German Oberfeldumstellung, whereas (41c) is a variant of the German Zwischenstellung cases in (39). Thus, a strictly phrase-structure-based view would predict that only (41b) should be possible, while the governor's occurrence internal to the governed cluster in (41c) should be illicit. However, the judgments in these cases are the exact reversal of the predictions of the theory. As van Noord and Bouma (1995) point out, sentences such as (41c) are judged to be grammatical in all dialects of Dutch; those similar to (41b) appear to be rather limited in acceptance and are possible for the most part only in Flemish dialects. A rather similar observation can be made with respect to the placement of separable prefixes. In contrast to most German dialects, separable prefixes in Dutch may "float" to earlier positions away from the main verb they cooccur with. For instance the prefix aan in aanspreken ('address') may occur not only in immediate adjacency with the main verb spreken (42a), but may also be separated from the latter by one or more intervening auxiliaries; cf. also den Besten and Edmondson (1983:193):31 (42) a. dat Jan Marie zou hebben aangesproken. that Jan Marie would have FREF.spoken 'that Jan would have spoken to Marie' b. dat Jan Marie zou aan hebben gesproken. that Jan Marie would PREF have spoken c. dat Jan Marie aan zou hebben gesproken. that Jan Marie PREF would have spoken Since prefix + verb combinations in West Germanic exhibit many of the properties of partial verbal clusters (for instance with respect to stress assignment, nonintervention of adverbs), these cases again arguably involve the discontinuous linearization of the verbal complement of a governor inside the verb cluster (here, hebben)—contrary to the predictions made by the Hinrichs-Nakazawa model of the verbal complex.

Verbal Complexes

241

Various solutions suggest themselves to the problem just encountered. In the tradition of transformational analyses, following Evers (1975), one could propose some manipulation of the constituent structure that achieves the desired linearization possibilities. However, since the ordering alternations are an exclusively local phenomenon, the existing HPSG mechanisms for nonlocal dependencies are arguably ill-suited for a reanalysis-based approach. Therefore a different line will be pursued here, which also has the significant conceptual advantage of allowing for a uniform analysis of government and constituency relations throughout West Germanic. In particular, we propose to analyze the problematic cases at face value as involving a discontinuous (partial) verb complex. As the requisite framework we adopt a variant of HPSG that possesses order domains, developed in Kathol (1995a), which in turn is based on ideas first advanced by Reape (1993). Order domains, appropriate for objects of sort sign, can be understood as totally ordered lists of information bundles which each contain phonological and syntactic-semantic information.32 Simplifying somewhat, at each application of a combinatorial schema, a corresponding order domain is built up involving the domains of the constituents thus far derived. Depending on general principles of domain construction, two basic scenarios need to be distinguished. The first possibility is that a constituent is entered into the resulting domain of the mother as a single phonologically encapsulated informational chunk. Consider for instance the combination of a verb and an NP object in (43): (43)

Here, the domain of the VP consists of two elements, one for the head (sieht) and one constructed from the complement (die Rose). In the domain resulting from the verb-object combination, the internal linear composition of die Rose can no longer be referenced. In the terminology of Kathol (1995a), the object has been domain-inserted into that of the head. As with the construction of local trees in G/HPSG, no assumption about order is mi'de at this level; instead, this will be the task of general

242

Andreas Kathol

LP constraints. As a result, LP constraints such as (27) must be thought to order domain elements, as opposed to arranging the PHON values of daughters in local trees. As the second domain construction possibility, the domain of the mother can be computed as the domain union obtained from the domains of the daughters. Domain union is equivalent to the shuffle operation; that is, given two lists A and B, a shuffling of A and B will contain the same members as A and B combined and preserve the ordering originally holding among members of each of the component lists. What is not guaranteed, however, is that A and B will necessarily be represented as contiguous sublists of the resulting list. The notion of shuffle was first introduced by Reape (1993) under the name sequence union for the analysis of nonfinite complementation structures in West Germanic. However, as will be discussed in more detail in section 5, his conception of constituency is rather different from the one argued for here. The heart of our approach to the linearization of verbal complexes lies in the assumption that the binary governor-governee combinations do not involve the insertion of the governee complex into the domain of the mother node—as is in effect assumed in Hinrichs and Nakazawa's treebased model. Instead, if the combination involves domain union, then as a result, the internal components of the governee will still be "visible" for the purposes of linearization. As a concrete example, consider the domain construction associated with a "canonical" German verb cluster, as in (44): (44)

When finden and konnen are combined, each contributes only one domain element; hence the effect of domain union will be the same as if finden had been inserted into konneris domain. However, when the resulting

Verbal Complexes

243

subcomplex is combined with wird, there are now three placement options: before finden konnen, between finden and konnen, and following finden konnen. In the case of canonical orderings, only the last is grammatical, as required by the specification [GVOR ->] on the subcomplex as inherited from konnen. Yet, that subcomplex can no longer be referred to as a separate element within the order domain. As a result, we need to revise our LP constraints in (27) to reflect this fact by making reference not to the entire governed complex, but only its head. The following LP constraint achieves the desired effect and hence supersedes the earlier formulation in (27a):33

(45) If the governor has to follow the head of the governed verbal complex, it necessarily has to follow all elements of the verbal complex; hence reference to the head in a head-final structure has the same effect as reference to the entire selected subcluster as in (27a.) At this point, we may also want to briefly consider how linearized verbal complexes are embedded into larger clausal structures. Kathol (1995a) proposes a linear model of German clause structure which is rather close in spirit to the traditional "topological fields" model (cf. for instance Drach, 1937; Engel, 1970; Hohle, 1986).34 In that formalization, a topological field constitutes an equivalence class of elements within a clausal domain. While some of these equivalence classes have cardinality restrictions (at most one Vorfeld element, exactly one domain element instantiating linke Satzklammer) others do not have such restrictions. We may then regard the verb cluster simply as one topological field, here encoded as the subsort35 uc on elements of sort dom(ain)-obj(ect).36 An example of a topological organization for a subordinate clause is given in (46):

(46)

As was argued by Kathol (1995b) and Kathol (1995a), the linear organization can also serve as the basis for a positional, albeit non-movement-based account of verb placement variability in German. Thus, a verb-initial clause such as in (47) is distinguished from the verb-final one in (46) solely on the basis of the positional assignment of the finite verb wird to cf

244

Andreas Kathol

instead of vc: (47)

Even though in the analysis adopted here, both fronted verbs and members of the verb cluster are part of the same clausal domain, the frontal occurrence does not interact with the linear constraints operative within the verb cluster. This becomes immediately apparent if one compares the relative order possibilities as in (48): (48) a. Kann er das Buck finden? can he the book find 'Can he find the book?' b. *daB er das Buck kann finden. that he the book can find In other words, a verb in frontal position is outside the scope of any constraints on order that apply to verb cluster elements. For that reason, we need to assume that the LP constraint in (45 a) only applies to elements that are topologically marked as part of the verb cluster. The requisite refinement given in (49) yields the final version of the verb cluster LP constraints. It also lists the head-initial mirror version of that LP constraint, and it is to those structures that we turn next:37

(49)

What are the valid serializations that comply with constraint (49b)? Considering first the case of Dutch, where it is possible—in fact the unmarked option—to build verbal complexes from sequences of headinitial combinations, we notice that such situations result in the precise mirror image of the German case (44). That is, if the governor must precede the head of a head-initial structure, then this will have the same linear effect as if the governor had been linearized with respect to the entire governed complex. In (50), an example is given in which arrows link

Verbal Complexes

245

each governor and the head of the governed subcomplex:

Once we have mixed precedence requirements, the linearization-based approach is clearly more flexible than one based on deriving order from tree encodings directly. Thus consider what happens when konnen's GVOR value is instantiated as ]. 41 However, we currently have little insight to offer as to the nature of the unacceptability of the sequences in (41b) moet gelezen hebben and (42b) zou aan hebben gesproken. One possible line to pursue may be based on the observation made in Abraham (1994:42) that elements bearing a nuclear accent generally seem to have a strong tendency to occur at the left or right edge of the verb cluster. Both in (41b) and in (42b), this Fokusrandprinzip ('principle of peripheral focus') is violated. However, this still does not address the (diverging) distinctions in acceptability reported for different varieties of Dutch for such examples. 42 This requirement is too strong to accommodate examples such as (ib) from Johnson (1988:118-119), in which a higher finite governor follows, rather than precedes, a cluster headed by a nonfinite modal. However, examples of this type are uniformly rejected by all of my informants. (i) a. dat hij het [heeft [kunnen zien]]. that he it has can-iNF see-iNF 'that he has been able to see it'

Verbal Complexes

267

b. dat hij het [[kunnen zien] heeft] that he it can-iNF see-iNF has 'that he has been able to see it' 43

These are typically referred to among German grammarians as "Acl" (for accusativum cum infinitivo 'accusative with infinitive'). 44 For discussion on this point, see Kiss (1992:226). 45 Here, we make the simplifying assumption that DAT-NOM passives should only be possible if the verb in addition takes an accusative object. As Webelhuth and Ackerman (1992) observe, the acceptability of accusative-less instances as in (i) exhibits a great deal of speaker- and lexeme-dependent variation. (i) a. Ich helfe dem Jungen. I help the boy-DAT 'I help the boy.' b. DerJunge bekommt geholfen. the boy-NOM receives helped 'The boy is helped.' 46

For accounts of impersonal passives in HPSG, see for instance Pollard (1994) and Kathol (1994). The former proposes a uniform account for both personal and impersonal variants of passive. 47 The issue of finding a uniform characterization of passive auxiliaries occurring in personal and impersonal passives will have to remain unresolved here.

REFERENCES Abraham, W. (1994). Kaynes Asymmetriehypothese und die Abfolge im V-Komplex. In J.-W. Zwart (ed.), Studies in Minimalism and Kayne's Antisymmetry Hypothesis, Groninger Arbeiten zur Germanistischen Linguistik, Vol. 37, 19-46. Groningen. Baker, K. (1994). An integrated account of "modal flip" and partial verb phrases fronting in German. In Proceedings of the 30th Regional Meeting, Chicago Linguistic Society. Bech, G. (1955). Studien uber das deutsche Verbum Infinitum. Danske Historiskfilologiske Meddelelser 35, 2. Bierwisch, M. (1990). Verb cluster formation as a morphological process. In G. Booij and J. van Marie (eds.), Yearbook of Morphology, 173-199. Foris Publications, Dordrecht. Borsley, R. (1989). An HPSG approach to Welsh. Journal of Linguistics 25, 333-354. Bouma, G. and G. van Noord (1998). Word order constraints on verb clusters in German and Dutch. In E. Hinrichs, A. Kathol, and T. Nakazawa (eds.), Complex Predicates in Nonderivational Syntax. Syntax and Semantics, Vol. 30, 43-72. Academic Press, San Diego.

268

Andreas Kathol

Bresnan, J., R. Kaplan, S. Peters, and A. Zaenen (1982). Cross-serial dependencies in Dutch. Linguistic Inquiry 13, 613-635. Chung, C. (1993). Korean auxiliary verb constructions without vp-nodes. In S. Kuno, J. Whitman, Y.-S. Kang, I.-H. Lee, J. Maling, and Y. J. Kim (eds.), Harvard Studies in Korean Linguistics V, 274-286. Hanshin, Seoul. (Proceedings of the 1993 Workshop on Korean Linguistics.) Chung, C. (1998). Argument composition and long-distance scrambling in Korean: An extension of the complex predicate analysis. In E. Hinrichs, A. Kathol, and T. Nakazawa (eds.), Complex Predicates in Nonderivational Syntax. Syntax and Semantics, Vol. 30, 159-220. Academic Press, San Diego. den Besten, H., and J. Edmondson (1983). The verbal complex in continental West Germanic. In W. Abraham (ed.), On the Formal Syntax of the Westgermania. Papers from the Third Groningen Grammar Talks, 155-216. Benjamins, Amsterdam/Philadelphia. Drach, E. (1937). Grundgedanken der deutschen Satzlehre. Diesterweg, Frankfurt. Engel, U. (1970). Regeln zur Wortstellung. In U. Engel (ed.), Forschungsberichte des Instituts fur deutsche Sprache Mannheim, Vol. 5, 1-148. Gunter Narr, Tubingen. Evers, A. (1975). The Transformational Cycle in Dutch and German. Dissertation, University of Utrecht. Haider, H. (1993). Deutsche Syntax—Generativ. Vorstudien zur Theorie einer projektiven Grammatik. Gunter Narr, Tubingen. Heinz, W., and J. Matiasek (1994). Argument structure and case assignment in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in HeadDriven Phrase Structure Grammar, 199-236. CSLI Publications, Stanford, CA. Hinrichs, E., and T. Nakazawa (1990). Subcategorization and VP structure in German. In S. Hughes and J. Salmons (eds.), Proceedings of the Third Symposium on Germanic Linguistics. Benjamins, Amsterdam. Hinrichs, E., and T. Nakazawa (1994a). Linearizing finite AUX in German verbal complexes. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 11-38. CSLI Publications, Stanford, CA. Hinrichs, E., and T. Nakazawa (1994b). Partial VP and split NP topicalization in German: An HPSG analysis. In E. Hinrichs, D. Meurers, and T. Nakazawa (eds.), Partial-VP and Sptit-NP Topicalization in German—An HPSG Analysis and Its Implementation. Arbeitspapiere des SFB 340, Vol. 58. Hohle, T. (1978). Lexikalistische Syntax: Die Aktiv-Passiv-Relation und andere Infinitkonstruktionen im Deutschen. Niemeyer, Tubingen. Hohle, T. (1986). Der Begriff "Mittelfeld," Anmerkungen uber die Theorie der topologischen Felder. In Akten des 7. Internationalen Germanisten-Kongresses, Gottingen 1985, Vol. 3, 329-340. Tubingen. lida, M., C. Manning, P. O'Neill, and I. Sag (1994). The Lexical Integrity of Japanese Causatives. Presented at the LSA 1994 Annual Meeting, Boston. (Unpublished manuscript, Stanford University.) Johnson, M. (1988). Attribute-Value Logic and the Theory of Grammar. CSLI Lecture Notes, No. 16. CSLI Publications, Stanford, CA.

Verbal Complexes

269

Kasper, R. T., A. Kathol, and C. Pollard (1995). A Relational Interpretation of Linear Precedence Constraints. Unpublished manuscript, Ohio State University. Kathol, A. (1994). Passives without lexical rules. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 237-272. CSLI Publications, Stanford, CA. Kathol, A. (1995a). Linearization-Based German Syntax. Dissertation, Ohio State University. Kathol, A. (1995b). Verb-'movement' in German and topological fields. In A. Dainora, R. Hemphill, B. Luka, B. Need, and S. Pargman (eds.), Papers from the 31st Regional Meeting, Chicago Linguistic Society, 231-245. Chicago Linguistic Society, Chicago. Kathol, A. (to appear). Agreement and the syntax-morphology interface in HPSG. In R. Levine and G. Green (eds.), Readings in HPSG. Cambridge University Press, Cambridge, UK. Kiss, T. (1992). Infinite Komplementation: Neue Studien zum deutschen Verbum Infinitum. Dissertation, Bergische Universitat-Gesamthochschule Wuppertal. Kroch, A. S., and B. Santorini (1991). The derived constituent structure of the West Germanic verb-raising construction. In R. Freidin (ed.), Principles and Parameters in Comparative Grammar, 269-338. MIT Press, Cambridge, MA. Meurers, W. D. (1994). A Modified View of the German Verbal Complex. Presentation given at the 1994 HPSG workshop in Heidelberg. Meurers, W. D. (1996). German Partial-VP Fronting Revisited—Back to Basics. Unpublished manuscript, University of Tubingen. (Paper presented at the Third International Conference on HPSG, Marseille, France.) Miiller, S. (1996). Yet another paper about partial verb phrase fronting in German. In Proceedings of COLING-96, Copenhagen, 800-805. Nerbonne, J. (1994). Partial verb phrases and spurious ambiguities. In J. Nerbonne, K. Netter, and C. J. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 109-150. CSLI Publications, Stanford, CA. Pollard, C. J. (1994). Toward a unified account of passive in German. In J. Nerbonne, K. Netter, and C. Pollard (eds.), German in Head-Driven Phrase Structure Grammar, 273-296. CSLI Publications, Stanford, CA. Pollard, C. J. (1996). On head non-movement. In H. Bunt and A. van Horck (eds.), Discontinuous Constituency, 279-305. de Gruyter, Berlin/New York. Pollard, C. J., D. Levine, and R. T. Kasper (1993). Studies in Constituent Ordering: Toward a Theory of Linearization in Head-Driven Phrase Structure Grammar. Grant proposal to the National Science Foundation, Ohio State University. Pollard, C. J., and I. A. Sag (1987). Information-Based Syntax and Semantics, Vol. 1. CSLI Lecture Notes Series No. 13. CSLI Publications, Stanford, CA. Pollard, C. J., and I. A. Sag (1994). Head-Driven Phrase Structure Grammar. CSLI Publications, Stanford, CA, and University of Chicago Press, Chicago. Reape, M. (1993). A Formal Theory of Word Order: A Case Study in West Germanic. Dissertation, University of Edinburgh.

270

Andreas Kathol

Reis, M. (1979). Ansatze zu einer realistischen Grammatik. In K. Grubmiiller, E. Hellgart, H. Jellissen, and M. Reis (eds.), 1-21. Niemyer, Tubingen. Rentier, G. M. (1994). A lexicalist approach to Dutch cross dependencies. In K. Beals, J. Denton, E. Knippen, L. Melnar, H. Suzuki, and E. Zeinfeld (eds.), Papers from the 30th Regional Meeting of the Chicago Linguistic Society, 376-390. Chicago Linguistic Society, Chicago. Sailer, M. (1993). Zuruckin die Felder. 3 Ansatze zur Beschreibung infiniter Konstruktionen in Deutschen. Unpublished manuscript, University of Tubingen. Stiebels, B., and D. Wunderlich (1994). Morphology feeds syntax: The case of particle verbs. Linguistics 32, 913-968. van Noord, G., and G. Bouma (1995). Dutch Verb Clustering without Verb Clusters. Unpublished manuscript, University of Groningen. Webelhuth, G., and F. Ackerman (1992). Form and Function: The German Passive. Unpublished manuscript, University of North Carolina and University of California, San Diego. Zwicky, A. (1990). Syntactic words and morphological words, simple and composite. In G. Booij and J. van Marie (eds.), Yearbook of Morphology, Vol. 3, 201-206. Foris Publications, Dordrecht.

A DEDUCTIVE ACCOUNT OF FRENCH OBJECT CLITICS ESTHER KRAAK Utrecht Institute of Linguistics OTS Utrecht University Utrecht, The Netherlands

1. INTRODUCTION

Over the past few years, there has been a development in the field of computational linguistics which has led to the emergence of a so-called "grammar logic." In this framework, insights from two logical traditions have been combined: Linear Logic (Girard, 1987) has provided the notion of a resource-conscious logic, where each assumption can and must be used only once, while from the categorial line of work initiated by Lambek (1958) has come the idea of structuring the assumptions in various ways. A detailed account of the grammar logic program can be found in Moortgat (1997). The purpose of the grammar logic is to provide us with a set of constants (connectives) of grammatical reasoning which are thought to underlie our linguistic abilities. Using these constants, analyses of a wide range of phenomena from a variety of languages can be construed in an entirely deductive way. In this article we present a complex test case for the framework, in the form of the intricate data on object clitics in French. The structure of the article is as follows: In section 2 we introduce the base logic on which our grammar fragment will be built. Section 3 then gives a short overview of the empirical facts which the fragment intends to cover, and in section 4 the base logic is shown to be adequate for analyzing Syntax and Semantics, Volume 30 Complex Predicates in Nonderivational Syntax

271

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

272

Esther Kraak

the most elementary of these constructions. In sections 5 and 6 the grammar is further refined, to deal with the more complex data concerning order requirements for multiple clitics and interactions between clitic pronouns and complex verbal constructions. Finally, in section 7 the approach presented here is compared with other analyses of the same phenomena.

2. CATEGORIAL LOGICS This section starts with an introduction to the fundamentals of categorial logic, which constitutes the basis of the grammar fragment that will be developed. In section 2.2, two key features of this basic formalism are presented that will play an important role in the analysis of French object clitic behavior: the use of unary control operators and of multiple modes of linguistic composition. To conclude, section 2.3 shows the working of the base logic with an example. 2.1. The Base Logic Consider the language ^"of category formulas, obtained by closing a set of atomic types s/ (the basic categories or basic types) under binary connectives /, •, \:

In words, a category A*B represents the composition of an element of type A with an element of type B. The slash can be regarded as the leftward- or rightward-looking "dual" of the product operator •: categories A/B and B\A express incompleteness with respect to composition. Elements of those categories need to be composed with an element of type B, on the right- or left-hand side, respectively, to yield an expression of type A. Because modal logic is ideally suited for reasoning about structures, in our case structured configurations of pieces of linguistic information, we interpret these category formulas in a Kripke style modal frame (W, R3y. Here, W represents the set of linguistic resources or signs, i.e., units of linguistic form and meaning which belong to a certain syntactic category and represent the words and phrases of a language. The ternary accessibility relation R3 in this frame represents linguistic composition, R3zxy meaning that the signs x and y can be composed to yield z. Adding a valuation v to the frame turns it into a model. This valuation takes

French Object Clitics

273

atomic categories to subsets of W (namely, those subsets containing all words and phrases of that category) and satisfies the following in case of compound formulas:

Intuitively, the relation R3 holds between two parts (x and v) and a whole (z) composed of those parts. In case of a slash category, one of the parts is unsaturated: in order to form a whole z of type C, an unsaturated part x of type C/B (resp., y of type A\C) must be composed with another part y to its right (resp., x to its left) which belongs to its argument category B (resp., A). To characterize the derivability properties of the categoryforming operators • and /, \, we will make use of the concept of residuatlon. Let j/ = (A, and 38 = B and g: B -» A is called residuated if the following holds:

For our purpose, the sets j/ and £% are the same: A and B represent the set of category formulas ^ and in both cases the ordering is the derivability relation ->. If we assume /= — *B and g = —/B, with some fixed B e ^ and take x and y in the definition of residuation to be A and C (e &~~), respectively, the above biconditional gives us

Analogously, if we assume f = A* — and g = A\- , with some fixed A e ^ and take x and y to be B and C (e &*), respectively, we get:

Combining these results yields the following laws of residuation, defining the derivability properties of the product and its left- and right-residual operators, the slashes:

The essential soundness and completeness result states that A -> B iff v(A) c u(B), for all valuations on all frames. These residuation laws, together with the inherent characteristics of the derivability relation, reflexivity (identity) and transitivity, constitute the

274

Esther Kraak

grammatical base logic. It is known as NL for nonassodatiue Lambek calculus, since it does not include the structural postulate of associativity, nor any other restriction. From this pure logic of residuation a whole landscape of categorial logics can be obtained, by relaxing its sensitivity for linguistically relevant structural parameters such as immediate dominance (constituent sensitivity) and linear precedence (order sensitivity). One way this can be done is by adding the structural postulates of Associativity and Commutativity to NL. Whereas in the base logic the interpretation of the accessibility relation R3 is completely free, adding these structural postulates introduces constraints on its interpretation. Notice that the more restrictions we impose on R3, the less discriminating it becomes. Below are the structural postulates and frame conditions that allow the linguistic resources to be restructured and reordered (permuted), respectively (for all u, x, y, z e W): Associativity A*(B*C) **(A • £)• C 3t.R3zxt & R3tyu **lu.R3uxy & R3zvu Commutativity A • B *+B*A R3zxy **R3zyx

Adding Associativity to NL gives us the Lambek system L; adding Commutativity to NL yields NLP (P for Permutation). If we adopt both postulates we get the system LP, which is also known as the Lambek-van Benthem calculus. Throughout this paper we will be using a system of natural deduction1 to present derivations of linguistic constructions; a successful derivation is called a proof. The things we want to prove are statements of the form F \- A, pronounced 'T derives A," where F is a structured configuration of category formulas and A is a single category formula. For each of the logical connectives /, •, \ there are two rules: an Introduction rule that introduces the connective to the right of the turnstile, and an Elimination rule for the removal of the connective.2 In the natural deduction rules below, the category formulas are annotated with semantic (lambda) terms, following the Curry-Howard correspondence between natural deduction proofs and terms of the A-calculus (note that Elimination corresponds to functional application, and Introduction to functional abstraction):

French Object Clitics

275

For instance, a rule such as I/ should be read as follows: if a structure F composed with a category formula B to its right-hand side derives a category formula A, then that structure F alone derives a complex category formula A/B. As we remember from the residuation laws, something is of category A / B if combining it with something of category B immediately to its right yields something of category A. This is expressed by the E/ rule, a version of the well-known rule of Modus Ponens in classical logic. It states that if some structure F derives a category A/B, and if some other structure A derives a category B, then the composition of these two structures F and A derives a category A. It is important to notice the parallellism between the logical product operator • to the right of h- and the structural product operator ° to the left. The structural connective builds a compound configuration of grammatical signs from more elementary configurations, whereas the logical operator expresses the type-logical properties of these configurations. As a simple example, we take the string of words "Mary loves Paul" and see if it derives an s, the category of a sentence. The top of the derivation (corresponding in a sense to the lexical entries) shows both "Mary" and "Paul" to be of category NP, and the verb "loves" of category (NP\S)/NP, meaning that "loves" yields a sentence if it is first combined with an NP (the object) to its right, and then with another NP (the subject) to its left. For ease of presentation, we use structural labels in the antecedent, and type formulas in the succedent:

Because the type of the verb indicates that it needs to be combined with its object first, we start by applying the E/ rule to "loves" and "Paul." It composes the words to form the string "loves Paul," which is of category NP\S—compare this category to the subcategorization list of a transitive verb, from which the object has been removed. The newly constructed phrase is not fully saturated yet, as it is still looking for a subject to its left. This is taken care of by the E\ rule, composing "Mary" and "loves Paul," which results in the string "Mary loves Paul" of the desired category s. The square brackets reflect the bracketing in the category of the verb, in that the verb and its object have been combined first, before the subject was added (the outermost brackets are dropped to improve readability). In a system where Associativity is assumed, the brackets can be omitted because all possible bracketings of a string can be generated, which are all

276

Esther Kraak

equivalent. We, however, assume a nonassociative base system, sensitive to constituent structure, for which the use of brackets is essential. As we already mentioned, structural rules of Associativity (rebracketing) and Comrnutativity (permutation) can be added to the nonassociative system NL, which will destroy the system's potential to discriminate between different structurings (dominance) and orderings (linear precedence) of the linguistic material. In natural deduction format, these rules assume the following form, where the double horizontal line indicates premise and conclusion can be interchanged, and F{A} means there is a substructure A somewhere in T:

By adding these structural rules to our system, we would not only be able to derive strings such as Mary loves Paul, but also Loves Mary Paul or Paul Mary loves as being of category s. In some cases this kind of freedom of constituent structure and/or word order can be desirable. For instance, we would like both Yesterday it was raining and It was raining yesterday to be derivable (with the same type assignment to "yesterday"}, which is not the case in (N)L. But at the same time we want to avoid derivability of such ungrammatical strings as Was yesterday it raining or Loves Mary Paul that can be derived in (N)LP. Simply assuming the structural rule of Commutativity causes our system to lose its order sensitivity completely, leaving us with no means to distinguish between grammatical and ungrammatical word order. The same holds, mutatis mutandis, for Associativity and structure sensitivity. We will see in the sections to come that it is nevertheless possible to combine different levels of structure- and order-sensitivity in one framework, by adopting a multimodal architecture which allows more fine-structuring of the notion of linguistic composition. Crucially, it will give us explicit control over the structural management of linguistic resources, instead of having the structural possibilities hard-wired into the system. 2.2. Structural Control In the previous section we have seen, in a nutshell, that although each of the Lambek systems in the structural landscape has its merits, no single one of them is capable of handling all the facts of linguistic analysis on its own. Their shortcomings can be remedied in two different ways, both providing control over the structural aspects of resource management: by

French Object Clitics

277

extending the framework with unary control operators that license or constrain structural operations on the resources (cf. Hepple, 1990; Kurtonina and Moortgat, 1996; Moortgat, 1997; Morrill, 1994; Morrill and Gavarro, 1992; Morrill et al., 1990; Nishida, 1991; Stabler, 1996; Versmissen, 1996) and by introducing a multiplicity of composition modes which all have different resource-management properties hard-wired into them (cf. Moortgat and Oehrle, 1994, 1996). 2.2.1. Unary Control Operators The central concept of residuation can be generalized to the case of ft-ary operators. The language of binary operators is extended by Kurtonina and Moortgat (1996) with a pair of unary residual operators O (diamond) and D* (box), for which the definition of residuation (fie < y iff x < gy) takes the following form:

We can see how this instantiates the general definition if we assume that O = / and D = g, and substitute A for x and B for y. By putting together the unary and binary category-forming operators we obtain a "mixed" language ^ defined inductively from a set of atomic categories j/ as

Category formulas of this language can be interpreted in terms of a mixed Kripke frame {W, R2, R2\ where the set of possible worlds W represents the set of linguistic resources, and linguistic composition is modeled by binary and ternary accessibility relations R2 and R3. The frame is turned into a model by adding a valuation v for the new unary operators (compare with that for the binary operators):

By analogy with the ternary relation .R3, we can think of binary R2 as a relation composing a whole, z, but this time out of a single part, x. It is

278

Esther Kraak

important to note that the valuation for D; is not the same as that for the box (necessity operator) in modal logic; the downarrow on the box is indicative of this difference. The accessibility relation is visualized in the diagram below, where the first and second graphs represent the relation R2 interpreting the unary category-forming operators O and D^, and the third graph represents the relation R^L interpreting the modal logic necessity operator D, given here for comparison. The diagram clearly shows the operators O and nl to be duals, as the composition relation R2 moves in opposite directions: whereas OA is the whole, composed by R2 from the part A, D^4 is the part which R2 composes to form the whole A:

The unary operators will serve as "features" that control the structural aspects (order and constituency) of linguistic composition. They can either be employed to constrain the use of permutation and rebracketing in the less discriminating logics or to license restricted use of these structural operations in the more discriminating ones. In this they show great similarity to the ! and ? operators of Linear Logic. The control features can be seen as consisting of a "question" (n 1 ) and an "answer" (O) part that cancel each other out, as expressed by the basic law Oni4 -> A (derivable in one residuation step from n^A -> \31A. What this law means, intuitively, is that a feature must be explicitly checked before it can be consumed in a grammatical deduction. In this sense the law bears a strong resemblance to the principles governing the checking of formal features that have been proposed by Stabler (1996). Rules for introducing and eliminating the diamond and box operator in natural deduction are given below.3 Analogous to the product operator, the diamond has a structural counterpart on the left-hand side of the turnstile, written here in the form of angled brackets { ). The antecedent (F> stands for a structure F which has this structural ( > connective as its main (outermost) operator.

French Object Clitics

279

It is easy to see that the rules for the box correspond to the residuation laws at beginning of section 2.2.1. The IO rule is not hard to understand either: if a structure F derives a category A, then a diamond can be introduced as main operator on both F (the structural diamond) and A (the logical diamond). The elimination rule for the diamond is somewhat more complex, but it is in fact a rule of substitution (like the elimination rule for the product; cf. the Appendix): the structural operator { ) can be replaced by the logical operator O. Kurtonina and Moortgat (1996) develop a theory of systematic communication between the systems NL, L, NLP, and LP on the basis of the diamond and box operators licensing and/or constraining the use of structural operations of Associativity and Commutativity. Through a series of embedding theorems, they show that it is possible to entirely recover the discriminatory potential of a weaker logic in a system which is less sensitive to the structural management of resources, and, in the opposite direction, that the flexibility of a stronger logic can be reintroduced in a system with a more fine-grained notion of structure-sensitivity. Like the binary connectives /, •, \ the unary operators O and D; are defined in the pure logic of residuation: there are no restrictions on the binary accessibility relation R2 interpreting the diamond and box.4 This base logic could be extended with postulates imposing constraints on R2, such as reflexivity and transitivity axioms (just as adding associativity and commutativity postulates would restrict R3). However, a logic based solely on residuation is appropriate for the unary operators used in the analysis to be presented: it will be shown in section 6.1 that a logic containing a reflexivity axiom would produce undesirable derivability results. Applications for unary control operators can be found in a variety of linguistic phenomena, for example in the realm of the locality domains that Hepple (1990) and Morrill (1994) have tried to account for with their domain modalities and bracket operators. Versmissen (1996) uses unary operators to give a categorial account of word order domains, and in the next section we will see a concrete example of the use of the diamond and box operators in an analysis of Dutch verb raising, following Moortgat and Oehrle (1996).

280

Esther Kraak

2.2.2. Multiple Forms of Linguistic Composition The second direction in which Lambek systems can be extended is by moving from mixed (W, R2, R3} frames, where connectives are interpreted in terms of a single form of linguistic composition, to multimodal frames (W, R2, R?), with i, j ranging over sets of indices J^ f. Each relation Ri j models a different way of composing the linguistic resources, and therefore the i e J^and j e^are referred to as the resource management modes. At the end of section 2.1 we already touched upon this idea of multimodality, which we will now present more formally. Multimodal systems contain a variety of (unary and binary) logical connectives • and •;. For each of these logical connectives there is a matching structure-building operator < >;- or 0;. As in the unimodal setting, each multimodal operator is related to its corresponding residual(s) by laws of residuation now taking the form

The indices i, j range over sets of modes in which the linguistic resources (words, phrases) can be combined, such as commutative mode, associative mode, etcetera. With each comes a mode-specific rule defining the behavior of the operator in that mode. For example, the commutative product •co is accompanied by the postulate A »co B -> B »co A, stating that if two signs are put together in commutative mode, reversal of their order is allowed. The multimodal language & can be inductively defined in a straightforward manner, and a valuation for the category formulas in the multimodal language can be given in terms of accessibility relations R2 and R^, with /,;' variables over the diverse modes of linguistic composition. The natural deduction rules for the multimodal connectives are modified, to ensure that the resource management modes are respected when introducing or eliminating any of the operators. They can be found in the Appendix. Addition of frame conditions linking the different .R, and Rj will enable the various logics in a multimodal system to interact with each other. These linking principles give the system a much greater control over resource management, making it more powerful than the sum of its parts. Linking of the different families can be realized in two ways: through Inclusion and Interaction principles. Inclusion principles order the different resource management modes. A certain mode i is ordered before another mode k by the following frame condition, with the corresponding Inclusion axiom and natural deduction

French Object Clitics

281

rule (for all x, y, z e W}\

Interaction principles mix the distinct modes. Two kinds of interaction can be distinguished: interactions between a binary Rj and a ternary Rif and interactions between two different Rt that are both ternary. Instances of the first type of interaction are the strong and weak distributivity principles K, Kl, K2 that distribute unary O over binary •. In a multimodal system, these principles can be restricted to specific mode interactions by defining the frame conditions for specific modes i,j. In axiomatic form, they will look like this:

Communication between two binary families can be established through the weak distributivity principles of Mixed Commutativity and Mixed Associativity, that can only apply in situations where the accessibility relations are in different modes (/ ¥= k). They are given here as postulates:

Whereas in the classical systems (NL, L, NLP, and LP) the structural rules of Associativity and Commutativity are either always available (hard-wired into the system) or never, in the multimodal setting these systems are put together in such a way that "substmctural travel" between them is possible. This means that elements of a logic with less discriminative power, say LP, can be imported into a more sensitive logic such as NL, through the mixed rules that define interactions between the various modes of linguistic composition. As this traveling is not allowed in general, but is restricted to situations where specific modes i and k are in construction, collapse of the more discriminating into the less discriminating logic is avoided. In other words, the interaction principles provide us with the structural flexibility needed to handle certain linguistic phenomena, without creating the problems of overgeneration that would ensue from adopting the unimodal rules of Associativity and Commutativity.

282

Esther Kraak

2.3. An Example: Dutch Verb Raising We conclude this section with an example taken from Moortgat and Oehrle (1996), showing how the multimodal framework enables us to deal with constructions that cannot be handled in a unimodal setting. We have chosen this example because the analysis it presents shows a strong resemblance to the one we will give for French object clitics in sections 4-6. Consider the following Dutch (subordinate clause) verb raising construction: (Jan weet dat Piet) Marie wil (Jan knows that Piet) Marie wants NP

plagen. to tease

VP/INF NP\INF

'Jan knows that Piet wants to tease Marie.' Given the type-assignments above, the grammatical string of words Marie wil plagen does not derive a VP in an order-sensitive system such as NL: the auxiliary wil is in between the infinitive plagen and its argument Marie, thus blocking application of the \E rule. On the other hand the string wil Marie plagen, which is ungrammatical as an embedded verb phrase, does derive a VP. In a multimodal system these problems of under- and overgeneration can be solved. In Moortgat and Oehrle (1996), two binary modes are assumed: normal phrasal composition mode ^, used for combining for instance a verb with its noun phrase arguments; and head adjunction mode »0, for putting together two elements that are both heads. These modes interact via the principles of Mixed Commutativity (MC) and Mixed Associativity (MA) below, which define the head adjunction product to take scope over the phrasal composition product: after application of these rules the »0 operator ends up governing the configuration with the •l operator.

The words of the string above are assigned to the categories NP, VP/OINF, and NEWEST, respectively: the infinitive combines with its NP argument in phrasal composition mode, whereas the auxiliary and the infinitive are linked together using the stronger of the two modes. The MC rule now enables derivation of Marie wil plagen as a VP, because the NP and the auxiliary can swap positions:

French Object Clitics

283

However, we are still left with the derivability of the ungrammatical string, and this is where the unary operators come into play. Moortgat and Oehrle introduce two unary operators, C^ and O0, whose control behavior is restricted by the following inclusion and distributivity postulates:

It is important to note that these rules allow O1 to be distributed only over •1, and O0 only over «0. The linguistic intuition behind these rules is best understood if we regard O1 as checking a phrasal head, and OQ as checking a lexical head. Distribution principles then regulate the flow of head feature information through the structure.5 By equipping the categories of verbs in the lexicon with a marker for lexical headedness DO , which needs to be counterbalanced by a corresponding diamond O0 (remember that OD^ derives A), derivation of the ungrammatical wil Marie plagen is blocked:

Reading from the bottom up,6 the goal of the derivation is to determine whether the string wil Marie plagen derives a properly headed verb phrase, indicated by the box operator on the VP. The rule I lifts this box over the turnstile where it assumes the form of its residual, the structural diamond operator < }\. This phrasal composition diamond needs to be distributed over a head adjunction configuration, but the mismatch of modes blocks the use of rule K. Application of the inclusion principle causes the diamond to switch from phrasal composition to head adjunction mode. Now it can be strongly distributed over the product, thus identifying both substructures of the head adjunction as lexical heads. The elimination rule splits up the structure, leaving us with the lexical entry axiom (in one more step) for wil on the left and Marie plagen on the right. The diamond then must continue its nondeterministic search for the lexical head in the substructure Marie plagen. However, the operator is now in mode 0, for

284

Esther Kraak

which distribution over the phrasal composition product connecting Marie and plagen is not allowed. The inclusion principle being one-way only, switching the diamond back to mode 1 is not an option, and derivation fails at this point. To conclude, we present the derivation of the grammatical construction. It shows that Marie can be composed with the infinitive, by applying the Mixed Commutativity rule, but only after (again reading from the bottom upward) the headedness of the entire phrase has been checked via distribution of the diamond. In other words, as long as headedness has not been checked, phrasal composition cannot occur within a head adjunction configuration, for this would cause the distributivity problem we have just seen in the ungrammatical case.

If the semantically interesting part of this derivation (here just the elimination steps for the slashes, as only these correspond to an operationfunctional application-in the semantics) is annotated with semantic terms, we see the verb wil take correct scope over the application of plagen to Marie:

The basic grammar logic that has been presented in this section constitutes the starting point for the grammar fragment to be developed in sections

French Object Clitics

285

4-6. The following section gives a brief overview of the data which the fragment aims to cover.

3. DATA

In this section we list the facts about the behavior of French object clitics for which we want our grammar fragment to account. They concern the relations object clitics have with their full NP counterparts, with the verbs they lean on and with other clitics. Our analysis of these data, based on the multimodal framework as we have just seen it, will then be presented in sections 4-6. • Clitics are pronouns in their weak form. French pronouns can be classified into weak and strong forms. The first group, also known as clitics, includes the subject pronouns je, tu, il, elle, ce, on, nous, uous, Us, elles; the direct object pronouns me, te, le, la, se, nous, uous, les; the indirect object pronouns me, te, se, lui, nous, vous, leur; locative y and genitive or quantitative en; while the latter consists of the corresponding strong forms moi, toi, lui, elle, nous, vous, ewe, elles, and cela. Not all clitics have a corresponding strong pronoun, such as, for example, v, which replaces a combination of the preposition a and a normal NP. (1) a. Je le lui donnerai. ('I will give it to him.') b. Eux n'auraient pas fait cela. ('They would not have done that.') • Object clitics are in complementary distribution with strong pronouns and normal NPS. Weak object pronouns occur in positions that cannot be occupied by a normal object NP; conversely, object clitics cannot occupy NP positions: (2) a. b. c. d.

Marie *Marie Marie *Marie

les connait. ('Marie knows them.') mes amis connait. ('Marie knows my friends.') connait mes amis. connait les.

Apart from constructions like (3a), where the NP is added to a sentence (but separated from it by a pause) to emphasize the object in that sentence, a clitic pronoun can never occur simultaneously with its NP counterpart or with its strong pronoun variant: (3) a. Marie les connait, mes amis. ('Marie knows them, my friends.') b. *Marie le voit Jean. ('Marie sees him Jean.')

286

Esther Kraak

• Clitics cannot occur at all in the absence of a host. Unlike strong pronouns and full NFS, clitics need an adjacent, phonologically strong word to support them. Without such a host the sentence is ungrammatical: (4) a. b. c. d.

Qui as-tu vul ('Who did you see?') *Les. Ewe. ('Them. Them.') Ou as-tu etel ('Where have you been?') *Y. A la maison. ('There. At home.')

• Object clitics occur preverbally. Apart from their occurrence in positive imperatives, object clitics can only occupy preverbal positions, the verb serving as their host: (5) a. b. c. d.

Jean lew donne des cadeaux. ('Jean gives them presents.') *Jean donne lew des cadeaux. Donne-les-lui. ('Give them to him.') *Les lui donne.

• A clitic's verbal host cannot be a past participle. Although every clitic or combination of clitics needs an adjoining verb to lean on, not all verb forms are possible hosts for clitics. In French, verbs in their finite, infinitival, or present participle form (including the gerundive construction with en) can fulfill this role, whereas past participles never do: (6) a. Je lui parlerai demain. ('I will speak to him tomorrow.') b. Marie veut la lui montrer. ('Marie wants to show it to her.') c. Jean regarde Paul en Vecoutant. ('Jean looks at Paul while listening to him.') d. *Elle a les manges. ('She has eaten them.') e. Elle les a manges. • Nothing can intervene between object clitic and verb. Apart from another object clitic, no lexical material can intervene between the object clitic and its host verb: (7) a. b. c. d.

Jean ne Va pas vu. ('Jean has not seen him.') *Jean le n'a pas vu. Marie le lui dira. ('Marie will say it to him.') *Marie le lui souvent a dit. ('Marie has often said it to him.')

• Clitics cannot be conjoined. Unlike full NFS and strong pronouns, clitics cannot be conjoined, nor can they have wide scope over a conjunction of verbs: (8) a. *Jean le et la voit. ('Jean sees him and her.') b. *Jean le voit et entend. ('Jean sees and hears him.')

287

French Object Clitics

• Clitics can occur on a host of which they are not a complement. Some verbs allow a clitic to be attached to them, even though it is an argument of the embedded verb. This is the case with verbs expressing causation, such as laisser and faire, or perception, like voir, entendre, and others, when the understood subject of the embedded verb is a dative NP: (9) a. Marie laisse manger unepoire a Jean. ('Marie lets Jean eat a pear.') b. Marie la lui laisse manger. ('Marie lets him eat it.') In sentence (9b) the clitics are attached to the finite verb, although they are the (understood) subject and the object of the infinitive lower in the structure. However, if the embedded verb has an accusative NP as its subject, its object cannot be attached to the matrix verb: (10) a. b. c. d.

Marie laisse Jean manger une poire. ('Marie lets Jean eat a pear.') * Marie la laisse Jean manger. Marie le laisse manger une poire. *Marie le la laisse manger.

Most verbs do not allow the so-called clitic climbing of example (9b), leaving the clitic(s) attached to the infinitive: (11) a. Jean veut les manger. ('Jean wants to eat them.') b. Jean les veut manger. • Clitics appear in a fixed order. An important property of clitics is that their relative order is not subject to any variation. It is generally assumed that French pronominal clitics are realized according to a kind of "template," describing the order in which clitics (and negative particle ne) must occur: Position:

I

II

NOM

ne

III me te nous vous se

IV

V

VI

VII

SACC

3DAT

y

en

Here NOM represents the subject clitics; 3ACC the third person accusative clitics le, la, les; and BOAT the third person dative clitics lui and leur. Usually the following assumptions accompany this template: clitics must occur in the indicated order, there can be only one clitic per position, normally there are at most two object clitics, and in a cluster of clitics the positions III and V cannot be filled simultaneously. The following examples reflect these restrictions on order and co-occurrence of the clitic

288

Esther Kraak

pronouns: (12) a. b. c. d. e. f.

Je ne le lui donneraipas. ('I will not give it to him.') *Je ne lui le donnerai pas. Jean me Va dit. ('Jean has said it to me.') *Jean le m'a dit. *Elle te lui presentera. ('She will introduce you to him.') Je I'y ai vu. ('I saw him there.')

The phenomena described here will have to be accounted for in any adequate theory of clitic pronouns. We will see in the following sections that the data from examples (2)-(5) can be taken care of by assigning clitics to the proper category. Interaction between different modes, controlled by corresponding unary operators, will be responsible for correct treatment of the facts in examples (6)-(12).

4. SIMPLE CLITIC CONSTRUCTIONS If we want our grammar fragment to cover the data presented above, it is essential that we assign the clitics to the right category. In this section we will build the appropriate type step by step, motivating our choice as we go along: the use of a higher order type, of hypothetical reasoning, of features, resource management modes, and unary control operators will all be reviewed. 4.1. Assigning a Category: Stepwise Refinement In accordance with the way pronouns are standardly treated, we assign the object clitics to a higher order category. To capture the fact that (in nonimperative sentences, to which we restrict ourselves here) they are situated immediately before their host verb, we assign them to a functor category looking rightward for a verbal argument. From this verbal argument an NP must be missing, namely the one that has been cliticized. This incompleteness will, through hypothetical reasoning, block co-occurrence of a clitic with its corresponding full NP form, thus ensuring the complementary distribution of the pair (clitic, NP) mentioned in section 3 (examples (2) and (3)). As the NP that has been cliticized does not necessarily come from a position adjacent to the verb, we do not use the normal slash operator but the extraction1 operator T • An item of category A^B will yield something of category A if it is combined with an item of category B somewhere.

289

French Object Clitics

(Note that this makes the slash operator a special instance of the extraction operator.) An informal natural deduction rule for the introduction of this operator can be found in the Appendix. If we assume the verbal argument of a clitic to be of category VP, the provisional category to which we assign the clitics is then of the form VP/(VPtNP)

Because we need to distinguish between finite, infinitival, and past participle verb phrases, which will be represented in this article simply by VPy (abbreviating NP\S), VP,-, and vppp, respectively, the clitic's category must be specified for a certain type of host verb as well. In the clitic types we will use the general category VP^ for verb phrases, where X is a variable over {/, i, pp}. To make certain that the grammatical function of a clitic is the same as that of the NP it replaces, we add a subscript for case to the missing NP in the clitic's category. This means that clitics take as their argument a verb phrase which is incomplete for an NP y , with Y ranging over {a, d} for accusative and dative NFS. In the derivations that follow, the variables X and Y will be directly instantiated to their proper values.8 The phonological bond between clitics and their hosts is a strong one: we saw in example (7) that nothing can intervene between the two. Following Moortgat and Oehrle's analysis of Dutch verb raising (section 2.3), we assume two modes for putting together linguistic material: "normal" phrasal composition mode •c and a stronger mode, equivalent to their head adjunction, which we will call clitic attachment •ca. Likewise, we assume two modes for the unary operators, Op for phrasal heads and O, for lexical heads (corresponding to Moortgat and Oehrle's Ol and O0, respectively), and the following rules of inclusion and interaction (we will not use the rule MC):

In light of the evidence from example (7), we assume the main operator of the clitic category to be in strong clitic attachment mode, whereas the missing hypothetical NP in the argument subcategory is combined with the verb in standard phrasal composition mode. The verb, being the head of the sentence, is equipped in the lexicon with an/ operator identifying it as the lexical head. Because clitics form a strong cluster with the verb, they

290

Esther Kraak

too have D/ as the main (outermost) operator in their lexical category. The general form of the category to which clitics are assigned is

The lexical entry that we are assuming for a transitive verb like voit ('sees') is the following:

Now that the foundation of our categorial treatment of object clitics has been laid, we can give a simple example to illustrate the working of the system. Consider the sentence Jean la voit, 'Jean sees her.' Derivation of this sentence is as follows (we use VP^ as abbreviation for NP^\ C S):

The goal of this derivation is to find out whether the string of words Jean la voit is a properly headed sentence (category D^ s). Via the rules for distribution of the diamond and a switch from Op to O,, the clitic and the verb both end up (reading from the bottom upward) with a O, diamond, which is the correct one considering their category. If we start from the top, the verb voit first combines with a hypothetical accusative NP, the one that has been cliticized. This hypothesis is then withdrawn through application of the I|c rule, resulting in a VP category missing an accusative NP somewhere. By taking this complex category as its argument, the clitic binds the argument position "occupied" by the hypothetical NP, and in this way fulfills the role of direct object to the verb. Now it also becomes clear how co-occurrence of the two forms is blocked: if the verb is combined with a real NP first, it is no longer possible to construe the proper argument for the clitic, as there is no hypothesis to be withdrawn.

French Object Clitics

291

Looking at the derivation, the impression might arise that it is necessary to present the system with a string in which the correct operators and brackets are already present. It has been shown by Moot (1996), however, that this is not the case. He uses labeled proof nets to show that the configuration of the antecedent is "projected" automatically in the unfolding of the lexical type assignments. What we see is that (analogous to the Projection Principle, for instance) structural information originating from the lexical assignments is projected through the derivation.

4.2. Coordination As it stands, the multimodal framework can handle the aspects of object clitic behavior that were demonstrated in section 3 with examples (2)-(5): object clitics do not occupy normal object NP positions and are in complementary distribution with full object NFS; they occur directly in front of their verbal host and can never appear without it. An interesting consequence of the way in which the system works is that the right predictions about the conjoinability properties of object clitic pronouns follow automatically from it. Consider example (8), repeated here as (13): (13) a. *Jean le et la uoit. ('Jean sees him and her.') b. *Jean le uoit et entend. ('Jean sees and hears him.') From example (13a) we see that it is not possible to conjoin two object clitics; example (13b) shows that an object clitic cannot have scope over a coordination of two verbs. To reflect the fact that the conjunction et can be used to combine all kinds of categories, as long as they are identical,9 we assign it to the category schema below, where X is a variable ranging over the entire set of categories. The fact that the conjunctive et constitutes the head of a conjunction is indicated by the box control operator in lexical head mode /:

As with the variables in the category for clitics, the variable X will be instantiated, in each specific derivation, in accordance with the category of the elements to be conjoined. In example (13b), for instance, X would be substituted by the lexical type of a transitive verb. To see how exactly the ungrammatical cases mentioned above are ruled out, let us try to derive

292

Esther Kraak

the first sentence, (13a):

Derivation fails at the point where the diamond in lexical head mode needs to be distributed over the phrasal composition (mode c) product, which is not allowed. And since the < )/ diamond cannot be turned back into a < )p diamond (because of the unidirectionality of the inclusion principle I) there is no way in which the derivation can proceed. The same happens with the second sentence, where the clitic le takes wide scope over a coordination of two verbs:

The basic idea behind the system, which causes these constructions to be ruled out as ungrammatical, is that the prosodically strongest mode of linguistic composition (clitic attachment mode) does not allow any material using the weaker mode of phrasal composition to occur within its domain, as long as headedness has not been checked. This is precisely the reason why things go wrong in the attempted derivations above: in both cases a phrasal composition structure is contained within a clitic attachment configuration.

5. CLITIC ORDER

Now that we are able to deal with sentences containing one clitic pronoun, the grammar must be refined in such a way that, in cases where

French Object Clitics

293

there are multiple clitics, it allows only derivation of those sentences in which the clitic order is correct, i.e, corresponds to the order template we have seen following example (11). In its present form any combination and order of object clitics can be derived, which is clearly an undesirable situation. Since the unary operators already identify the clitic-verb cluster as the head of a sentence, they are ideally suited for conveying information about the order within that cluster. To this effect, we inscribe the diamond and box operator with a feature representing a position in the order template. We let this feature range over the integers 0 to 8, corresponding to the position numbers in the following order schema:

The order -< defined for the position features is as follows: 0 •< 1 -< 2 •< 3 x 4 x 5 x 6 x 7 x 8 and O x l x 2 x 3 ' x 6 x 7 x 8. What this means is that the first and second person accusative clitics me, te, nous, vous, with order feature 3', can never precede (nor follow) any of the other object clitics.10 They can, however, follow a nominative clitic and precede en and y and, evidently, the verb. The feature-inscribed unary operators will have the form j and H^ for the logical operators, and < >* for the structural counterpart of the diamond, where j ranges over the modes as before, and x ranges over the set of position features.11 Lexical entries for the third person accusative clitic la and the transitive verb uoit, for instance, now assume the form

Rules of introduction and elimination for the feature-inscribed unary operators remain the same, as do the inclusion rule and the weak distributivity rules Kl and K2. An important role in enforcing the correct clitic order will be played by the strong distribution rule K, as this rule distributes the diamonds within the clitic-verb cluster. We refine it as follows:

Together with its side condition, rule K states that, whenever a diamond in lexical head mode is distributed over a product in clitic attachment mode,

294

Esther Kraak

the left conjunct must get a diamond whose position feature strictly precedes that of the right conjunct. This will ensure that the left-to-right order of the clitics must always correspond to the order prescribed by the template, and that no two clitics with the same order feature can co-occur. Two derivations will illustrate the working of this rule. As a first example, consider the partial derivation of the grammatical sentence Jean la lui donne ('Jean gives it to her'):

The aim is to find out whether this particular string of words (composed as shown) derives something of category s which is properly headed and of which the clitic-verb cluster is well ordered. This last requirement is expressed by an order feature 0 in the box operator on the succedent category s. It can be regarded as a kind of start symbol for the clitic order, expressing the fact that any grammatical combination of clitics is still allowed to occur. After the box operator has changed into its residual diamond ( >° (via rule I Dp), has been distributed over the phrasal composition product to the right conjunct la lui donne (rule K2), and its mode has been switched from phrasal to lexical (rule I), the ( >/ diamond is distributed over the clitic attachment cluster via two applications of rule K. The first one distributes a diamond with order feature 4 to the left conjunct la, and a diamond with order feature 5 to the right conjunct lui donne, respecting the condition on K. In the second step this diamond with order feature 5 must be distributed over the clitic attachment product again, this time yielding a diamond with order feature 5 for the left conjunct lui, and one with order feature 8 for the right conjunct donne. We can tell that the order in the clitic-verb group is grammatical from the

French Object Clitics

295

top of the derivation: it shows the correct lexical entries for the object clitics and the verb. This is not the case with the ungrammatical—as a result of non-well-ordering—sentence *Jean lui la donne:

This derivation proceeds like the previous one, up to the point where rule K must be applied. The left conjunct of the outermost clitic attachment is lui, which gets a diamond with order feature 5. Because of the condition on rule K, the right conjunct la donne must get a diamond whose position feature follows 5, so for instance 6. The second application of K then distributes this diamond with order feature 6 to la and one with feature 8 to the verb donne.12 However, this distribution of diamonds does not result in a successful derivation: the lexical entry for la does not have E/ as its main operator, which is the outcome of the applications of rule K, but H/ instead. We could also have started the above derivation from the top, beginning with the correct lexical entry for la, but then the problem would arise at the point where rule K must be applied for the second time:

It is not possible for rule K to be used without violating the condition on its application, which says that the order feature on the first conjunct must

296

Esther Kraak

precede the order feature on the second. Since the left conjunct lui has an outermost diamond operator with order feature 5, and the diamond on the right conjunct la donne is inscribed with order feature 4, it is clear that the condition cannot be met and that failure of derivation must result. So we can see how the correct distribution of diamonds to clitics and verb is not deterministically chosen, but is in fact enforced by the interaction between rules and lexical entries: any wrong choice of diamond (with an order feature that does not correspond to the order feature in the category of the word) will bring about a failure of derivation. Only in those cases where the order within the clitic-verb group matches the order in the schema at the beginning of this section does our system produce a successful derivation, which is exactly what we aimed to achieve.

6. CLITICS AND VERBAL COMPLEXES Having shown the working of our system for simple clitic sentences, we will now consider cases in which the verb phrase consists of a finite verb followed by a past participle or an infinitive. 6.1. Modal and Perception Verbs Our first case concerns sentences that contain a clitic and a verbal complex consisting of a modal verb (for instance uouloir, devoir, aimer) and an infinitive. The following types are assigned to the infinitives:

Modal verbs never serve as hosts to clitics, for they do not allow clitic extraction out of the infinitival VP domain. This is evident from the following sentences: (14) a. Marie veut la manger. ('Marie wants to eat it.') b. *Marie la veut manger. In order to account for these data, the category for the modal verb must be such that the domainhood of its argument VP is respected. To this effect, the argument subcategory of the modal verb is equipped with a

French Object Clitics

297

complex of unary operators:

In words, what this combination of O and [Hip does is as follows: first, it should be noted that the diamond introduced here is of a different type than the unary operators used to convey information about headedness and clitic order. Crucially, the distribution rules do not apply to it, but otherwise it obeys the by now familiar rules of introduction and elimination. The function of this diamond is to impose a locality constraint13 on the infinitival VP argument of veut. It does so via the rule IO, that can only be applied if the entire antecedent structure is contained in unary brackets { >. If this is not the case, the diamonds on either side of the turnstile cannot be removed (reading from bottom to top) and derivation fails. After the domainhood of the argument VP has been successfully verified, GDp —just as in the other derivations we have seen—checks it for correct headedness and clitic order, through the introduction rule for the box and the distribution rule K. When every head in the infinitival verb phrase has been appropriately assigned a (structural) diamond operator with the correct order feature, all that remains is to determine whether the antecedent derives VP(, the category of an infinitival VP. In the following derivation, showing the grammaticality of the sentence Marie veut la manger, the clitic la will be instantiated as taking an infinitival argument (the rightmost part of the derivation is straightforward, as the reader can check):

We see that the structural diamond operator < > defining la manger as a closed-off domain is a necessary condition for derivational success. It is not conveniently put wherever it happens to be needed, but is entirely projected from the lexicon. Derivation of the ungrammatical *Marie la veut manger, where the direct object la has been cliticized outside the infinitival

298

Esther Kraak

domain of the embedded verb, is blocked as the hypothetical NP bound by the clitic cannot penetrate the diamond brackets around manger in order to be combined with the infinitive:

It is interesting to mention at this point that a logic with the modal reflexivity postulate A -» OA would enable us to derive the ungrammatical sentence above. By means of this postulate, the domain brackets ( > could simply be introduced around the structure [manger oc x] halfway up the deduction. Then it would become possible to apply rule IO, leading to a successful derivation. But, as we have seen, the reflexivity postulate is not included in the base logic for our control operators, so this problem of overgeneration does not occur in the present framework. Another group of verbs that do not allow arguments of the infinitival VP embedded under them to be cliticized comprises laisser and perception verbs such as uoir, entendre, ecouter, when these take an accusative NP argument which functions as the understood subject of the infinitive. Consider the sentences of example (10), repeated here with the verb uoit: (15) a. b. c. d.

Marie uoit Jean manger une poire. ('Marie sees Jean eat a pear.') *Marie la voit Jean manger. Marie le voit manger une poire. *Marie le la uoit manger.

The direct object une poire of the embedded verb manger cannot be cliticized onto the finite verb, as is shown by the ungrammaticality of sentences (15b) and (15d). However, cliticization of manger's subject Jean is allowed in sentence (15c). Note that the corresponding clitic le has accusative case, not nominative. In line with proposals made by Dowty

French Object Clitics

299

(1985), we therefore take the understood subject of the infinitive to be an accusative argument of the matrix verb. These considerations lead us to assign this group of verbs to the following category:

The argument VP constitutes a closed domain as in the case of the modal verbs, thus not allowing extraction of any vp-internal object. Since the understood subject of the embedded verb is not vp-internal but belongs to the matrix verb, it can freely be cliticized onto that verb, which is shown in the following derivation:

An attempt to derive sentences (15b, d) will result in failure for exactly the same reason as the derivation of *Marie la veut manger earlier in this section did: at a certain point we are left with the subgoal (manger) oc x \OMp VP,-. As the label x of the hypothetical accusative NP is unable to enter manger's domain, the sequent fails to meet the criterion for applicability of rule IO, demanding that the entire antecedent be contained in unary brackets. Thus the types for voit and veut correctly imply the ungrammaticality of sentences like (14b) and (15b, d) above. With the category for the perception verbs and laisser we can also derive instances of so-called "downstairs" cliticization where the clitics binding arguments of the infinitive are not extracted out of the VP,

300

Esther Kraak

domain: (16) a. Marie laisse Jean la manger. ('Marie lets Jean eat it.') b. Marie le laisse la manger. Because the clitic la is strongly attached to the infinitive, it remains within that verb's domain. This means that the entire VP argument of the matrix verb is contained within unary domain brackets, and therefore that the problem we encountered while attempting to derive *Marie la veut manger does not arise. Derivation of example (16b) for instance proceeds exactly like that of sentence (15c), if we replace (manger oc une poire) by (la oca manger). After successful checking of the antecedent's domain status and removal of the domain brackets, we are left to prove the goal [la oca manger] h- GD^ VP,-, an exercise we will again leave to the reader. 6.2. Clitic Climbing Finally, we consider cases of so-called "clitic climbing," in which clitics attach themselves to the matrix verb although they are complements of some verb embedded under it. The first major group of verbs exhibiting this phenomenon consists of the tense auxiliaries avoir and etre, which serve as hosts for clitics belonging to the past participle. Clitic climbing is obligatory, as can be seen from sentences in which clitics appear on the past participle: they are always ungrammatical (cf. (6d, e)). (17) a. Marie Va mange. ('Marie ate it.') b. *Marie a le mange. We assign the past participles of verbs like dormir, manger, and donner to the types given below:

Cliticization onto the tense auxiliary is in clitic attachment mode as before. We do not assume the auxiliary to form an equally strong cluster with the participle, since elements like pas and peut-etre can intervene between the two. However, to combine tense auxiliary and past participle in phrasal composition mode would not reflect the intuition that these two verbs together do constitute a stronger type of complex, for clitics are not allowed to occur in between. They must climb over the entire verbal complex instead, as we saw in example (17b). This intuition can be

French Object Clitics

301

formalized by introducing a third mode of linguistic composition, which we will call clause union mode. The clause union product and its residuals, the slashes, will be subscripted «, and the following interaction rules are added to the system:

In words, the distribution rule Ku for the clause union mode states that a clause union configuration consists of two verbs—two lexical heads. For both conjuncts receive a diamond in mode /, indicating their lexical head status, and with order feature 8, which the reader will remember is the position of the verb in the order schema for the clitic-verb cluster. We will see below how this forces clitics to climb over the entire verbal complex formed by the clause union product. The Mixed Associativity postulate defines clause union mode to take scope over phrasal composition mode. The tense auxiliaries est and a which trigger clitic climbing are assigned to the following category, where the slash operator in mode u reflects the clause union between auxiliary and participle:

With this category we can derive a sentence like Marie Va mange. Derivation fails for ungrammatical cases such as *Marie a le mange where the clitic is attached to the participle, because both conjuncts of the clause union product must get a diamond with order feature 8, for the verb position. Consequently, the clitic will not receive the diamond it needs: because of the condition on rule K the only possible order feature that can be distributed to le is 8, which does not match the order feature 4 in the clitic's lexical box operator. In addition, according to that same rule the second conjunct mange should be given an order feature following 8, which is impossible as 8 is the last position in the schema.

302

Esther Kraak

The other group of verbs showing obligatory clitic climbing contains faire, and the perception verbs and laisser insofar as these take a dative complement which functions as the understood subject of the embedded infinitival verb. For instance (see also (9a, b): (18) a. Marie fait manger une poire a Jean. ('Marie makes Jean eat a pear.') b. Marie la fait manger a Jean. c. Marie la lui fait manger. d. *Marie fait la manger a Jean. The clitic la corresponding to the (vp-internal) direct object of manger must attach itself to the front of the finite-infinitival verb cluster as in sentences (18b) and (18c); cliticization onto the infinitive is not allowed, as can be seen from the ungrammaticality of (18d). The fact that clitic climbing is obligatory in these cases is indicative of a clause union relation between the two verbs. We therefore assign this group of clitic climbing triggers to the following category:

As with the versions of the perception verbs and laisser that host cliticization of their vp,-external accusative NP argument, here the dative NP "subject" of the embedded verb cliticizes onto the main verb, of which it is a complement. But unlike the former verbs, clitic climbing verbs such as faire allow the vprinternal NP to be extracted from its domain and cliticized onto the triggering main verb, as in examples (18b, c). The category assumed here for the clitic climbing verbs produces the desired derivational results for these constructions. That the clause union mode does indeed force a clitic to climb, instead of allowing it to occupy a downstairs position, becomes clear when we try to derive (18d), *Marie fait la manger a Jean. At a certain point in the derivation, the diamond in lexical head mode must be distributed over the structure [fait ou [la oca manger]]. Following rule Ku, both conjuncts fait and la manger receive a diamond with order feature 8. But now we can see that we are in the same situation as before when we attempted to derive the ungrammatical sentence (lib), * Marie a le mange: the clitic cannot get a diamond carrying the order feature 4 which it needs in accordance with its lexical category, as such a distribution would violate the condition on rule K. So, with this last refinement of the framework, all the facts about the behavior of French object clitics presented in section 3 can now be accounted for by our grammar fragment.

303

French Object Clitics

7. COMPARISON WITH OTHER ANALYSES In this section, the analysis of preceding pages is compared with and Head-Driven Phrase Structure HPSG analysis of French clitics in

French object clitics presented in the other analyses in Categorial Grammar Grammar. In particular, we look at the Miller and Sag (1995).14

7.1. Clitics as Pronominal Affixes Following the argumentation of Miller (1991), who applies the criteria for distinguishing affixation from postlexical cliticization (given by (Zwicky and Pullum, 1983)) to the case of clitics in French, Miller and Sag, (1997) consider clitics to be pronominal affixes instead of independent syntactic words. They propose a lexical rule of pronominal affixation that turns a word into a cliticized word by removing one or more complements from the word's COMPS list, thereby reducing its valence. These complements must be of type afftixal-synsem). Since complements are identified with members of the ARG-ST list, a cliticized word will have one or more elements of type aff in ARG-ST. Morphological realization of the pronominal affixes is then triggered by the presence of these aff-type elements. Thus the rule guarantees complementary distribution of clitics and their full NP counterparts. Unlike Miller and Sag, we see no reason to draw a firm distinction between syntax and morphology. Rather, we like to think of the relation between the two as a sliding scale. This is not to claim that the distinction does not matter. Instead of embracing a strictly modular view, however, we take it that syntax and morphology are intricately intertwined. In our analysis this is expressed by the coexistence of different modes of linguistic composition, reflecting varying strength of prosodic attachment. The strongest mode, connecting clitics with their host, occupies a position nearer the "morphological side" of the scale, whereas the weakest mode, combining verbs and their arguments, is situated on the "syntactic side." Since we consider clitics as having their own lexical entries, this variation in compositional strength can be entirely projected from the lexicon. It is also worth noting that the complementary distribution of a clitic and its corresponding full NP, for which Miller and Sag need the abovementioned lexical rule, follows in our account from the interaction between the clitic's higher order type and the general mechanism of hypothetical reasoning triggered by that type. As the clitic binds the hypothetical position of the noun phrase it replaces, it will automatically be identified with that missing NP, thus blocking their simultaneous occur-

304

Esther Kraak

rence. It is important to see that in our view it is not the verb which selects for a "clitic" argument, as is the case in Miller and Sag's lexical rule. Instead, it is the clitic which serves as the functor, selecting a verb that is unsaturated for the NP complement whose role is fulfilled by the clitic. Combining the clitic with its verb argument (binding the hypothetical argument position) yields a verb whose valence is reduced with that same NP complement. Since we work in a resource-conscious logic where each assumption can and must be used only once, there is always exactly one clitic that binds an argument position. Additional constraints enforcing this uniqueness, as needed by Miller and Sag in order to prevent affixal haplology, are therefore superfluous.

7.2. Clitic Order

Miller and Sag treat the restrictions on order and co-occurrence of clitics within the clitic cluster by characterizing it as an ordered sequence of slots. In addition, they assume a substantial number of cluster conditions, stating for each of the slots whether and how it is to be filled, depending on the presence of certain elements on the ARG-ST list of the cliticized word. This is in line with their view of cliticization as an entirely morphological phenomenon. As we mentioned earlier, our account of cliticization does not embrace this modular view. Instead, the make-up of the clitic cluster results from the interaction of lexical type-assignments for the clitics with a language-specific order schema and generally valid deductive principles. In this sense, our position is more in line with work by Kathol (1995), who analyzes word order in terms of so-called topological fields. Sentences are regarded as consisting of an ordered sequence of fields or domains, in each of which only certain types of constituents occur. Every word is explicitly marked in the lexicon with a feature for the specific field(s) where its occurrence is allowed. This corresponds to our representation of clitics as lexical items marked for a specific position within the clitic cluster15 whereas in Miller and Sag's view it is the ARG-ST list of the clitic host which constitutes the locus of information relevant to clitic ordering. Another (categorial) treatment of cliticization which makes use of order features is that of Morrill and Gavarro (1992), who propose an account of Catalan clitics similar to the analysis put forward in the present article. They treat clitic order with a feature for cliticization class. This feature, in the form of an integer, is connected with both subtypes s in the higher order type which Morrill and Gavarro assume for the clitics. The idea is

305

French Object Clitics

that the value of this feature on the argument category of the clitic is higher than that of its result category. Thus, clitics that appear near the front of the clitic cluster will have lower values in their argument categories than those appearing closer to the verb. The verbs themselves, following the clitic cluster, have the highest value. The main difference with the analysis presented in this article is that, in the multimodal grammar logic, checking of these features is effected through deductive principles, whereas in Morrill and Gavarro's account it is done by means of unification.16

7.3. Clitic Climbing: Argument Composition The analysis of clitic climbing proposed by Miller and Sag17 is described by them as a case of argument composition: a clitic climbing trigger combines with an unsaturated verb and then with the arguments selected by that verb. In the lexical entry of such a trigger, the COMPS requirements of the verb argument are added to the trigger's own COMPS list, and their occurrences are identified via reentrancy. As Miller and Sag point out, the notion of argument composition is related to the concept of division type-shift in Categorial Grammar, which can be schematized as A/B ->

(A/C)/(B/C)

The analogy becomes clear if we take A/B to be the initial type of the clitic climbing trigger, B/C the type of its (unsaturated) verb argument, and C the complement of that verb argument. The higher order type for the trigger verb, on the right-hand side of the arrow, requires to be combined first with the unsaturated verb B/C, and after that with the verb's complement C. Accounts of clitic climbing based on a rule of division have been proposed in the field of Categorial Grammar by Casadio (1993) for Italian and Nishida (1991) for Spanish. However, these categorial analyses adopt a version of division valid only in systems which include the structural rule of Permutation, leading to overgeneration in the form of ungrammatical word order. A drawback of all analyses of clitic climbing founded on division is that, compared to the multimodal approach advocated in this article, they constitute a derivational detour. Why this is so can best be illustrated with an example. Consider the sentence Jean a vu Marie ('Jean saw Marie'), where the tense auxiliary a is a clitic climbing trigger. Deriving division in a proper (but unimodal) natural deduction setting, we arrive at the

306

Esther Kraak

following analysis for this sentence:

We see the division category for a being derived via hypothetical reasoning18 (the initial type and the higher order type are boxed for the sake of clarity). Then the newly obtained category is combined with unsaturated vu (its first argument, VPpp/NPa) and after that with the participle's complement Marie (its second argument, Npa). The immediate succession of an Introduction and an Elimination step acting on the same connective, in the middle of the derivation, indicates the possibility of simplification. Here, the difference between analyses based on rules (such as division) and deductive approaches like ours becomes clear: the former lack a notion of normalization of proofs, in the sense that the proofs they produce are not necessarily minimal derivations without any detours. In the deductive account, which does possess this normalization property, the steps needed above for the derivation of the division type can be pruned away, resulting in the following derivation:

French Object Clitics

307

As was shown in section 6.2, this analysis yields the desired results for the clitic climbing data. Among other things, it accounts in a principled way for the fact that, in French, cliticization on a past participle is not allowed. The analyses based on division mentioned above, on the other hand, need to stipulate application of this rule to be blocked specifically for verbs in past participial form. There is also another, more serious problem with the division-based approaches to clitic climbing, evidence of which is given in the first derivation of this section. In the process of constructing the higher order type, the structural rule of Associativity is needed to enable application of the I/ rule withdrawing the hypothetical NPa. Since the structural rules in their unimodal version are either always available or never, the consequence is that full Associativity must hold, with the consequent loss of constituent sensitivity. The multimodal analysis presented in the second derivation assumes restricted (mixed) Associativity only as an interaction principle, relating modes that are in themselves nonassociative, and therefore does not cause the system to collapse into constituent-insensitivity.

8. CONCLUSION It has been shown in this article that a deductive approach to cliticization is fruitful in that it throws new light on the "fine-structure" of grammatical resource management. Comparing the proposed multimodal approach with the resource management assumptions underlying competing analyses (both categorial and HPSG-style), two points have been made.19 First, from the point of view of derivational economy, argument composition approaches to clitic climbing are suboptimal: they represent detours that in the present deductive setting are eliminated as a result of proof-normalization. Second, the argument composition analyses discussed here violate linear order sensitivity (Casadio, 1993; Nishida, 1991) and/or constituent structure sensitivity (Miller and Sag, 1995; Monachesi, 1995). This means that, both within the unimodal categorial and the HPSG-based analyses, extra non-logical principles have to be imposed to constrain grammatical resource management. Such non-logical principles can be avoided as a result of the more delicate multimodal resource management regime advocated here. The framework used here for the analysis of French object clitics needs to be extended in a number of ways. I have already touched upon the idea

308

Esther Kraak

of adding a feature logic component to handle those features that were simply represented by indices in the above. The fragment can also be made to cover the rest of the French clitics: subject clitics, reflexives, en and y. Moreover, preliminary investigations into the realm of Italian object clitics strongly suggest that clitic systems from other Romance languages can be accounted for as well. Cross-linguistic variation can be captured by choosing different packages of resource management rules for the different languages. These three topics constitute my aims for future research.

9. APPENDIX: THE FRAMEWORK Logical Rules

French Object Clitics

309

Resource Management Rules

NOTES The natural deduction format is used here for ease of presentation, but it does not constitute a proper basis for proof search. However, equivalence of the natural deduction, axiomatic, and Gentzen sequent presentations has been proven (cf. Moortgat and Oehrle, 1996, and Moortgat, 1997). The latter has the property of decidability: it will tell us whether there is a derivation for a given string or not. Every possible step is examined, guaranteeing a complete search producing all possible derivations if there exist any, and a negative answer otherwise. 2 The product rules will not be needed here, as we will not make use of lexical categories containing the product operator. For the sake of completeness, however, they are included in the Appendix. The category formulas in these rules are not annotated with semantical terms, as the unary control operators are considered to be semantically inert. 4 In this respect they differ from the box operator proposed by Morrill et al. (1990), which requires a reflexive and transitive interpretation. 5 In this sense, they provide the deductive basis for the Head Feature Principle in GPSG and HPSG. However, the distribution principles are more flexible than the HFP in that they allow multiple headedness in head adjunction mode. This difference is essential for Moortgat and Oehrle's analysis of Dutch verb raising and, as we will see shortly, also for an adequate treatment of French objects clitics. Throughout this article we will use the term bottom up in the following sense: the bottom of a derivation is its conclusion, the goal to be proven, whereas the top is constituted by the premises—the lexical entry axioms. 7 Moortgat (1997) shows that this operator can be defined in the present def framework as follows: A\B =A/OE\B, where the unary operators give access to

310

Esther Kraak

modally controlled versions of the rules for Associativity and Commutativity, which look like this (the condition holds for each of the rules):

with some A, of the form (A) When the desired configuration has been reached, the diamond and box operators cancel each other out, such that OD^B can be used as a normal item of category B. 8 A topic for further research in this respect would be to see whether it is possible to combine the multimodal framework with a feature logic along the lines of Dorre and Manandhar (1995). 9 We restrict ourselves here to coordination of like types. 10 Note that this schema differs from the template following example (11) on two points. First, whereas the template does not allow combination of l,2DATwith 3 DAT clitics, our schema does. This leaves room for sentences such as Jean me lui semble fidele ('Jean seems to me to be faithful to her'), considered grammatical by a number of speakers. Second, in the schema we have separated 1,2ACC from SDAT clitics, blocking derivability of such sentences as *// me lui presente ('He presents me to her'). Because SACC clitics can co-occur with 3DAT clitics, they are positioned in the same branch of our order schema. As a consequence, the 1,2ACC clitics are separated from the 3ACC clitics, whereas the template does not prohibit simultaneous occupation of positions III and IV. As far as we know, however, there seems to be no evidence refuting our analysis. 1] It is also possible to express other features in the unary operators, such as case or verb form. We will not pursue this line of thought here, but see Versmissen (1996) for an implementation. 12 In sentences without clitics, the verb gets a diamond with the correct order feature 8 via a special rule (called 0; see the Appendix) which takes a diamond with order feature 0 to a diamond with order feature 8. The idea behind this rule is that, trivially, the empty clitic sequence is well ordered. 13 Comparable to the function of the "bracket" operators of Morrill (1994), who assumes a less restrictive logic, however. 14 Miller and Sag (1995) was the representative HPSG analysis at the time when contributions to the present volume were written. The reader should notice that in the revised analysis of Miller and Sag (1997)-summarized in Abeille et al. (this volume)-the use of lexical rules has been abandoned. 15 For a more general categorial implementation of the concept of word order domains, see Versmissen (1996). 16 There are other differences as well, concerning the unary operators Morrill and Gavarro introduce for the treatment of medial cliticization and clauselocality. The logic of these operators is less restrictive than the pure logic of residuation for our diamond and box operators. As was pointed out in section 6.1

French Object Clitics

311

(cf. also the end of section 2.2.1), such a less restrictive logic can lead to overgeneration. 17 And proposed by Monachesi (1995) for clitic climbing in Italian. 18 Note that lexical assignment of the trigger verb to the higher order type is not an option, as this would result in multiple type-assignments to enable selection of verbal arguments with different complement structures. ^EDITORS' NOTE: The editors of this volume have pointed out to the author that her criticism of HPSG-style analyses is based on some purported "translation" of these analyses into categorial grammar-something that the authors whom Kraak cites have never proposed. To the editors of this volume it, therefore, seems far from clear how HPSG analyses that employ argument-composition techniques are to be fit into this same scheme of evaluation as analyses stated in the framework of categorial grammar. After all, Head-Driven Phrase Structure Grammar constitutes a framework with radically different logical foundations and with radically different assumptions about the underlying grammar formalism and the grammatical constraints imposed on individual grammars.

REFERENCES Abeille, A., D. Godard, and I. A. Sag (1998). Two kinds of composition in French complex predicates In Syntax and Semantics (E. Hinrichs, A. Kathol, and T. Nakazawa, eds.), Vol. 30, 1-41, Academic Press, San Diego. Casadio, C. (1993). A Categorial Approach to Cliticization and Agreement in Italian. CLUEB, Bologna. Dorre, J., and S. Manandhar (1995). On constraint-based Lambek Calculi. In P. Blackburn and M. de Rijke (eds.), Specifying Syntactic Structures. CSLI Publications, Stanford, CA. Title appeared in 1997. Dowty, D. (1985). On recent analyses of the semantics of control. Linguistics and Philosophy 8, 291-331. Girard, J.-Y. (1987). Linear Logic. Theoretical Computer Science 50, 1-102. Hepple, M. (1990). The Grammar and Processing of Order and Dependency: A Categorial Approach. Ph.D. dissertation, Centre for Cognitive Science, University of Edinburgh. Kathol, A. (1995). Linearization-Based German Syntax. Ph.D. dissertation, Ohio State University. Kurtonina, N., and M. Moortgat (1995). Structural control. In P. Blackburn and M. de Rijke (eds.), Specifying Syntactic Structures. CSLI Publications, Stanford, CA. Title appeared in 1997. Lambek, J. (1958). The mathematics of sentence structure. American Mathematical Monthly 65, 154-170. Miller, P. H. (1991). Clitics and Constituents in Phrase Structure Grammar. Ph.D. dissertation, Universiteit Utrecht.

312

Esther Kraak

Miller, P. H., and I. A. Sag (1995). Une analyse lexicaliste des affixes pronominaux en Frangais. Revue Quebecoise de Linguistique 24(1), 135-171. Miller, P. H., and I. A. Sag (1997). French Clitic Movement without Clitics or Movement. Natural Language and Linguistic Theory 15, 573-639. Monachesi, P. (1995). A Grammar of Italian Clitics. Ph.D. dissertation, ITK/TILDIL Dissertation Series 1995-3, Tilburg University. Moortgat, M. (1996). Multimodal linguistic inference. Journal of Logic, Language and Information, 5(3-4), 349-385. (Also in R. Kempson (ed.), Bulletin of the IGPL 3(2,3), 371-401, special issue on Deduction and Language.) Moortgat, M. (1997). Categorial type logics. In J. van Benthem and A. ter Meulen (eds.), Handbook of Logic and Language. Elsevier, Amsterdam. Moortgat, M., and G. Morrill (1991). Heads and Phrases. Type Calculus for Dependency and Constituent Structure. Unpublished manuscript, OTS, Utrecht. Moortgat, M., and R. Oehrle (1994). Adjacency, dependency and order. In Proceedings 9th Amsterdam Colloquium, 447-466. Moortgat, M., and R. Oehrle (1996). Structural abstractions. In V. M. Abrusci and C. Casadio (eds.), Proofs and Linguistic Categories. Applications of Logic to the Analysis and Implementation of Natural Language. Proceedings of the 1996 Roma Workshop, CLUEB, Bologna. Moot, R. (1996). Proof Nets and Labeling for Categorial Grammar Logics. Masters thesis, Universiteit Utrecht. Morrill, G. (1994). Type Logical Grammar. Kluwer, Dordrecht. Morrill, G., and A. Gavarro (1992). Catalan clitics. In A. Lecomte (ed.), Word Order in Categorial Grammar, 211-232, Editions Adosa, Clermont-Ferrand. Morrill, G., N. Leslie, M. Hepple, and G. Barry (1990). Categorial deductions and structural operations. In G. Barry and G. Morrill (eds.), Studies in Categorial Grammar. Edinburgh Working Papers in Cognitive Science, Vol. 5, 1-21, Centre for Cognitive Science, Edinburgh. Nishida, C. (1991). A non-transformational analysis of clitic climbing in Spanish. In Proceedings of WCCFL 9, 395-409. Stabler, E. P. (1996). Acquiring and Parsing Languages with Movement. Forthcoming. Versmissen, J. A. G. (1996). Grammatical Composition: Modes, Models, Modalities. Logical and Linguistic Aspects of Multimodal Categorial Grammars. Ph.D. dissertation, OTS Utrecht. Zwicky, A., and G. Pullum (1983). Cliticization vs. inflection: English n't. Language 59.3, 501-513.

ITALIAN RESTRUCTURING VERBS: A LEXICAL ANALYSIS PAOLA MONACHESI Utrecht Institute of Linguistics OTS Utrecht University Utrecht, The Netherlands

1. INTRODUCTION In Italian, there are some verbs which act as a class with respect to certain phenomena such as clitic climbing, long NP-movement, tough constructions, and auxiliary selection. They are usually called restructuring verbs after Rizzi (1982), who proposed a restructuring rule in order to account for their properties. In the first part of this article, I propose an analysis of these verbs in terms of argument composition. It is based on the idea that the arguments of the complement verb can be raised to become arguments of the restructuring verb. The analysis will be cast within the framework of Head-Driven Phrase Structure Grammar (HPSG) (Pollard and Sag, 1987, 1994) and will capture intuitions similar to those underlying Rizzi's Restructuring Rule (Rizzi, 1982) and the Clause Reduction analysis of Aissen and Perlmutter (1983). A comparison will be made with an alternative account based on the use of nonlocal features and the Nonlocal Feature Principle, which is the mechanism used within HPSG to account for Unbounded Dependency Constructions. However, I show that the argument composition analysis is indeed the most appropriate to provide a uniform account of the properties of the restructuring verbs. Syntax and Semantics, Volume 30 Complex Predicates in Nonderiuational Syntax

313

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

314

Paola Monachesi

The second part of the article addresses the issue of dialectal variation. I consider restructuring verbs in Salentino, a southern Italian dialect, and I focus on one of their properties: clitic climbing. In particular, I consider the behavior of the verb ulire 'want' in the variety of Salentino spoken in the provinces of Brindisi and Taranto. This verb is rather peculiar in that it allows clitic climbing out of a finite clause. I show that the argument composition analysis proposed for Standard Italian can account for this property and I suggest that within HPSG the parameterization of linguistic variation occurs in the lexicon.

2. BASIC FACTS

Rizzi (1982) identifies three classes of restructuring verbs in Italian: (1) Restructuring verbs • modal verbs (e.g., potere 'can,' dovere 'must,' volere 'want') • aspectual verbs (e.g., cominciare 'to begin,' finire 'to finish,' continuare 'to continue') • motion verbs (e.g., venire 'to come,' andare 'to go,' tornare 'to come back') As already mentioned, restructuring verbs exhibit certain properties, among which is clitic climbing. This is one of the most obvious diagnostic revealing complex predicate formation: (2) Anna lo vuole compare. Anna cl.(acc) wants to buy 'Anna wants to buy it.' In this construction, a clitic which originates as dependent of a complement verb can climb and attach to the trigger verb. Therefore, it seems that the two verbs act as a unit with respect to clitic placement and that what we have is a complex verb. In Standard Italian, clitic climbing is optional with these verbs; the following sentence where the clitic stays within the lower clause is also grammatical:1 (3) Anna vuole comprarlo. Anna wants to buy cl.(acc) 'Anna wants to buy it.'

Italian Restructuring Verbs

315

In Italian, there is a rather peculiar clitic, namely, the third person plural pronoun low. There are certain differences between low and the other clitics; however, they pattern alike with respect to climbing in the presence of restructuring verbs. In fact, these verbs also trigger the climbing of low 'to them,' as can be seen in the following example: (4) Non sapevano cos a potesse low capitare. not know what could to them happen 'They didn't know what could happen to them.' It should be noticed that low does not appear before the finite verb, as is the case for the monosyllabic clitic in (2). It is well known that morphosyntactic features determine the direction of attachment for monosyllabic clitics; they are proclitic if the verb is finite and enclitic if the verb is nonfinite or imperative. On the other hand, the placement of low is insensitive to the finiteness of the verb. In fact, this clitic usually follows the verb.2 As in the case of monosyllabic clitics, climbing of low is optional with these verbs. This is shown in the following example, where the clitic stays within the lower clause: (5) Non sapevano cosa potesse capitare low. not know what could happen to them 'They didn't know what could happen to them.' Restructuring verbs have also the property of triggering long NP-movement. If the clitic si, which acts as a passivizing element, is attached to the restructuring verb, the direct object of the verb in the embedded clause acts as subject of the restructuring verb: (6) Queste case si vogliono vendere a caw prezzo. these houses SI want to sell at high price 'These houses are wanted to be sold at a high price.' In this example, queste case originates as direct object of vendere and becomes the subject of si vogliono. Another property which is usually associated with this class of verbs is that they trigger auxiliary change. In fact, restructuring verbs which have avere as auxiliary can change it into essere if the embedded verb requires essere as auxiliary: (7) a. Rocco ha voluto partire. Rocco has wanted to leave 'Rocco has wanted to leave.' b. Rocco e voluto partire. Rocco is wanted to leave 'Rocco has wanted to leave.'

316

Paola Monachesi

Furthermore, in the presence of restructuring verbs, the Italian tough construction which is strictly bounded, appears to be unbounded: (8) Questa canzone e facile da cominciare a cantare. This song is easy to begin to sing In this case, the object of the lower verb questa canzone acts as superficial subject of the higher verb. 3. THE SYNTACTIC STRUCTURE OF RESTRUCTURING VERBS Before showing how an approach in terms of argument composition can account for the different properties of restructuring verbs, I will address the controversial issue of their syntactic structure. Rizzi (1982) proposed a rule of Restructuring to account for clitic climbing and for the other properties exhibited by restructuring verbs. He argued that sentences to which Restructuring has been applied have a different structure from those ones which have not undergone this rule. Restructuring turns a bisentential structure into a simple sentence. Therefore, if no clitic climbing occurs, Rizzi associates a bisentential structure to the sentence Anna vuole comprarlo given in example (3): (9) Anna vuole [s comprarlo]. While for a case of clitic climbing like Anna lo vuole comprare exemplified in (2), there will be a simple structure where the two verbs are reanalyzed into a verbal complex: (10)

Anna [v lo vuole comprare].

Rizzi considers clitic climbing as a sign that Restructuring has been triggered and he suggests that while there is motivation to let the embedded verb and its complements form a constituent if no clitic climbing occurs, this is not the case if clitic climbing is triggered. Rizzi gives four tests, on the basis of which he argues that the embedded verb and the following material do not form a constituent if clitic climbing is triggered. They are summarized here: • Pied-piping applies to an infinitive with following material, but not in a sentence where clitic climbing has occurred: (11) a. Questi argomenti, a parlarti dei quali verrb these topics, to talk cl.(dat) of which will come al piu presto, mi sembrano interessanti. very soon, cl.(dat) seem interesting 'These topics, to talk to you about which I will come as soon as possible, seem very interesting to me.'

Italian Restructuring Verbs

317

b. *Questi argomenti, a parlare del quali ti these topics to talk of which cl.(dat) verrb al piu presto... will come very soon... • Clefting can apply to an infinitive with following material, but not in a sentence where clitic climbing has occurred: (12) a. E proprio a riportargli i soldi che sto andando. is just to bring back cl.(dat) the money that am going 'It is really to bring him back the money that I am about to.' b. *E proprio a riportare i soldi che gli sto andando. is just to bring back the money that cl.(dat) am going • Right Node Raising (RNR) can apply to an infinitive with following material, but not in a sentence where clitic climbing has occurred:3 (13) a. Mario vorrebbe, ma a mio parere non potra mai, Mario would like but to my opinion not can never pagargli il debito. pay cl.(dat) the debt 'Mario would like, but in my opinion will never be able, to pay his debt to him.' b. * Mario gli vorrebbe, ma a mio parere non potra mai Mario cl.(dat) would like but to my opinion not can never pagare il debito. pay the debt • Complex NP shift can postpose an infinitive with following material, but not in a sentence where clitic climbing has occurred: (14) a. Fra qualche giomo verrb a Firenze ad esporti in few day will come to Florence to submit cl.(dat) la mia idea. the my idea 'In a few days, I will come to Florence to submit my idea to you.' b. *Fra qualche giomo ti verrb a Firenze ad esporre in few day cl.(dat) will come to Florence to submit la mia idea. the my idea The constituency tests proposed by Rizzi are based on the idea that constituents can move. The ungrammatical examples where clitic climbing has occurred show that the embedded verb and its complements do not form a constituent since they cannot move. Given an example such as (11),

318

Paola Monachesi

Rizzi considers the following structures as input to wh-mouement: (15) a. Questi argomenti [s verro a parlarti dei quali] alpiu presto. b. Questi argomenti [s ti verro a parlare dei quali] al piu presto. In the former case, the string a parlarti dei quali is a constituent and can be preposed, while the latter case shows an instance of clitic climbing. The string a parlare dei quali is not a constituent and it is for this reason that the ungrammatical (lib) is ruled out. If the embedded verb and its complements do not form a constituent, there are two possible configurations which are available in the case of clitic climbing: the main verb and the embedded verb can merge together in a sort of compound structure or there is a flat structure where the infinitival verb and its complements are sisters of the restructuring verb. Both structures are exemplified below, where the abbreviation CPL stands for complements:

As Rizzi points out, the choice between the two structures is quite difficult since there is not much clear evidence in favor of one or the other. However, it should be noticed that lexical material such as adverbs can intervene between the two verbs, arguing against a verbal compound: (17) Anna lo vuole immediatamente comprare. Anna cl.(acc) wants to immediately buy 'Anna wants to buy it immediately.' Therefore, I will assume that restructuring verbs are associated with the following structures:

Italian Restructuring Verbs

319

The flat structure will be associated with the clitic climbing configuration, while the hierarchical structure will represent those cases where the clitic remains attached to the lower verb.4 It should be noticed that Moore (1991) (a.o.) proposes instead that in Spanish, the embedded verb and the following material do form a VP constituent in cases of clitic climbing. He addresses the issue of the constituency tests and he says that not all of Rizzi's tests are possible in Spanish, but the clefting test is:5 (19) a. Lo que quiero es escribirles una carta. cl. that want is write cl. a letter 'What I want is to write them a letter.' b. *Lo que les quiero es escribir una carta. cl. that cl. want is write a letter The test produces the same result as in Italian, showing that the infinitival and the following material do not form a VP if clitic climbing has occurred. However, he argues against the clefting test, by saying that VPS may not cleft, since clefting may not strand an auxiliary: (20)

*Lo que ha fue llegado tardi. what he has was arrived late

This is also the case in Italian, but following Napoli (1981) it can be argued that this fact constitutes evidence for a flat structure also in the case of auxiliary verbs.6 He also shows that, in Spanish, right node raising is possible with upstairs cliticization. Furthermore, he discusses the coordination test, which is not mentioned by Rizzi; in Spanish it is possible to have a coordination of two VPS with a restructuring verb which has undergone clitic climbing: (21) Juan le quiere mandar un regalo y comprar un coche. Juan cl.(acc) wants send a gift and buy a car 'Juan wants to send her/him a gift and buy her/him a car.' This is also the case in Italian and it constitutes a potential problem for a flat structure: (22) Rocco le vuole spedire un regalo e scrivere una lettera. Rocco cl.(acc) wants to send a gift and write a letter 'Rocco wants to send a gift to her and write a letter to her.' Abeille and Godard (personal communication) suggest analyzing similar cases which occur with French auxiliaries as instances of Non Constituent Coordination. Their suggestion could be adopted also in the case of restructuring verbs; in fact, cases of Non Constituent Coordination are

320

Paola Monachesi

well attested in Italian: (23) Rocco scrivera una lettera a Martina e una cartolina a Vito. Rocco will write a letter to Martina and a postcard to Vito 'Rocco will write a letter to Martina and a postcard to Vito.' Therefore, if example (22) is assimilated to them, it will not constitute a serious problem for an analysis in terms of flat structure.

4. THE ANALYSIS

The data I have presented in section 2 show that restructuring verbs act as a class with respect to a variety of phenomena. In Monachesi (1993a) and in Monachesi (1995), I have suggested that an analysis in terms of argument composition can provide a uniform account of the different properties of these verbs. The idea behind this approach is that the arguments of a verbal complement are raised to become arguments of the restructuring verb, thus creating a verbal complex. The mechanism crucially relies on the notion of structure sharing which is widely used within HPSG. In this case, there will be structure sharing between the complements of the embedded verb and those of the main verb. The mechanism of argument composition was first introduced in HPSG by Hinrichs and Nakazawa (1990) for the treatment of the German verb cluster, and it is currently employed in the analysis of Dutch (Rentier, 1994; van Noord and Bouma, 1994), French (Miller and Sag, 1993; Abeille and Godard, 1994; Miller and Sag, 1995; Abeille et al, 1998), and Polish (Przepiorkowski and Kupsc, 1997). Its effect is similar to that of division in Categorial Grammar (Moortgat, 1988), which has also been used to account for Spanish clitic climbing (Nishida, 1991). In the following sections, I show how the argument composition analysis can account for the relevant properties of Italian restructuring verbs. However, it could be argued that certain properties of these verbs could be better treated as instances of long distance movement operations. Within HPSG, it would be possible to make use of nonlocal features and the Nonlocal Feature Principle, which is the mechanism used to deal with Unbounded Dependencies Constructions. Clitic climbing can be thus analyzed as a case of unbounded dependency. The nonlocal features can be employed to encode the information that a clitic, which originates as argument of the embedded verb, can appear attached to the higher verb. In section 5, I discuss the consequences of such an analysis and show the problems it faces. In particular, specific locality constraints need to be

Italian Restructuring Verbs

321

postulated and certain generalizations are missed if the analysis is extended to account for the various phenomena which are triggered by restructuring verbs. 4.1. Restructuring Verbs and Argument Composition In section 3, I have suggested that two different structures should be associated with restructuring verbs. In fact, restructuring verbs can subcategorize for a VP, but they can also inherit the complements of their infinitival verbal argument. It is the latter configuration that will be necessary to allow clitic climbing. Within HPSG, a lexical rule can be used to establish a relation between these two sets of words:7 (24)

Argument Composition Lexical Rule (ACLR)

The lexical rule relates words, namely restructuring verbs which subcategorize for a saturated VP, with other ones which have the same properties, except that they do not subcategorize for a VP, but for a verbal argument and the complements of the latter. The partition of SYNSEM in phrase synsem (p-ss) and word synsem (w-ss) together with the appropriate feature declaration represents a way to distinguish between words and phrases, alternatively the attribute LEX + / - could be used. In this rule, the notation © stands for the append relation, while the tag GO indicates structure sharing between the elements in the COMPS list of the infinitival and that of the restructuring verb. The condition CUTS { } ensures that argument composition occurs only if the embedded verb does not constitute a cliticized verb form. Complements will be ordered according to Linear Precedence constraints; in particular, the following constraints are relevant in this context: • A lexical head precedes all its complements. • Verbal complements are ordered according to the obliqueness hierarchy. • A lexical verbal complement precedes a phrasal complement.

322

Paola Monachesi

The lexical rule presented above should relate only verbs which belong to the restructuring class. In this work, I will assume that the set of the relevant verbs is specified lexically since there is a big variation among speakers as to which verbs belong to the restructuring class.8 As noticed in Rizzi (1982), for every Italian speaker there is a welldefined class of verbs which allows clitic climbing, and opposed to it there is another class of verbs which clearly does not trigger it. In between the two classes there is a group of verbs for which judgments vary from speaker to speaker, as in the case of the conative verbs (e.g. cercare, tentare, provare 'to try') and sometimes there are isolated verbs which according to some speakers trigger clitic climbing, such as the verb sembrare 'seem.' It seems that membership in the class of trigger verbs is rather idiosyncratic, and this is even more evident if dialects are taken into consideration, as will become clear in section 6.3.1. Attempts to derive the class of trigger verbs semantically, though interesting, are not very convincing since they are not sufficiently developed to capture the idiosyncratic variation which is typical of restructuring verbs. For example, Burzio (1986) suggests that trigger verbs are characterized as being semantically impoverished. He presents the following minimal pair: (25) a. Lo voglio leggere. cl.(acc) want read 'I want to read it.' b. *Lo desidero leggere. cl.(acc) wish read 'I wish to read it.' According to him, volere acts as a trigger verb because it is semantically weak, while this is not the case for desiderare? However, he does not spell out this idea in more detail, so it is difficult to see what its scientific content is. The concept of semantic impoverishment is reconsidered by Emonds (1996), but also in his case no attempt is made to make this idea more concrete. A more formalized attempt to derive the class of trigger verbs semantically is Napoli (1981). She claims that the rule of Restructuring proposed by Rizzi should apply whenever the structural conditions are met, after which a rule of semantic interpretation applies. If the latter rule is blocked for some reason, the sentence does not receive an interpretation and it is therefore judged as not acceptable. According to her, it is through the rule of semantic interpretation that it is possible to determine which verbs act as a trigger and which do not. While I believe that there might be some semantic characterization that defines the set of trigger verbs, I think that the work of Napoli is not supported by enough empirical evidence. How-

Italian Restructuring Verbs

323

ever, her hypothesis is interesting and it would be relevant to devote some research to formalize it into a more up-to-date framework and to test it against a broader set of data.10 In the following sections, I consider the different properties which are triggered by restructuring verbs: clitic climbing, climbing of loro, long NP-movement, tough constructions, and auxiliary selection, and I show how the argument composition lexical rule presented above can provide a uniform account of them. 4.2. Restructuring Verbs and Clitic Climbing The analysis of clitic climbing is based on the assumption that Italian clitics behave as lexically attached inflectional affixes, as argued at length in Monachesi (1995). Therefore clitics will not be considered lexical items which are located in a specific position by the rules of syntax, but featural information which is provided in the lexicon and used in morphology and phonology for the realization of the cliticized verb form.11 However, it should be noticed that Italian clitics are in complementary distribution with full complements, as exemplified below: (26)

*Valentina lo mangia il gelato. Valentina cl.(acc) eats the ice-cream 'Valentina eats the ice-cream.'

Therefore it is necessary to have a mechanism which can provide the necessary featural information and which can ensure the complementary distribution between clitics and full phrases. The following lexical rule, proposed in Abeille et al. (1998) for French and adopted in Monachesi (1995) for Italian, can serve this purpose. It will license cliticized verb forms:12 (27) Complement Cliticization Lexical Rule (CCLR)

The effect of this rule is that the complements of the verb are removed from the COMPS list one at a time and are added as members of the CUTS set.13 If an argument is realized as a clitic it cannot be realized also as a

324

Paola Monachesi

full phrase, accounting in this way for their complementary distribution. The CLTS feature will encode the agreement information relative to the clitic and will act as interface to morphology. Clitic ordering will be dealt with in morphology; in particular the rigid order of Italian clitics can be accounted for by assuming that clitics belong to different position classes. Such a notion can be expressed by means of template morphology (Simpson and Withgott, 1986). The following template should be adopted for Italian: (28) Template for the Italian clitic cluster Position:

I mi ti gli le vi

II ci

III si (ref)

IV lo la li le (ace)

V si (imp)

VI ne

Italian clitics will be thus assigned to different position classes and they will appear in the order prescribed by the template. The lexical rule presented above plays a key role in the analysis of cases of cliticization such as the one exemplified below, where the clitic has not climbed: (29) Anna vuole comprarlo. Anna wants to buy cl.(acc) 'Anna wants to buy it.' The Complement Cliticization Lexical Rule applies to words which are licensed by the lexical entry for comprare, the relevant part of which is shown below:

The effect of the rule is that the direct object will be removed from the COMPS list and it will be added as member of the CLTS set:

The necessary featural information will be available for spelling out the cliticized verb form in phonology. It should be noticed that, in those cases where the clitic has not climbed, the restructuring verb subcategorizes for

Italian Restructuring Verbs

325

a VP. Sentence (29) will receive the following representation:

On the other hand, cases of clitic climbing are accounted for by means of an interaction of the Complement Cliticization Lexical Rule and the Argument Composition Lexical Rule (presented in the previous section), which is repeated here:14 (24) Argument Composition Lexical Rule (ACLR)

Given a sentence such as (33), which exemplifies a case of clitic climbing, (33) Anna lo vuole comprare. Anna cl.(acc) wants to buy 'Anna wants to buy it.' the rule will apply to words which are described by a lexical entry, the relevant part of which is represented as:

The lexical rule will establish a relation with other words, with similar properties, but with crucially different subcategorization requirements. In fact, the complements of the lower verb will be inherited by the restructur-

326

Paola Monachesi

ing verb, creating thus a "clause union" effect. A partial description of vuole after the application of ACLR is:

The CCLR can then apply to license cliticized verbs. Its effect is that the relevant complement will be removed from the COMPS list and added as value of the CLTS set:

The following flat structure will be associated with sentence (33):15

Given a sequence of restructuring verbs, the ACLR would predict that clitics can attach to any of them. This is indeed the case: (38) a. Anna lo vuole poter comprare. Anna cl.(acc) wants to can buy b. Anna vuole poterlo comprare. Anna wants to can cl.(acc) buy c. Anna vuole poter comprarlo. Anna wants to can buy cl.(acc) 'Anna wants to be able to buy it.' In (38a), the ACLR is triggered twice, applying to the two restructuring verbs. Therefore, the finite verb vuole will inherit the complement of the lowest verb which will be realized as a clitic. The result will be a com-

Italian Restructuring Verbs

327

pletely flat structure:

As for (38b), the ACLR will be triggered once; the middle verb subcategorizes for the complement of the embedded verb which is realized as a clitic. The middle verb and the lower one form a constituent, namely a VP which is the complement of the higher one:

The last example is similar to the one presented in (29). In particular, the Complement Cliticization Lexical Rule applies to the lower verb in order to license the cliticized verb comprarlo. Italian clitics tend to cluster together; therefore if there are two clitics which originate as complements of the same verb, they must climb together:16 (41) a. *Valentina lo voleva spedirgli. Valentina cl.(acc) wanted to send cl.(dat) 'Valentina wanted to send it to him.' b. Valentina glielo voleva spedire. Valentina cl.(dat) cl.(acc) wanted to send 'Valentina wanted to send it to him.'

328

Paola Monachesi

Sentence (4la) is ungrammatical because the clitics are split. The condition that the curs set of the embedded verb must be empty, as stated in the output of the lexical rule in (24), enforces that the two clitics must climb together. In fact, the main verb can inherit the complements of the embedded verb only if the verbal complement does not constitute a cliticized verb form. This implies that the curs set of the lower verb must be empty, while this is not the case in (4la) where it contains the information related to the dative clitic. Therefore, the restructuring verb will not be able to inherit the subcategorization requirements of the verbal complement and so it cannot license the presence of the object clitic lo. Clitic climbing is triggered only by certain verbs; therefore a sentence like the following is ungrammatical because promettere does not belong to the class of restructuring verbs: (42)

*Anna lo promette di comprare. Anna cl.(acc) promises to buy 'Anna promises to buy it.'

An advantage of the argument composition approach is that it is lexically constrained; therefore the mechanism is triggered only by restructuring verbs. In the case of the sentence above, the main verb is not lexically marked as a restructuring verb; therefore it can only subcategorize for a saturated VP and not for the arguments of the verbal complement. The presence of the clitic attached to the main verb is thus not licensed, which is the desired result. 4.3. The Bisyllabic Clitic loro Restructuring verbs allow not only for the climbing of monosyllabic clitics, but also for the climbing of the bisyllabic clitic loro 'to them.' Loro, which is the third person plural dative pronoun, has very peculiar properties. On the one hand, it behaves like the other Italian clitics since it occurs in a different position from that of the related complement, it cannot be coordinated, and it cannot be modified, but on the other hand it differs from them in several respects, showing properties of lexical words.17 As already mentioned, the morphosyntactic features of the verb determine the direction of attachment for monosyllabic clitics. On the other hand, the placement of loro is insensitive to the finiteness of the verb; in fact, it always follows: (43) a. Dava loro notizie delle due figlie sposate. gave to them news of the two daughters married '(She) gave them news about the two married daughters.'

329

Italian Restructuring Verbs

b. Una valanga di letters di ammiratrici che gli chiedono an avalanche of letters of fans that cl.(dat) ask di dare loro un figlio. to give to them a child 'An avalanche of letters from fans that ask him to give them a child.' While no material can intervene between monosyllabic clitics and the verb, certain adverbs like spesso 'always' and mat 'never' can separate loro from the verb: (44) // mercato non dam mai loro le somme the market will not give never to them the amounts necessarie per effettuare il pagamento. necessary to make the payment 'The market will never give them the amount necessary to make the payment.' When an accusative and a dative monosyllabic clitic combine in a cluster, they will follow the order dative-accusative. On the other hand, when loro combines with an accusative clitic, the order will be accusative-dative: (45) Martina deve spedirla loro. Martina must send cl.(acc) to them 'Martina must send it to them.' In the previous section, I have shown that, in case of clitic climbing, monosyllabic clitics cannot be split (cf. (41)); however loro does not need to cluster with the other clitics in case of climbing. In addition, while Italian monosyllabic clitic are unstressed, loro bears word primary stress. The following table summarizes the differences between monosyllabic clitics and loro:18

(46) Monosyllabic clitics Position Adjacency

Proclitic/enclitic to the verb Cannot be separated from the verb Ordering Dative-accusative Clitic climbing Cannot be split when they climb Phonology Do not bear stress

Loro Follows the verb Certain adverbs can intervene Accusative-dative Can be split Bears stress

330

Paola Monachesi

In Monachesi (1995), I have argued that it is possible to account for the similarities and the differences between monosyllabic clitics and low, if they are assigned a different status. In fact, the former show affix like behavior while the latter exhibits word like behavior. Therefore, low can be considered an element of category XP[ + CLitic].19 If low is considered a saturated XP, it is possible to account for the fact that it cannot have complements. However, low should be distinguished from other XPS, in that it has different properties; therefore the feature [ + CLitic] will be used to this purpose.20 If low is an XP, its position will be determined by Linear Precedence constraints. The following constraint, which is independently motivated, accounts for the fact that low follows the verb:21 (47) Linear Precedence Constraint 1 (LPI) • A lexical head precedes all its complements. However, this LP constraint would not be sufficient to account for a sentence such as the following, where low precedes the direct object: (48) Una valanga di lettere di ammimtrici che gli chiedono an avalanche of letters of fans that cl.(dat) ask di dare low un figlio. to give to them a child 'An avalanche of letters from fans that ask him to give them a child.' An additional LP constraint can be proposed in order to account for the fact that low precedes all the phrasal complements: (49) Linear Precedence Constraint 2 (LP2) • A complement marked [ + CL] precedes all the other phrasal complements. Given these LP constraints, it is possible to account for the order of low in sentence (48). I refer to Monachesi (1995) for an analysis of the other properties of low discussed in this section. In what follows, I will be mainly concerned with the fact that low can undergo clitic climbing. As already mentioned, low can undergo optional clitic climbing with restructuring verbs: (50) a. Non sapevano cosa potesse low capitare. not know what could to them happen b. Non sapevano cosa potesse capitare low. not know what could happen to them 'They didn't know what could happen to them.'

Italian Restructuring Verbs

331

I will show that the interaction of the Argument Composition Lexical Rule with the LP constraints proposed above can account for the climbing of low. As already shown in the previous section, the output of the ACLR licenses restructuring verbs which subcategorize for a verbal complement and for the arguments of the verbal complement. I repeat the lexical rule here, for convenience:

The output of the lexical rule will thus license a word with the following COMPS list:

Therefore, in example (50) if argument composition has been triggered, the main verb will inherit the complements of capitare. Given the XP status of low, it will be ordered by the LP constraints; in particular, LPI will order low after the main verb. Since capitare is a v, and does not represent a phrasal complement, LP2 will not apply; therefore two orders will be allowed which are both grammatical, one as in (50a), exemplified below:

and another one where low follows the verbal complement as in (50b).22 However, it should be noted that a normal XP is not allowed to climb; it is not possible to have an XP (in this case a Giovanni) in the position

332

Paola Monachesi

occupied by loro: (53)

*Rocco non sapeua cosa potesse a Giovanni capitare. Rocco not know what can to Giovanni happen 'Rocco didn't know what could happen to Giovanni.'

In order to rule out this sentence it is necessary to revise in the following way the LP constraint (introduced in section 4.1) which says that a lexical verbal complement precedes a phrasal complement: •

A lexical verbal complement precedes a phrasal complement which is marked [ — CL].

In this way, sentence (53) will be correctly ruled out since a Giovanni is marked [ - CL]. In fact, in this case there is a verbal complement (capitare) which follows an XP [ - CL] and this is contrary to what is stated in the LP constraint. On the other hand, a sentence such as (50a) where the verbal complement follows the XP loro will be allowed. This is because loro is marked [ + CL], therefore it escapes the constraint mentioned above. As already mentioned, loro does not have to cluster with the other clitics. It is possible to have a monosyllabic clitic attached to the restructuring verb while loro follows the infinitival verb: (54) La pub fomire loro una dichiarazione del ministero. cl.(acc) can provide to them a declaration from the Ministry 'A declaration of the ministry can provide them with it.' This is crucially different from the case where two monosyllabic clitics are present; in fact in that situation clitics cannot be split. This different distribution provides further support for the word status of loro. Since loro is not an affix (unlike monosyllabic clitics), but an XP, there will be no cliticized verb form fomire loro in the lexicon. If the infinitival verb has not combined with a clitic, the curs set will be empty, argument composition will be thus possible, and the restructuring verb will inherit the complements of the infinitival verb. The following COMPS list will be associated with it:

The Complement Cliticization Lexical Rule will license the cliticized verb form la pub where the clitic realizes the object NP. On the other hand, since loro is an XP, it will be ordered according to the Linear Precedence constraints. In this particular case, it complies with LPl, which states that a

Italian Restructuring Verbs

333

complement should follow the head, accounting in this way for the fact that it does not cluster with the monosyllabic clitics. The following representation shows this:

Climbing of loro is possible only with restructuring verbs; therefore the following sentence is ungrammatical because decidere does not belong to this class: (57)

*Rocco decide loro di spedire la lettera. Rocco decides to them to send the letter 'Rocco decides to send the letter to them.'

Since the argument composition mechanism is lexically constrained, only restructuring verbs can inherit the complements of the lower verb. This is not the case of decidere; therefore the presence of loro in that position is not allowed.

4.4. Restructuring Verbs and Long Np-movement Restructuring verbs act as a class also with respect to another phenomenon, namely, long NP-movement. If the clitic si, which acts as a passivizing element, is attached to the restructuring verb, the direct object of the verb in the embedded clause is promoted to subject of the restructuring verb: (58)

Queste case si uogliono uendere a caro prezzo. these houses SI want to sell at high price 'These houses are wanted to be sold at a high price.'

In this example, queste case originates as direct object of vendere and is promoted to subject of si uogliono.

334

Paola Monachesi

The parallelism between this type of construction and the cases of clitic climbing discussed in section 4.2 is straightforward. I have already mentioned that only restructuring verbs can trigger clitic climbing, as shown in (42), repeated here: (42)

*Anna lo promette di comprare. Anna cl.(acc) promises to buy 'Anna promises to buy it.'

The same holds for long NP-movement; the object cannot be promoted to subject of the main verb if it does not belong to the restructuring class: (59)

*Le nuove case popolari si sono promesse di costruire. the new houses council SI are promised to build 'The new council houses are promised to be built.'

Furthermore, if more than one restructuring verb is present, the clitic can appear on the highest verb, as in (38a), repeated here: (38a) Anna lo vuole poter comprare. Anna cl.(acc) wants to can buy 'Anna wants to be able to buy it.' A similar situation can be found in the long NP-movement configuration if more than one restructuring verb is present: (60) Queste case si uogliono poter vendere a caro prezzo. these houses SI want can sell at high price 'These houses are wanted to be able to be sold at a high price.' In this case, queste case, which is the object of vendere, acts as subject of the main verb. Recall that when two clitics are subcategorized by the same verb, if they climb, they cannot be separated, as can be seen in (41a), repeated here: (4la)

*Valentina lo voleva spedirgli. Valentina cl.(acc) wanted to send cl.(dat) 'Valentina wanted to send it to him.'

Similarly, if long NP-movement interacts with clitic climbing, it is not possible to have a sentence where the clitic has not climbed: (61)

*Queste case si uogliono uendergli a caro prezzo. these houses SI want to sell cl.(dat) at high price 'These houses are wanted to be sold to him at a high price.'

Italian Restructuring Verbs

335

In this example, both queste case and gli are complements of the lower verb vendere', it is not possible to have a structure where the NP is preposed while the clitic stays behind. In fact, the clitic should also climb, as in the following grammatical sentence: (62) Queste case gli si vogliono vendere a cam prezzo. these houses cl.(dat) SI want to sell at high price 'These houses are wanted to be sold to him at a high price.' Given the similarities between the two constructions, it would be desirable to have an analysis that could account for both. In section 4.4.2,1 will show that the analysis in terms of argument composition which I have proposed for clitic climbing carries over naturally to the corresponding cases of long NP-movement. In particular, this is achieved through an interaction between the Argument Composition Lexical Rule and the Middle si Lexical Rule which I introduce in the following section.

4.4.1. MIDDLE si Italian clitic si has several different uses. It occurs with an impersonal interpretation, a middle interpretation, an ergative interpretation, and a reflexive-reciprocal interpretation, and it can also be an inherent reflexive clitic. I refer to Monachesi (1995) for an analysis of these different uses of si. In this section, I will only be concerned with middle si; in this case the clitic acts as a grammatical marker and the sentence receives a middle passive interpretation: (63) Gli spaghetti si mangiano spesso. the spaghetti SI eat often 'Spaghetti is often eaten.' This construction, which is fully productive, shares with passive the property that the direct object functions as a superficial subject; in particular it agrees with the verb and it can be replaced by a null subject.23 Therefore, it seems that clitic si acts as a passivizer and its function is similar to that of passive morphology. However, a difference between the middle si construction and the passive construction is that the former does not admit a by-phrase; being in this way similar to agentless passives.24 A lexical rule to account for middle si constructions (MIDSI-LR) can be proposed. This rule, which is similar to the lexical rule for passive, licenses

336

Paola Monachesi

middle verb forms:25 (64) Lexical rule for middle verb forms (MIDSI-LR)

The rule relates transitive verbs with middle passive verb forms.26 It has the effect of removing the subject from the SUBJ list and of promoting the object of the transitive verb to subject of the middle verb. Furthermore, it specifies that the subject in middle si constructions should receive nominative case and should be in the third person. The curs feature will trigger the presence of the clitic attached to the verb.27 4.4.2. THE ANALYSIS OF LONG NP-MOVEMENT In this section, I show that the interaction of the Argument Composition Lexical Rule with the Middle si Lexical Rule accounts in the desired way for those cases of long NP-movement, exemplified in (58), repeated here: (58)

Queste case si vogliono vendere a cam prezzo. these houses SI want to sell at high price 'These houses are wanted to be sold at a high price.'

The restructuring verb will be subject to the application of the ACLR; it will thus inherit the object of the complement verb:

The words licensed by the ACLR can in turn be the input of the Middle si Lexical Rule. The lexical rule will relate words with the COMPS list specified above with other ones where the object has been promoted to subject of the middle verb form:

Italian Restructuring Verbs

337

Therefore, in example (58), queste case which is the object of vendere will be inherited by uogliono and will be eventually promoted to subject of the restructuring verb. Since queste case is a full phrase, the Valence Principle, which is the principle of the grammar responsible for checking off subcategorization requirements that have been satisfied, can normally apply and remove the SUBJ slot. Sentence (58) receives the following representation:

As already mentioned, it is possible to have sentences where more than one restructuring verb is present. In this case, the object of the last verb is promoted to subject of the main verb, as in (60), repeated here: (60) Queste case si vogliono poter vendere a caro prezzo. these houses SI want can sell at high price 'These houses are wanted to be able to be sold at a high price.' The corresponding clitic climbing example, where the clitic climbs more than one clause, was accounted for by applying the Argument Composition Lexical Rule to the two restructuring verbs. Similarly, in the case of long NP-movement exemplified above, a double application of ACLR to the restructuring verbs makes queste case the object of vogliono. The application of the Middle si Lexical Rule makes the object queste case have the function of the subject. However, recall that in the clitic climbing cases, if more than one verb is present, it is also possible for the clitic to attach to the middle verb, as in

338

Paola Monachesi

(38b), repeated here: (38b) Anna vuole poterlo comprare. Anna wants to can cl.(acc) buy 'Anna wants to be able to buy it.' On the other hand, in a long NP-movement configuration, it is not possible for the NP to intervene between the two restructuring verbs, as shown here: (68)

*Si vogliono queste case poter vendere a caro prezzo. SI want these houses can sell at high price 'These houses are wanted to be able to be sold at a high price.'

In this case, the NP has the function of subject; therefore sentence (68) would be ruled out by the LP constraint that requires subjects to be ordered before the verb. Restructuring verbs subcategorize for a verbal complement whose CUTS set is empty; in this way sentences such as: (41 a)

* Valentino lo voleva spedirgli. Valentinocl.(acc) wanted to send cl.(dat) 'Valentino wanted to send it to him.'

are correctly ruled out, as shown in section 4.2. Similarly, sentences such as (61), repeated below, where there is an interaction among clitic climbing, long NP-movement, and middle si are also correctly ruled out in the same way: (61)

* Queste case si vogliono vendergli a caro prezzo. these houses SI want to sell cl.(dat) at high price 'These houses are wanted to be sold to him at a high price.'

The cliticized verb form uendergli does not match the requirements on the verbal complement since the curs set will not be empty; in fact it contains the information related to the dative clitic. Therefore, the restructuring verb cannot subcategorize for the object of the embedded verb; this implies that the appropriate conditions for the application of the Middle si Lexical rule are lacking. As a consequence, in this case no object preposing will be permitted and this is the desired result. On the other hand, sentences such as (62) are correctly allowed: (62) Queste case gli si vogliono vendere a caro prezzo. these houses cl.(dat) SI want to sell at high price These houses are wanted to be sold to him at a high price.' In this case, the output of ACLR will license restructuring verbs which subcategorize for the complements of the embedded verb. The following

Italian Restructuring Verbs

339

specification describes the COMPS list of the verb vogliono after the application of ACLR:

In addition, the Middle si Lexical Rule will relate words with the COMPS list specified above with other ones where the object has the function of the subject, while the clitic si is added as member of the CUTS set:

In addition, the Complement Cliticization Lexical Rule applies to words which are licensed by the lexical entry for vogliono, the relevant parts of which are shown above. Its effect is that the PP complement will be removed from the COMPS list and added as member of the CUTS set:

The sentence will receive the following structure:

340

Paola Monachesi

Like clitic climbing, long NP-movement is triggered only by restructuring verbs; therefore sentences such as (59) are not grammatical since promettere is not a restructuring verb: (59)

*Le nuoue case popolari si sono promesse di costruire. the new houses council SI are promised to build 'The new council houses are promised to be built.'

They are ruled out for the same reasons as sentences such as (42), repeated here: (42)

*Anna lo promette di comprare. Anna cl.(acc) promises to buy 'Anna promises to buy it.'

The reason is that verbs which do not belong to the restructuring class will not be subject to the application of the Argument Composition Lexical Rule. Since argument composition is not allowed, the application of the lexical rule for middle si construction will not be triggered and so there is no way that the object of costruire can become the subject of promettere. Sentences where a verb that does not belong to the restructuring class intervenes between sequences of restructuring verbs, such as: (73)

*Le nuove case popolari si vogliono promettere the new houses council SI want to promise di poter costruire. to be able to build 'The new council houses are wanted to promise to be able to be built.'

are ruled out for the same reasons.

4.5. Tough Constructions Restructuring verbs act as a class also with respect to tough constructions. While in Italian this construction is strictly bounded, it appears to be unbounded if a restructuring verb is present. In tough constructions, the embedded object surfaces as superficial subject, as in the following example from Rizzi (1982): (74) Questo problema e difficile da risolvere. 'This problem is difficult to solve.' It should be noticed that this construction differs from the English equivalent, since in English the direct object can be extracted from almost

Italian Restructuring Verbs

341

any level of embedding:28 (75)

That book will be impossible for you to convince the class to try to finish before Monday.

This is not the case in Italian, where the construction is strictly bounded: (76)

*Questo lavoro e facile da promettere difinire per domani. this work is easy to promise to finish by tomorrow

On the other hand, if a restructuring verb is present, it is possible for the object of an embedded infinitival verb to be promoted to subject of the copular construction: (77)

Questa canzone e facile da cominciare a cantare. this song is easy to begin to sing

The analysis proposed for restructuring verbs can provide an explanation for the possibility of (77) and the impossibility of (76). In example (77), the Argument Composition Lexical Rule will apply to cominciare, which will inherit the complement of the infinitival. Therefore, the object of cantare becomes the object of cominciare. This means that this case of tough construction is reduced to the simple case exemplified in (74). On the other hand, the ungrammaticality of (76) is due to the fact that promettere is not a restructuring verb and therefore the ACLR cannot apply; thus promettere cannot subcategorize for the object of the lower verb. Assuming an analysis of tough constructions along the lines of that proposed by Abeille et al. (1998) for French, adjectives such as facile will subcategorize for a VP with an object in its COMPS list whose index is shared with the subject of the copular construction:29

In example (77), the object of the lower infinitival will become the object of the restructuring verb by means of the argument composition mechanism. Given the specification for facile presented above, this object will be coindexed with the subject of the copular construction accounting in this way for the sentence.30

342

Paola Monachesi

4.6. Auxiliary Selection Restructuring verbs trigger auxiliary change; namely, restructuring verbs which have avere as auxiliary can change it into essere if the embedded verb requires essere as auxiliary: (79) a. Rocco ha voluto partire. Rocco has wanted to leave 'Rocco has wanted to leave.' b. Rocco e voluto partire. Rocco is wanted to leave 'Rocco has wanted to leave.' While this change of auxiliary is optional in the examples above, this is not the case when clitic climbing occurs. As shown in the following sentences, it is the auxiliary of the embedded verb that is realized: (80) a. * Rocco ci ha voluto andare. Rocco cl.(loc) has wanted to go 'Rocco has wanted to go there.' b. Rocco ci e voluto andare. Rocco cl.(loc) is wanted to go 'Rocco has wanted to go there.' More generally, it seems that it is the last verb in the verbal complex that determines which auxiliary is selected, as shown by the following examples where several restructuring verbs are present: (81) a. Rocco li ha voluti andare a cercare. Rocco cl.(acc) has wanted to go to find 'Rocco has wanted to find them.' b. Rocco ci sarebbe voluto cominciare ad andare. Rocco cl.(loc) would be wanted to begin to go 'Rocco would have wanted to begin to go there.' In (81a), volere and cercare require avere as auxiliary, while andare requires essere. In example (81b) both volere and cominciare require avere, while andare requires essere. It should be noted that if the other auxiliary is selected, the sentences above are not acceptable: (82) a. *Rocco li e voluti andare a cercare. Rocco cl.(acc) is wanted to go to find 'Rocco has wanted to find them.' b. *Rocco ci avrebbe voluto cominciare ad andare. Rocco cl.(loc) would have wanted to begin to go 'Rocco would have wanted to begin to go there.'

Italian Restructuring Verbs

343

However, the generalization mentioned above does not seem to hold if the restructuring verb requires essere as auxiliary: (83) Rocco lo e andato a cercare. Rocco cl.(acc) is gone to find 'Rocco has gone to find him.' In this case, cercare selects avere and it is the last verb in the verbal complex, but instead essere is realized. Therefore, it seems that in the presence of a verbal complex, which is reflected by the fact that clitic climbing occurs, the following generalizations hold with respect to auxiliary selection: •

If the restructuring verb selects essere then this is the auxiliary which is realized. • If the restructuring verb selects avere, then it is the auxiliary which is selected by the last verb, which is realized. If auxiliary selection is lexically encoded for each verb, through a feature AUX, it would be possible to account for the generalizations mentioned above. Restructuring verbs which have undergone argument composition and which select essere would simply realize this auxiliary. In the case they require avere, it should be feasible, with an appropriate system of features, to inherit the auxiliary requirements of the last verb in the verbal complex. However, the fact that it is possible to have optional auxiliary change also when the clitics do not climb, is unexpected: (84) a. Rocco ha voluto andarci. Rocco has wanted to go cl.(loc) 'Rocco has wanted to go there.' b. Rocco e voluto andarci. Rocco is wanted to go cl.(loc) 'Rocco has wanted to go there.' In fact, according to the generalizations expressed above, it seems possible to realize the auxiliary of the lower verb only in the presence of a verbal complex.31 An alternative possibility, which could be explored in order to account for the auxiliary selection facts mentioned above, is to take into consideration the role of unaccusativity.32 An analysis in this direction is that of Butt (1995), which proposes an LFG account of this property of restructuring verbs. She suggests a mechanism of argument fusion according to which, when verbs like andare (which require the auxiliary essere) are part of a complex predicate, their theme argument is fused with the highest embedded argument. Furthermore, she assumes that, when this process occurs, the auxiliary essere should be selected. Even though this analysis can

344

Paola Monachesi

account for the relevant data, it is somehow stipulative since it does not explain why it is the auxiliary essere that surfaces when argument fusion occurs and not avere.

5. AN ALTERNATIVE ANALYSIS

In the previous sections, I have presented an analysis of the different properties of restructuring verbs in terms of argument composition. The basic insight behind it is that the restructuring verb inherits the complements of the infinitival one and the two (or more) verbs form a verbal complex. Therefore, a clause union effect is obtained. In this section, I will discuss an alternative analysis of the properties of restructuring verbs, one which will take into consideration the nonlocal character of clitic climbing, long NP-movement, and tough constructions. As already mentioned, it could be possible to account for these phenomena in terms of nonlocal features and the Nonlocal Feature Principle, which have been proposed in Pollard and Sag (1994) to deal with Unbounded Dependency Constructions. In Monachesi (1993c), I have presented an analysis of clitic climbing which indeed relies on this mechanism.33 The treatment makes use of lexical rules to update the subcategorization requirements of the verbal head and of nonlocal features to encode the information that a clitic can appear at some point in the tree. Given a sentence such as (38a), repeated here: (38a) Anna lo vuole poter comprare. Anna cl.(acc) wants to can buy 'Anna wants to be able to buy it.' It will receive the following representation, if an analysis in terms of nonlocal features is assumed:

Italian Restructuring Verbs

345

A verb like comprare will be subject to the Lexical Rule for Object Clitics. Its effect is that one element will be removed from the COMPS list of the verb and added as value of the oc set. This value will percolate by means of the Nonlocal Feature Principle, the percolation will stop once a TO-BIND|OC feature is found. The TO-BIND|OC feature will be present on the cliticized verb form. If the values of the TO-BIND|OC feature and the INHER|OC feature are identical, the nonlocal dependency is bound off according to the Nonlocal Feature Principle: it will be subtracted from the set of nonlocal feature values that are passed up to the mother.34 However, this type of account faces certain problems which are avoided in an analysis in terms of argument composition. An obvious problem is that the analysis in terms of nonlocal features overgenerates, allowing clitic climbing also with verbs that do not trigger it. Therefore, a sentence such as the following, where the main verb is not a restructuring verb, will be accepted, unless specific conditions are added: (42) *Anna lo promette di comprare. Anna cl.(acc) promises to buy 'Anna promises to buy it.' The approach does not account in a straightforward way for the fact that only restructuring verbs can trigger clitic climbing. This is because Unbounded Dependency Constructions are subject to different locality constraints than clitic climbing constructions. In order to rule out sentences such as (42), it is necessary to add specific constraints on the path of the oc feature. The percolation of the feature must be stopped if there is a verb that does not trigger clitic climbing and must be allowed if there is a verb that triggers it. In Monachesi (1993c) this was achieved by imposing the condition that verbs that do not allow clitic climbing should select for a complement with INHER|OC { }. While this could be a possible way to obtain the correct result, it also suggests that this approach does not naturally capture the fact that clitic climbing is triggered only by a specific class of verbs since specific conditions need to be imposed to obtain the right result.35 Furthermore, the constraints need to be imposed both on verbs that do not trigger clitic climbing and on those that trigger it; this is clearly not very economical. In addition, some generalizations are missed if the analysis is extended to account for long NP-movement: (58) Queste case si vogliono vendere a cam prezzo. these houses SI want to sell at high price These houses are wanted to be sold at a high price.'

346

Paola Monachesi

As already mentioned in section 4.4, long NP-movement and clitic climbing are subject to similar constraints; it is thus desirable to have an analysis that could account for both constructions. Therefore, sentence (58) should be analyzed by means of an interaction between the mechanism based on nonlocal features and the rule proposed in section 4.4.1 for middle si constructions. However, there are some problems connected to this. Recall that MIDSI-LR relates two sets of words with different subcategorization requirements; its effect is to promote the direct object of a transitive verb to the subject of the middle verb form. However, given the nonlocal approach presented above, the information about the object of the embedded transitive verb, namely, queste case, would be encoded in the INKER |oc of the verb volere. One could postulate another rule for middle si constructions which relates restructuring verbs; the rule should operate on information encoded in a nonlocal feature and change the grammatical function of the element present in the INHER|OC set, making it a subject. Therefore, there would be two lexical rules in order to account for middle si constructions. One would relate restructuring verbs and would operate on material encoded in the nonlocal feature while MIDSI-LR would relate verbs that do not belong to this class, as in (63), repeated here: (63) Gli spaghetti si mangiano spesso. the spaghetti SI eat often 'Spaghetti is often eaten.' In this case, the rule would effect the subcategorization requirements of the verb; thus, the same phenomenon would be analyzed in two different ways, which is not a desirable result. A similar problem would occur in the analysis of tough constructions: (77) Questa canzone e facile da cominciare a cantare. this song is easy to begin to sing Given an analysis of tough constructions such as that sketched in section 4.5, the adjective facile subcategorizes for a VP with an object in its COMPS list whose index is shared with the subject of the copular construction. However, if nonlocal features are used to account for the properties of restructuring verbs, cominciare will encode the information relative to the object of cantare, not in the COMPS list, but in the INHER|OC set. Therefore, an alternative entry which is associated with the adjective facile should be foreseen to account for the interaction of restructuring verbs with tough constructions. In this entry, there should be a coindexation between the subject of the copular construction and the element contained in the INHER|OC set.

Italian Restructuring Verbs

347

It seems therefore that an approach in terms of nonlocal features does not account for clitic climbing in an adequate way since it does not naturally capture the fact that only a specific class of verbs triggers it. Furthermore, it does not account for long NP-movement and for the interaction of restructuring verbs and tough constructions in an elegant way. On the other hand, an analysis of restructuring verbs which is based on the mechanism of argument composition provides a more adequate treatment since it can naturally account for the intermediate distance character of the constructions involved.36

6. RESTRUCTURING VERBS IN SALENTINO The second part of this paper will be concerned with dialectal variation. I will take into account restructuring verbs in Salentino, a southern Italian dialect discussed in Calabrese (1993). I will focus on one property of restructuring verbs, namely, clitic climbing. In particular, I will consider the behavior of the verb ulire 'want' in that variety of Salentino spoken in the provinces of Brindisi and Taranto. This verb has the peculiar property of triggering clitic climbing out of a finite clause. I will show that, within HPSG, the parameterization of linguistic variation occurs in the lexicon and, in particular, that a lexical approach to clitic climbing allows for a proper treatment of the variation represented by this dialect.

6.1. Clitic Climbing in Salentino As already discussed, clitic climbing with restructuring verbs is optional in Standard Italian, but this is not the case in certain southern Italian dialects, such as Salentino, which is spoken in certain parts of Puglia. In particular, in this section I will be concerned with that variant of the dialect spoken in the northern part of the province of Lecce. In this dialect, restructuring verbs trigger obligatory clitic climbing, as can be seen from the following example:37 (86) a. *Lu pottsu kkattare krai. cl.(acc) canlsgbuyinf tomorrow b. *Pottsu kkattarlu krai. canlsg buyinj cl.(acc) tomorrow 'I can buy it tomorrow.'

348

Paola Monachesi

These data can receive an account under the argument composition analysis proposed for restructuring verbs in Standard Italian. The same mechanism will account for clitic climbing in both cases. Verbs that trigger clitic climbing will thus subcategorize for a verbal complement and for the arguments of the latter, as shown in the following lexical entry for pottsu:38

Therefore, while in Standard Italian, it is necessary to formulate a lexical rule in order to account for the optionality of clitic climbing, in Salentino a lexical entry would suffice. Each clitic climbing trigger verb will be described by a lexical entry similar to the one presented above. I assume the Complement Cliticization Lexical Rule, which I have adopted for Standard Italian and which licenses cliticized verb forms to be operative in this dialect as well. The restriction on curs in the lexical entry (87), namely, curs must be empty, accounts for obligatory climbing. A sentence such as (86b) where the clitic is attached to the lower verb is ruled out. In fact, in that case the CLTS set will contain the information related to the clitic; therefore it will not be empty and will thus not fulfill the condition imposed by the lexical entry above. It should be mentioned that there is variation with respect to which elements belong to the class of restructuring verbs in Standard Italian and in Salentino. For example, while a verb like volere 'want' is a restructuring verb in Standard Italian, this is not the case in Salentino (at least in the variant spoken in the province of Lecce). However, a crucial difference between Standard Italian and Salentino is the fact that in Italian volere allows for the presence of an infinitival complement if there is coreference between the subject of the matrix clause and that of the embedded clause: (88)

Carlo uuole venire domani. Carl wants to come tomorrow 'Carl wants to come tomorrow.'

Italian Restructuring Verbs

349

This is not the case in Salentino, where a verb like volere cannot subcategorize for an infinitival complement: (89)

*Lu Karlu ole inire krai. the Carl wants 3sg to comeinf tomorrow 'Carl wants to come tomorrow.'

More generally, in Salentino there is a class of verbs, namely, verbs of ordering, desiring, warning, etc. (of which ulire 'want' is a member), that cannot be followed by an infinitival complement. They select instead a finite clause which is introduced by the element ku, as can be seen in sentence (90):39 (90) Lu Karlu ole ku bbene krai. the Carl wants35g KU comes inrf3ig tomorrow 'Carl wants him to come tomorrow.' In particular, Calabrese (1993) suggests that ku introduces the complement of verbs which express an attitude toward, or an attempt to bring about, an event which is yet to come. The status of the element ku, which introduces finite clauses after this particular class of verbs, is rather controversial since it exhibits both properties of inflectional affixes and of complementizers.40 The idiosyncratic behavior of ku can be easily understood if diachronic data is taken into consideration; in fact ku was originally a complementizer. As Calabrese notices, ku derives from Latin quod, which was a complementizer: (91) Luvat me quod studia vigent. pleases me that studies flourish 'I am pleased that studies are flourishing.' However, as will become clear from the following discussion, synchronically, ku is becoming an inflectional affix connected to tense morphology. Yet, the process is not completed; therefore it still maintains some of its original properties of complementizers. 6.1.1. ON THE STATUS OF THE ELEMENT ku Evidence in favor of the status of ku as inflectional affix comes from the following properties: • It attaches to the verb and only clitics can separate it from the verb. • It constrains tense morphology. • It cannot have wide scope over coordination.

350

Paola Monachesi

The first property is exemplified as follows: (92) a. *Oyyu ku lu Maryu bbene krai. want lig KU the Mario comesind, 3sg tomorrow 'I want Mario to come tomorrow.' b. Oyyu ku nnu le kkatti. want lig KU not cl.(acc) buyind>2sg 'I want you not to buy them.' Sentence (92a) shows that no phrasal complement can separate ku from the verb; only clitics may intervene, as in (92b).41 The element ku constrains tense morphology; in fact, when it is present, there is a violation of the morphological agreement of tenses between the main clause and the embedded clause, which is generally assumed to be obeyed in Standard Italian. For example, the presence of a past tense in the main clause does not imply a past tense in the clause introduced by ku; in fact, only the present and the perfect may be used in this case, as shown in the following example: (93) la ulutu la Maria ku bbae/ia ™*cond.past, 1sg wanted the Maria KU go P r e s . / p e r f , 'I wanted Maria to have gone there before.'

3sg

f/iuta ddai. there

More generally, ku seems to introduce clauses characterized by a nondeictic tense; in fact, as noticed by Calabrese (1993), the tense of the embedded clause must be bound by the tense of the matrix clause in order to get a time reference.42 Additional evidence in favor of the affixal status of ku comes from coordination; in fact, in this case, the element must be repeated in front of each conjunct, as the following example from Melillo (1975) shows: (94)

Tokka ku ffacimu festa e kku ne prisamu. must KU make^ip, party and KU cl. enjoyind, 1pl 'We should have a party and enjoy it.'

However, ku exhibits also additional properties which could argue in favor of its status of complementizer: •

It is in complementary distribution with the complementizer ka.

•

It is incompatible with a wh-expression.

•

It precedes negation.

Italian Restructuring Verbs

351

The first property is shown by the following example, which demonstrates that ku and the complementizer ka cannot cooccur in the same sentence: (95)

*Voleva ka la Maria ku bbene. wanted 3ig that the Mary KU comesind >3sg 'He wanted that Mary would come.'

This seems to suggest that they are in complementary distribution because they are both complementizers. The position of negation also argues in favor of the complementizer status of ku, as can be seen from example (92b), repeated here: (92b) Oyyu ku nnu le kkatti. want ug KUnot cl.(acc)buyjW>2jg 'I want you not to buy them.' It should be noticed that in Standard Italian negation follows the complementizer che: (96) Martina dice che non vena alia festa. Martina says that not will come to the party 'Martina says that she will not come to the party.' Furthermore, ku is incompatible with a wh-expression: (97) a. *Me sta ddumannu nt/e ku add3u fare. cl. aux. wonder1sg what KU have.nrf ls_ to do lagind,1sgb. Me sta ddumannu ntje add3u fare. cl. aux. wonder what have to do 'I am wondering what to do.' In this respect it shares the same behavior as the complementizer ka: (98)

*Me sta ddumannu nt je ka add3u fare. cl. aux. wonder1sg what that have lcff to do 'I am wondering what to do.' 13

6

13

5

The following table summarizes the properties of ku:

(99) Inflectional affix properties Attaches to the verb; only clitics can intervene Constrains tense morphology No wide scope over coordination

Complementizer properties Complementary distribution with the complementizer ka Incompatible with wh-expression Precedes negation

352

Paola Monachesi

Therefore, it seems that ku has an intermediate status between that of an inflectional affix and of a complementizer. 6.2. An HPSG Analysis In this section, I propose an analysis of the data presented above. In particular, I will be concerned with the status of the element ku and how it should be represented within HPSG. Given that it shares both properties of inflectional affixes and of complementizers, I can see three possible ways of accounting for it: 1. ku is considered an inflectional affix; therefore the analysis proposed for object clitics in Standard Italian can be extended to account for it. 2. ku is considered a complementizer; therefore it will have the status of a marker and it attaches to a VP. 3. ku is a complementizer, thus a marker, but it attaches to a v. If the first possibility is considered, it would be easy to account for the fact that ku attaches to the verb and nothing but clitics can intervene. However, it is not evident how to account in a principled way for the fact that ku is in complementary distribution with the complementizer ka. On the other hand, if ku is considered a VP marker it would be straightforward to account for the complementary distribution with ka, but one would predict that it is possible for adverbs to intervene between ku and the verb, while this does not seem to be the case.43 The third option, which views ku as a complementizer, thus a marker which attaches to v, seems thus the most appropriate way to account for the hybrid status of this element. In fact, it incorporates its complementizer properties and would also account for the fact that ku is a verbal affix. According to Pollard and Sag (1994), complementizers are considered markers.** In fact, they mark the constituent in which they occur; they are distinguished by the attribute MARKING which they structure share with the mother node. Markers can select the phrase with which they combine through the feature SPEC. Therefore, the following will be the entry for ku:

In addition, verbs of order, desire, etc. should subcategorize for a finite sentence introduced by ku. However, it should be noted that, in Salentino,

Italian Restructuring Verbs

353

finite clauses introduced by ku (or ka) cannot be considered control structures, as argued in Calabrese (1993). He provides examples such as the following, where he shows that the empty category in the subject position of these clauses doesn't need to be bound by an antecedent in the matrix clause: (101) Lu Maryu ole ku ej bbete bbona. the Mario wants3sg KU is3sg good3F5g 'Mario wants her to be good.' Calabrese considers it a free pronominal that can be related to a referent in the extralinguistic context or to an antecedent in the discourse. Furthermore, he argues that sentences such as the following cannot be cases of Exceptional Case Marking: (102) Oyyu lu Maryu ku le kkatta krai. want lig the Mario KU cl.(acc) buysind, 3sg tomorrow 'I want Mario to buy them tomorrow.' In fact, the subject preceding ku must be assigned nominative case, as shown in the following example where the nominative pronoun idda 'she' occurs: (103) Oyyu idda ku bbene krai. want ljg she (nom) KU comes/nrf 35g tomorrow 'I want her to come tomorrow.' Therefore, the lexical entries for verbs of order, desire, etc. should contain the following description:

A sentence such as (102) will thus receive the following structure:

354

Paola Monachesi

It should be mentioned that the percolation of the MARKING value from the marker up to the s is ensured by the Marking Principle from Pollard and Sag (1994): (106)

MARKING Principle In a headed phrase, the MARKING value is token identical with that of the marker daughter if any, and with that of the head daughter otherwise.

Therefore, the resulting sentence will have ku as value for MARKING and will thus satisfy the condition in the lexical entry for ulire in (104). As already mentioned, no overt subject can intervene between ku and the verb: (92a) *Oyyu ku lu Maryu bbene krai. want 1sp KU the Mario comesind, tomorrow lig i r i U j3sg JAg 'I want Mario to come tomorrow.' This fact can be easily accounted for under the analysis proposed. Since ku selects for a v and not for an s, it will not be possible to have the subject in that position. In addition, under this treatment no adverbs are expected to occur between ku and the verb, which is the correct result. It should be recalled that ku is in complementary distribution with the complementizer ka; therefore a sentence such as (95), repeated here, where ku which is introduced by the verb ulire and ka cooccur, is not grammatical: (95) *Ulia ka la Maria ku bbene. wanted3ig that the Mary KU comesind, 3sg 'He wanted that Mary would come.' The complementizer ka will be also considered a marker which has the following lexical entry:

Since ka subcategorizes for an s with the value for MARKING equal to unmarked, the sentence above would be ruled out because it will have ku

Italian Restructuring Verbs

355

as value. The following representation exemplifies this:

Therefore, examples such as (95) would be ruled out in the same way as an English sentence such as the following: (109)

*/ think that that John left.

In a similar way, it is possible to rule out sentences such as the following, where ku and ka cooccur, but where the main verb subcategorizes for an embedded clause introduced by ka: (110)

*Sperava ka la Maria ku bbene. hoped3sg that the Mary KU comesind, 3sg 'He hoped that Mary would come.'

As discussed in Monachesi (1995) it would be more problematic to rule out sentences such as the one above if ku is considered an inflectional affix. 6.3. Clitic Climbing and Finite Clauses Unlike Standard Italian, clitic climbing does not occur with a verb such as ulire 'want'; in fact a sentence such as the following would be ungrammatical in Salentino of Lecce: (111)

*Le ole ku kkatta. cl.(acc) wants3igKU buyind, 3sg 'He wants to buy them.'

The ungrammaticality of this sentence could in some sense be expected since generally clitic climbing occurs out of a nonfinite clause, but in this case there is a finite one. However, as will become clear in the next section, this cannot be the reason. Instead, I suggest that clitic climbing cannot occur in (111), because there is no obligatory control. In fact, as I mentioned in the previous section, in Salentino, finite clauses introduced by ku cannot be considered control structures. I show in the next section that the reason for not having clitic climbing in sentence (111) cannot be related to the presence of a finite verb and that indeed obligatory control is a crucial condition for the occurrence of clitic climbing.

356

Paola Monachesi

6.3.1. SALENTINO OF BRINDISI AND TARANTO Calabrese (1993) mentions in a footnote data from a variety of Salentino which is spoken in the provinces of Brindisi and Taranto. This variety of Salentino is similar in all relevant respects to the variety previously discussed, but it has an additional peculiar property. In fact, if the matrix verb is ulire 'want' and if there is coreference between the subject of the matrix clause and the subject of the embedded clause, the element ku can be omitted: (112) a. Voggyu ku kkattu nu milune. wantl5g KU buy,^lxg to buy it.'

The data presented above seem to suggest that, in this variety of Salentino, clitic climbing is allowed with the verb ulire when there is obligatory control. The obligatory control configuration is reflected in the omission of the element ku. The lexical analysis of clitic climbing in terms of argument composition proposed for Standard Italian can be adopted to account for the data presented above. The fact that, in this variant of the dialect, ulire 'want' is the only verb in its class to trigger clitic climbing is accounted for by the fact that argument composition is lexically constrained. Parametric variation is thus handled in the lexicon. Therefore, if there is coreference between the subject of ulire and that of the embedded verb, the lexical entry should contain the following description:

The verb will subcategorize for a finite verb and its complements; this operation of argument composition is indicated by the tag 2. The coreference between the subjects is represented by the presence of the tag CD. It should be noted that the conditions imposed in the lexical entry above are similar to those in the output of the Argument Composition Lexical Rule proposed for Standard Italian.

358

Paola Monachesi

In this way, it is possible to account for a sentence such as (114a) where clitic climbing occurs: (114a) Lu uoggyu kkattu. cl.(acc) want^buy^^ 'I want to buy it.' It will receive the following structure:

This representation is similar to that provided for cases of clitic climbing in Standard Italian; the only crucial difference with respect to Standard Italian is that in this case two finite verbs will form a verbal complex.47 On the other hand, clitic climbing is blocked if ku is present: (115b) *Lu uogguy ku kkattu. cl.(acc) want1sg KU buy^^ 'I want to buy it.' Sentences such as (115b) are ruled out by the condition MARKING unmarked specified in (116). In other words, argument composition, thus clitic climbing, is possible only if the verbal complement has not combined with the marker ku. Clitic climbing with the verb ulire is optional; in fact clitics can also attach to the embedded verb, both if ku is present and if it is omitted. Therefore, there will also be a lexical entry where ulire subcategorizes for a saturated VP:

Italian Restructuring Verbs

359

In this case, it is required that the subject of the embedded verb and that of the main verb should be referential. It should be noted that the conditions in the lexical entry above are similar to those imposed on words licensed by the input of the Argument Composition Lexical Rule proposed for Standard Italian. The entry will account for those cases where the clitic has not climbed:48 (114b) Voggyu lu kkattu. want ljg cl.(acc) buylW>1,g 'I want to buy it.' (115a) Voggyu ku lu kkattu. want 1sg KU cl.(acc) buy/B