Autolexical Theory: Ideas and Methods 9783110889376, 9783110126266

Table of contents :
Introduction
An Autolexical account of voicing assimilation in Manipuri
Autostructural analysis of semantic roles
Word order in Romanian subjunctive clauses
Control in Autolexical Syntax
On deriving the government constraint for incorporation and inflection
Some pleasures and pitfalls of Autolexical Syntax
Toward a tri-modular analysis of -ly adverbs
Pronominal cliticization in West Flemish
Two types of mismatch between morphology and semantics
On restricting the role of morphology in Autolexical syntax
Author index
Subject index

Citation preview

Autolexical Theory

W G DE

Trends in Linguistics Studies and Monographs 85

Editor

Werner Winter

Mouton de Gruyter Berlin · New York

Autolexical Theory Ideas and Methods

edited by

Eric Schiller Elisa Steinberg Barbara Need

Mouton de Gruyter Berlin · New York

1996

Mouton de Gruyter (formerly Mouton, The Hague) is a Division of Walter de Gruyter & Co., Berlin.

© Printed on acid-free paper which falls within the guidelines of the ANSI to ensure permanence and durability.

Library of Congress

Cataloging-in-Publication-Data

Autolexical theory : ideas and methods / edited by Eric Schiller, Elisa Steinberg, Barbara Need. p. cm. — (Trends in linguistics. Studies and monographs ; 85). Includes bibliographical references and index. ISBN 3-11-012626-5 1. Autolexical theory (Linguistics) I.Schiller, Eric. II. Steinberg, Elisa, 1944III. Need, Barbara, 1959IV. Series. P158.15.A93 1995 415-dc20 95-24901 CIP

Die Deutsche Bibliothek —

Cataloging-in-Publication-Data

Autolexical theory / ideas and methods / ed. by Eric Schiller ... - Berlin ; New York : Mouton de Gruyter, 1995 (Trends in linguistics : Studies and monographs ; 85) ISBN 3-11-012626-5 NE: Schiller Eric [Hrsg.]; Trends in linguistics / Studies and monographs

© Copyright 1995 by Walter de Gruyter & Co., D-10785 Berlin All rights reserved, including those of translation into foreign languages. No part of this book 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. Printing: Arthur Collignon GmbH, Berlin. Binding: Lüderitz & Bauer, Berlin. Printed in Germany.

Contents

Introduction

1

Eric Schiller An Autolexical account of voicing assimilation in Manipuri

11

Shobhana L. Chelliah Autostructural analysis of semantic roles

31

Jan Terje Faarlund Word order in Romanian subjunctive clauses

87

Donka F. Farkas Control in Autolexical Syntax

103

David Kathman On deriving the government constraint for incorporation and inflection

131

Steven G. Lapointe Some pleasures and pitfalls of Autolexical Syntax

189

Jerrold M. Sadock Toward a tri-modular analysis of -lyadverbs

207

Robinson H. Schneider Pronominal cliticization in West Flemish

241

Hans Smessaert T w o types of mismatch between morphology and semantics

291

Gregory T. Stump On restricting the role of morphology in Autolexical syntax

319

Anthony C. Woodbury Author index

365

Subject index

369

Introduction Eric Schiller

Autolexical Syntax represents a radical departure from the approaches to modeling grammar which were developed in the 1960s and became dominant in the 1980s. It employs a theory of parallel grammatical representations, w h e r e information is organized on a number of levels, known as dimensions. Each dimension contains information relating to a single aspect of grammar, e.g., syntax, morphology, logico-semantics. Every dimension has the power of a context-free phrase-structure grammar, although the overall power of the grammar is greater than that of a context-free grammar. The dimensions are static and do not involve derivations. Instead, the representations of each dimension are compared in the interface, w h e r e any discrepancies are checked against a set of interface principles. The information derives from the lexicon, where lexical entries list properties of each item at all of the various dimensions.

(1) v

Ν D

Ν Ν

Ν

Ν

the W

man

's

at

D I the

W

AF

W

W

W

Ν I door

Syntax

W Morphology

2

Eric Schiller

( Γ ) ' s 'is' Syntax: a verb which combines with a maximal projection to form a verb phrase.1 Logico-semantics: an operator which indicates present tense, combining with a predicate to create a one-place predicate. Morphology: an affix which attaches to a fully inflected word.2 A simple example (1), from Sadock (1991: 50) illustrates the Autolexical mechanism with the appropriate lexical entry in CI'). The syntactic and morphological dimensions are represented by a context-free phrase-structure grammar similar to that used in Generalized Phrase Structure Grammar (Gazdar et al. 1985), but with the important difference that many of the devices posited in that framework are not required in Autolexical Syntax since the dimensions are fully autonomous and need only account for the distribution of formatives. The graph illustrates a typical mismatch, with one morphological word containing two syntactic elements. The node labels used in the graph represent categories in an X system. The thick lines are used to connect nodes, with thin lines used to connect lexical items to nodes.3 One can think of the model as a parallel processing computer (and indeed the framework can be implemented in this fashion) with each dimension processed independently, and with the representations generated by each component compared by an independent routine (the interface) to see that the combination of representations is well-formed. Tinkering with the details of the processing of independent dimensions will not always affect the overall function of the model. Naturally some mechanisms for the internal workings of each dimension will be superior to others, either in terms of correctly modeling real-world data from natural languages, or in terms of efficiency of execution. There is no need to burden the individual dimensions with fancy apparatus, since by maintaining the autonomy of dimensions the internal workings of each should be simple and straightforward. The Autolexical framework allows a piece of information to be represented along some dimensions, while being entirely absent in others. Consider, for example, the case of the word it in It's raining. The word it is needed on the syntactic dimension, because English requires explicit subjects in all non-imperative constructions. Thus without the word, the sentence *Is raining is ungrammatical on purely syntactic grounds. Yet in terms of the

Introduction

3

logico-semantics (function-argument structure and variable binding), the word is not required, and it plays no role on that dimension; thus the representation would be as in (2). The logicosemantic representation contains two formatives. The element 's (the contracted form of the verb is ) is an operator which instantiates the present tense, which combines with a single predicate (raining) to form a proposition that it is presently raining. (2)

S Ν \ Ν

Syntax

I Ν I it

I N" Morphology

Logico-semantics

The syntax consists of an ordinary declarative sentence with a noun phrase (it) combining with a verb phrase containing a verb (i's) and a noun phrase draining). On the morphological dimension, the pronoun (if), a fully inflected ("') form, combines with an affix Cs) to form another fully inflected noun, while the verb stem (V"°) combines with an affix i-ing) to form a fully inflected verb form. All of this information is available at a glance, and nothing is buried in transformational steps. The properties of each lexical item are listed in the lexicon. The entries for the sentence represented in (2) would be as in (3): (3) Lexical entries for (2) a. it Syntax: ν Morphology: N"1 Logico-semantics: nil

4

Eric Schiller

b. 's 'present tense operator' Syntax: V Morphology: Af Logico-semantics: Ο c. rain 'rain' Syntax: Ν Morphology: V"° Logico-semantics: F"1 d. -ing Syntax: nil Morphology: Af: V[PresPart][v"° Logico-semantics: nil

1

These entries can be interpreted as follows. The first item (3a) is syntactically a pronoun (or, more precisely, a pro-noun phrase, since it substitutes only for complete noun phrases, and not for N-bars or simple nouns). Morphologically it is a fully inflected form.4 From a logico-semantic viewpoint, it plays no role whatsoever. The affix represented in (3b) acts as a verb in the syntax and as a present tense copula in the semantics. Morphologically, it combines with a fully inflected word to form a word. While the general spirit of Autolexical analysis is easy to demonstrate, discussion of the details of the framework are harder to articulate, because at the time of the workshop from which these papers are drawn, many of the fundamental questions were still unresolved. But fortunately, such questions were at the heart of the papers presented, and now made public.5 The first basic question is the number and nature of the individual dimensions. This question is touched upon in virtually every single paper. The role of phonological information in an Autolexical approach is by no means clear, and a seminar at the University of Chicago in the spring quarter of 1989 failed to achieve any consensus. One approach is that of Shobhana Chelliah's paper in this volume. One representation of semantic roles (theta-roles, thematic relations, "deep case", etc.) is discussed in the paper by Faarlund. Faarlund opts for an additional component with a tiered structure reminiscent of Autosegmental Phonology. In my own work, e.g., Schiller (1989, 1991), I have consistently preferred to treat the semantic roles as features on nodes in the logico-semantic component. I believe that this is a superior method, because I

Introduction

5

have yet to find any mismatch between semantic roles and the logico-semantic tree, and if the roles are treated as features which percolate throughout the semantic tree, the subcategorization arising from lexical entries is achieved at no cost. In either case, the lexicon must contain information about semantic roles. Faarlund adds another layer of complexity to the lexicon, while I place the information in the logico-semantic entry, e.g., cut is a three-place predicate which subcategorizes for two default items (subject and object) and a third argument which bears the instrumental role. As noted in Sadock's presentation of problems and prospects of the Autolexical approach, there are no strict criteria by which one can determine whether an additional dimension is justified. On the one hand, by chopping the grammar up into smaller and smaller pieces one does gain a certain efficiency and simplicity in each component, but at the same time the proliferation of dimensions complicates the analysis as the information contained in each dimension becomes more and more isolated. The approach taken by all of the contributors seems to be to let the data rule, and let real-world facts determine the appropriate direction to be taken. If there is a tendency to over-proliferate now, perhaps better analyses will reign in the grammar later. If solutions can be found within a more limited number of components, that is all the better. Yet there is one area in which there is now a consensus that a dimension ought to be split into two. The paper by Woodbury, together with Lapointe (1988), has encouraged a significant number of Autolexicalists (including Jerrold Sadock and this author) in the view that both morphosyntax and morphophonology are required. In some languages, e.g., the isolating languages of Southeast Asia and many Creole languages, these two dimensions will be isomorphic, but it is quite reasonable to see some of the differences among languages of the world as reflecting differences in the functional loads of some of the dimensions. Morphology is a field which has been of great interest to generative grammarians in the past decade or so, and the papers by Farkas, Lapointe, Smessaert, Stump, and Woodbury provide fascinating analyses of some complex problems in a variety of languages, while the paper by Schneider brings the Autolexical technique to bear on a classic problem of English morphology-that of -ly adverbs. The recent paper by David Kathman is a late addition to this volume, but in the opinion of the authors and other experts in Autolexical theory it was of such importance that not to include it

6

Eric Schiller

would have been an egregious error. Kathman presents a proposal, influenced by the handling of control phenomena wholly within the logico-semantic dimension, incorporating classical notions of c-command but avoiding the complications entailed in applying it to control facts on the syntactic level. Taken together, the papers of this volume provide a good sample of work undertaken in the early days of the development of the Autolexical framework. By the time these words are read, some of the fundamental issues raised here will have been resolved, while still other problems will no doubt appear on the horizon shortly. To bring the reader up to date (as of January 1995) is a difficult task, as Autolexical research is being carried out in a variety of institutions by a growing number of individuals. The following comments are based on a seminar held in the spring quarter of 1991 at the University of Chicago, papers presented at conferences up to and including the 30th meeting of the Chicago Linguistic Society in April of 1994, and reflects work in progress. The following are offered as food for thought, rather than as definitive statements on the direction of the research. One of the joyful exercises of the Autolexical research program is the re-evaluation of much of the classic literature of the past three decades. Many of the insights of the transformational frameworks have been lost, or set aside by more recently developed frameworks. In the seminar, classic accounts of raising transformations and the cycle were reconsidered, and it was shown that an Autolexical account handles the facts which motivated these analyses (while, it must be admitted, sharing some of the problems that remain unsolved in those frameworks). As has been suggested in other theoretical contexts, the "passive transformation" can be analyzed as a lexical rule which creates (or accesses pre-existing) participial forms which have lexical entries which govern surface combinatorics. There is no derivational rule, but the link between the meanings of passive sentences and active sentences can be demonstrated by rules of semantic interpretation involving λ operations of the type found in Montague Grammar. But there is a major difference between the two approaches, in that these meaning postulates are not part of Autolexical grammar, but are outside the grammatical framework and merely constitute logical proofs that the sentences are equivalent in meaning. Explicit analyses along these lines are contained in Sadock (1992).

Introduction

7

Questions of control have also been adressed in presentations by David Kathman (in this volume) and Jerrold Sadock. A reflexive semantic quantifier (p) has been introduced to handle both reflexivization facts and control phenomena. The intermodular default system is undergoing revision based on the acceptance of a specific semantic category for common nouns proposed in Gupta (1980). A case-checking mechanism is being developed for the interface, in ongoing work by David Kathman. Details of these and other innovations can be found in Schiller (1992). Other papers which have extended and modified the framework are papers by Schiller - Need (1992), Lapointe (1992), and Sadock (1992), all presented at the 1992 meeting of the Chicago Linguistic Society. An important modification to the overall framework was introduced in Sadock - Schiller (1993), presented at the 29th meeting of the Chicago Linguistic Society. In that paper, the various formulations of the homomorphism constraints were replaced by the Generalized Interface Principle, which states that "Paradigmatic and syntagmatic features at all levels should correspond as closely as possible." In addition, the paper argued that as a general rule, linear order is maintained when a conflict between it and scope relations arise. At the 1994 meeting of the Chicago Linguistic Society, Autolexical theory was applied to issues of variation in language, particularly dealing with different representations for alternatives available to speakers and possible relevance to historical change (Need - Schiller 1994). There are three completed dissertations which are written using Autolexical Syntax: Graczyk (1991), Leer (1991), and Schiller (1991). For perspectives from authors who work primarily within the Government and Binding approach, the papers from the 1991 meeting of the Chicago Linguistic Society by Yoon - Yoon (1991) and Mark Baker (1990) are of interest.The framework continues to develop at a rapid pace, and it is hoped that the publication of the papers in this volume will contribute to an understanding of the basic principles and issues of Autolexical Syntax. The editors wish to express their thanks to all of the participants in the 1989 Workshop on Autolexical Syntax. Special thanks to Tista Bagchi and Erica Zeinfeld for entering and formatting editing changes. The camera-copy was produced at t h e Linguistics Computing Facility at the University of Chicago.

Eric Schiller

8 Notes 1.

More formal notations will be presented in (3) below. These informal definitions are mine, not Sadock's.

2.

W and AF are informal devices to indicate fully inflected items and bound morphemes, respectively. More formal notations will be introduced below.

3.

As with many aspects of Autolexical theory, the choice of node labels, and indeed of the internal workings of each component in general, are to some extent up to the practitioner, although the work based at the University of Chicago has recently taken on a more homogeneous character with regard to notation. In the modern age, computerization has provided linguists with a wider range of typographical symbols, especially in Macintosh-based environments. Since the dimensions are autonomous, all that matters is that the dimension contain a representation of the information which can be accessed by the interface. The variety of approaches taken to such substantive matters as the number and types of dimensions are well-illustrated by the papers in this volume.

4.

I have indicated it as a noun, but perhaps it is a member of the closed morphological class of particles (P).

5.

Unfortunately, William H. Eilfort's paper on the need for an illocutionary dimension was not submitted for publication. A summary of the ideas can be found in Eilfort - Schiller C1990).

References Baker, Mark 1990

"Pronominal inflection and the morphology-syntax interface", in: Michael Ziolkowski - Manuela Noske - Karen Deaton (eds.), 25-48. Canakis, Costas - Jeanette Denton - Grace Chan Ceds.) 1992 Proceedings of the 28th Annual Regional Meeting of the Chicago Linguistic Society: General Session. Chicago: Chicago Linguistic Society. Eilfort, William - Eric Schiller 1990 "Pragmatics and grammar: Cross-modular relations in Autolexical theory", in: Michael Ziolkowski - Manuela Noske - Karen Deaton (eds.), 125-136. Gazdar, Gerald — Ewen Klein — Geoffrey Pullum - Ivan Sag 1985 Generalized phrase structure grammar. Cambridge, MA: Harvard University Press.

Introduction

9

Graczyk, Randolph 1991 Incorporation and cliticization in Crow morphosyntax. [Ph.D. dissertation, University of Chicago.] Gupta, Anil 1980 The logic of common nouns: An investigation in quantified modal logic. New Haven: Yale University Press. Lapointe, Steven G. 1988 "Distinguishing types of morphosyntactic cooccurrences: mismatch resolution, agreement, and government", in: Diane Brentari - Gary Larson - Lynn MacLeod (eds.), Proceedings of the 24th Annual Regional Meeting of the Chicago Linguistic Society: Parasession on Agreement in Grammatical Theory. Chicago: Chicago Linguistic Society, 181-201. 1992 "Life on the edge: arguments in favor of an Autolexical account of edge inflections", in: Costas Canakis — Jeanette Denton - Grace Chan Ceds.), 318-332. Leer, Jeff 1991 Schetic categories of the Tlingit verb. [Ph.D. dissertation, University of Chicago.] Need, Barbara - Eric Schiller 1994 "An Autolexical account of variation", in: Katharine Beals — Jeannette Denton — Robert Knippen — Lynette Melnar — Hisami Suzuki - Erica Zeinfeld (eds.), Papers from the 30th Regional Meeting of the Chicago Linguistic Society: The Parasession on Variation in Linguistic Theory. Chicago: Chicago Linguistic Society, 218-231. Pollard, Carl - Ivan Sag 1987 Information-based syntax and semantics Vol. 1: Fundamentals. CSLI Lecture Notes #13. Stanford: Center for the Study of Language and Information. Sadock, Jerrold M. 1991 Autolexical Syntax: A theory of parallel grammatical representations. Chicago: University of Chicago Press. 1992 "Cyclic rules without derivations", in: Costas Canakis — Jeanette Denton - Grace Chan (eds.), Proceedings of the 28th Annual Regional Meeting of the Chicago Linguistic Society: The Parasession on the Cycle in Linguistic Theory, 237-262. Sadock, Jerrold M. - Eric Schiller 1993 "The generalized interface principle", in Katharine Beals — Gina Cooke - David Kathman - Sotaro Kita - Karl-Erik McCulIough - David Testen (eds.), 391-402.

10

Eric Schiller

Schiller, Eric 1989

The case for Autolexical case. [Unpublished manuscript, University of Chicago.] 1991 An Autolexical account of subordinating serial verb constructions. [Ph.D. dissertation, University of Chicago.] 1992 The art of Autolexical analysis. [Unpublished manuscript, University of Chicago.] Schiller, Eric - Barbara Need 1992 "The liberation of minor categories: Such a nice idea!", in: Costas Canakis - Jeanette Denton - Grace Chan (eds.), 484-498. Yoon, James - Jeongme Yoon 1990 "Morphosyntactic mismatches and the function-content distinction", in: Michael Ziolkowsk — Manuela Noske Karen Deaton (eds.), 453-467. Ziolkowski, Michael - Manuela Noske - Karen Deaton (eds.) 1990 Proceedings of the 26th Annual Regional Meeting of the Chicago Linguistic Society: General Session. Chicago: Chicago Linguistic Society

An Autolexical account of voicing assimilation in Manipuri Shobhana L. Chelliah

1. Introduction In recent studies on the interaction of phonology with other components of grammar (Kiparsky 1983; Pesetsky 1985; Marantz 1988; and Cohn 1989, among others) it has been noted that the theory of Lexical Phonology (Kiparsky 1982, 1983; Mohanan 1986) does not make the correct predictions for certain combinations of affixes where the morphophonological bracketing of a word is at variance with the bracketing necessary for semantic interpretation of the word. In this paper, I will discuss cases of such bracketing paradoxes found in Manipuri1 and argue that such mismatches indicate that word formation and word interpretation should not be carried out in conjunction with each other. The Manipuri data will be reanalyzed in the Autolexical framework (Sadock 1985, 1988) which allows for autonomous morphophonological and semantic representations which are mapped onto each other by means of a restricted interface grammar. This analysis, which will express the correct morphophonological and semantic structure of Manipuri words, will be extended to give an account of two types of bracketing paradoxes observed in English.

2. Level ordering in the lexical phonology of Manipuri I will now present the data from Manipuri, giving first the phonological evidence present for the existence of level ordering in the lexical phonology of the language. Consider the voicing assimilation rule where syllable-initial voiceless unaspirated stops are voiced between voiced segments. As illustrated in ( l a ) , the Voicing Assimilation rule applies with the suffixation of the infinitive

12

Shobhana L. Chelliah

marker -pa when it is suffixed to a root that ends in a voiced segment. When the root ends with a voiceless segment the initial stop of the suffix does not become voiced. Similar examples are given in (lb-d) with the affixation of the genitive, dative and associative markers respectively. (1) a. stem+-pa 'inf b. stem+-ki 'gen' c. stem+-ta 'dat' d. stem+-Ara 'ass'

öabo 'to eat' thagi 'of the moon' finc/a 'on the hill' migo 'with man'

p/Apa 'to be small' phuritki 'of the shirt' lempaktg 'on the ground' khutkd 'with the hand'

The Voicing Assimilation rule also applies when the initial voiceless stop of a root occurs between two voiced segments in compounds. Thus in (2a) the initial voiceless stop of poy 'wander' voices when compounded with a stem that ends in a voiced stop. Similarly the Voicing Assimilation rule applies on the stem £in 'border' in (2b). (2) a.

Idmboybd lam-poy-pd

b.

path-wand er-agen 'wanderer'

kumjin kum-öin

year-border 'early part of the year'

The Voicing Assimilation rule, however, fails to apply with prefixation. For example, in (3a), with the affixation of the first person pronominal prefix i- the stem pa 'father' does not voice. Other examples of the Voicing Assimilation rule's failure to apply with prefixation are given in (3b) with the prefixation of the third person pronominal prefix ma-, in (3c), the attributive marker a- , and in (3d), with the prefixation of ma-'method of (verb)-ing'. (3)

root

pronominal prefix i-, ma- + root

a. pa 'father' b. pu 'grandfather'

ipa 'my father' mapu 'his/her grandfather'

c. afanba

d.

9-tdn-p9

att-short-inf 'that which is short'

mapa mo-pa

mode-read 'way of reading'

As shown in example (4a-d), the Voicing Assimilation rule also does not apply on compounds where the initial stem is the head 2

Voicing assimila tion in Manip uri

13

of the compound. So in (4a), the k of ka 'rise' does not voice even though the environment for the application of the Voicing Assimilation rule is present. Similar examples are given in (4b-d). (4) a. ika b. i-ka water-rise 'flood water'

mitop c. mi-top man-distinct 'stranger'

hdykon d. hay-kon fruit-place 'orchard'

lupow lu-p9w head-rough 'dandruff

The fact that the Voicing Assimilation rule applies only in a particular morphological environment can be neatly represented in the theory of Lexical Phonology where word formation processes and phonological processes are organized in a series of hierarchically ordered levels. In the Lexical Phonology model, the correct characterization of the Voicing Assimilation rule can be made by placing the morphological environments where the Voicing Assimilation rule applies on an earlier level than the environments where the rule fails to apply, and by pairing the Voicing Assimilation rule with this earlier level. This level ordering is shown in (5). (5) Level

Morphology

Phonology

Level 2

suffixation

VAR, other Level 2 phonology

Level 3

prefixation

Level 3 phonology

This level ordering assures that the application of the Voicing Assimilation rule is restricted to level 2 (L2) where suffixation takes place. The Voicing Assimilation rule will not apply at level 3 CL3) where prefixation takes place, since it is turned off at the end of its own level of application. In this way, the application of the rule to morphological environments created at the later level is blocked. Similar facts are given in (6) and (7). (6a) and (6b) give examples of the rule of Lateral Deletion, where I deletes after k (the 1 of the perfect aspect marker -la in [6a] and the derivational marker -lom in [6b] delete). A rule applies subsequently whereby intervocalic k changes to glottal stop.3 (6c) and (6d) are examples of where the Lateral Deletion rule fails to apply; here I (in -Idk) does not delete but surfaces as r. As illustrated in example (6d), the initial I of -la does delete under the influence of the preceding k of -lak. Characterizing the Lateral Deletion rule is thus complicated by the

14

Shobhana L. Chelliah

presence of the Velar Stop Deletion rule where k is deleted before 1. Examples in (6c-d) illustrate the application of the Velar Stop Deletion rule, where the k of -thok is deleted with the suffixation of the directional suffix -tek. (6) a.

c.

yo?db9 yok-h-po rear-perf-inf 'rear up'

b.

la?dmmi lak-ldm-li come-seq-prog 'carried here'

puthor9?i pu-thok-Iak-i carry-out-dsource-pres 'carries out'

d.

tor)thord?9gd £oi)-thok-ldk-l9-kd j ump-out-dsource-perf-ass 'jumping out'

However, the Velar Stop Deletion rule is seen not to apply in cases like (6a) and (6b) where instead, the Lateral Deletion rule applies. The Velar Stop Deletion rule also fails to apply in certain compounds as illustrated in (7a-c). (7) a.

thoklakpo thok-lak-p9 verb-verb-vinfl happen-come-inf 'to emerge'

b. k"ikloy c. öaklem khik-loy äak-lem noun-noun noun-verb thatch-bamboo rice-remain 'bamboo thatch' 'left over rice'

Consider the necessary ordering between the rules of Lateral Deletion and Velar Deletion. In the derivation of a form like öopt!1 or9?9g9 'jumped out', if the Lateral Deletion rule is assumed to apply before the Velar Stop Deletion rule, the incorrect form given in (8a) is derived. If the Velar Stop Deletion rule is assumed to apply before the Lateral Deletion rule, the incorrect form given in (8b) is derived. (8) a. /£orj-thok-lak-l9/ the Lateral Deletion rule applies * [£og-thok-3k-3] environment for the Velar Stop Deletion rule no longer available h b. /fcog-t ok-l9k-la/ the Velar Stop Deletion rule applies *[toi)-tho-l3-l3] environment for the Lateral Deletion rule no longer available

Voicing assimilation in Manipuri

15

Thus mere ordering of the Lateral Deletion rule before the Velar Stop Deletion rule or ordering of the Velar Stop Deletion rule before the Lateral Deletion rule will not derive the correct result. However, both rules do apply to the form. The formalism used to characterize the application of these rules must be able to ensure that the Velar Stop Deletion rule applies with the affixation of -lak but is "turned off" with the affixation of -/a Furthermore, the Lateral Deletion rule must be "turned on" only with the affixation of -Id. The correct characterization is possible in Lexical Phonology. First, we may note that the Velar Stop Deletion rule applies with the affixation of aspect morphemes -Id, -li but not with the directional morpheme -1 ak and that the Velar Stop Deletion rule applies with the affixation of the directional morpheme -ldk but not with the aspect marker -lo or in some compounds. In order to make certain that the Lateral Deletion rule does not apply with the suffixation of -lok, the rule can be paired with the affixation of the aspectual marker -la (and also certain directional markers on which it applies as listed in [11]) at a level different from that of -lok affixation. Similarly, in order to avoid the application of the Velar Stop Deletion rule to the aspectual and directional markers on which the Lateral Deletion rule applies, the Velar Stop Deletion rule can be paired with the affixation of -lak at a different level from the application of the Lateral Deletion rule. Since the level-ordered lexicon reflects the order in which morphemes are concatenated, an additional way to determine the level ordering of -ldk and -Id affixation is to consider the linear order of these morphemes. From a form like cor)tbord?9g9 it is clear that -ldk affixation occurs before the affixation of -Id. Thus, -ldk affixation is assumed to occur, along with the application of the Velar Stop Deletion rule, at a level before the affixation of -Id. A derivation illustrating the application of the Lateral Deletion rule and the Velar Stop Deletion rule following this level ordering is given in (9). (9)

earlier level

coi)-thok-ldk

later level

coi)tboldk-l9

toi)tholdk

C0Qth0l9k9

[ corjt h or3?9ga]

Velar Stop Deletion rule applies Lateral Deletion rule applies further morphology and phonology apply

16

Shobhana L. Chelliah

The level ordering of the rules of the Voicing Assimilation rule, the Lateral Deletion rule and the Velar Stop Deletion rule can now be determined. First, in the linear order of morphemes given in (10), -la is affixed before the infinitive marker -pa (10a) and before the associative marker -ka (10b); thus -la affixation must occur before the affixation of these markers. (10) a.

carabadi öa-la-pa-ti eat-perf-inf-ex 'if (I) eat'

b.

taraga ca-la-ka eat-perf-ass 'after eating'

The Voicing Assimilation rule, which applies with the affixation of the infinitive and associative markers, can occur at the same level as the Lateral Deletion rule. Also, as shown in the form £opthora?aga (6d), -lak is affixed after the derivational marker thok in the linear order. Since -lak must be affixed before the affixation of -la, I will assume that -thok and -lak are affixed at a level before the affixation of -la. Finally, the Voicing Assimilation rule must be blocked from applying with prefixation by allowing prefixation to occur after the application of that rule. Based on these facts about the interaction of morphological and phonological processes, the level ordering of the lexical phonology of Manipuri is as seen in ( l l ) . 4 (11) Level

Morphology

Phonology

LI

-1 ak, -thok

Velar Stop Deletion

L2

infinitive -pa case markers aspect -la, -11 directionals -la, -lu, -lam

Voicing Assimilation Lateral Deletion

L3

prefixation compounds on which Voicing Assimilation does not apply

L3 phonology

Voicing assimilation in Manipuri

17

3. Bracketing paradoxes in Manipuri This level ordering, while correctly pairing morphological and phonological processes, leads to bracketing paradoxes between the phonological and semantic structure assigned to words since in Lexical Phonology semantic interpretation is carried on alongside of word formation. It has previously been noted (Hoeksema 1985; Aronoff 1988, among others), that when compounds are inflected, the inflection attaches phonologically to only one stem of the compound and that this sometimes results in bracketing paradoxes. In Manipuri, bracketing paradoxes of this sort will result for any compound noun that is derived at L3 in order to escape the Voicing Assimilation rule, when this compound is followed by a case marker. Recall that case markers have voiced allophones and are affixed at L2. Consider for example the compound given in (12) with a corresponding phonological derivation given in (12'). (12) ikagi i-ka-ki

(12') L2

ka-ki

L3

ka-gi ikagi

noun-verb-casem water-rise-gen 'of flood water' (12") phonological bracketing: semantic bracketing:

Voicing Assimilation

[ikagi] [i+[ka+ki]] [[i+ka]+ki]

In (12), compounding does not occur at L2 since the L2 rule of Voicing Assimilation does not apply between the stems / and kac, however, the case marker must be affixed to ka at L2 to allow for the application of the Voicing Assimilation rule. Thus, as shown in (12"), the phonological bracketing for ikagi is [i+[ka+ki]] but the bracketing for correct semantic interpretation is [[i+ka]+ki] where the genitive marker is affixed to the whole compound. Bracketing paradoxes will also occur when noun stems which are modified by one of the L3 pronominal prefixes or verb stems derived from nouns by use of the attributive suffix must additionally take an L2 case marker. An example of such a paradox is given in (13), where the Voicing Assimilation rule applies with the suffixation of the dative marker -fa, requiring the phonological bracketing as indicated in (13'). As shown in (13"), the semantic bracketing is at variance with this phonological bracketing.

18

Shobhana L. Chelliah

(13)

mgyurndd md-yum-td 3PP-house-dat 'to his house'

(13') L2

(13")

phonological bracketing: semantic bracketing:

yum-td Voicing Assimilation yumdo md-yumdd [mayumda] [m9+[yum+t9]] [m3+yum]+t3]j

Is there a way of resolving the paradoxes given in (12) and (13) within the theory of Lexical Phonology? Kiparsky (1983) offers a solution for a bracketing paradox in English that is similar to the ones illustrated in (12) and (13) for Manipuri. Consider examples like ungrammatically. As a level 1 ( L I ) suffix, -ity attaches to grammatical first. Un-, an L2 prefix , must only attach to adjectives and therefore should not be able to attach to the noun grammaticality. How is un- able to peek through the noun and see the adjective stem? In Kiparsky's view these violations must be seen as exceptions to the Bracket Erasure Convention which stipulates that internal brackets are erased at the end of every level. Kiparsky's analysis is that, as the Projection Principle is relevant in morphology, subcategorization requirements of affixes must be met at every level. Thus, since un- subcategorizes for an adjective, the requirements of the Projection Principle override all other conditions and allow for un- to see the adjective stem in grammaticality However, such a solution is not effective in solving the violations given in (12) and (13). There is no subcategorizational restriction against a prefix being affixed to a fully inflected form, or any subcategorizational requirement for a stem to be compounded with a fully inflected word. Thus there is no motivation for reanalysis due to the subcategorization requirements of an affix and Kiparsky's solution to English violations is inapplicable in Manipuri.

4. An alternate analysis What is necessary to characterize the formed and semantic structure of Manipuri words is an analysis which, while representing the level-ordered interaction of morphology and phonology, allows for

Voicing assimilation in Manipuri

19

rules that semantically interpret affixes to apply independently of this level ordering. Such an analysis is made possible in the theory of Autolexical Syntax, which posits multiple representations for word structure in grammar. In Autolexical Syntax each component of grammar (syntax, semantics, morphology) is represented as an autonomous module, is unordered with respect to the other modules, and works on an independent set of principles. The lexicon is non-modular and contains structural information about the nature of an item in each of the modules. The representations given by each of the modules are brought together to form a grammatical expression of the language through an "interface" grammar. In this part of the grammar, representations produced by each module are checked for structural similarity. Although standard Autolexical Syntax has a working definition of the semantic, syntactic, and morphological components, it does not include a description of a phonological or morphophonological c o m p o n e n t . T h e d e s c r i p t i o n of t h a t component presented here is an extension of standard Autolexical Syntax. 5 Consider now how the Manipuri facts can be characterized in such a model. First, a specification is given in the lexical entries of affixes a b o u t t h e type of phonology t h e y undergo. T h e morphological entry will indicate whether the affix is a stem, prefix, or suffix. In Manipuri it will not be sufficient to list in the lexicon the stem or word level status of a morpheme as the phonology is sensitive not to whether an affix is derivational or inflectional but to whether it is a suffix or a prefix. This is illustrated in (14a) for the prefix 9-, in (14b) for the inflectional suffix -ta and in (14c) for the derivational suffix -Idk. (14) a. b. c.

Lexical entry for 9- 'attributive' morphology: prefix phonology: Level 3 Lexical entry for -td 'dative' morphology: suffix phonology: Level 2 Lexical entry for -Isk 'action comes from a distant source' morphology: suffix phonology: Level 1

20

Shobhana L. Chelliah

This additional information in the lexicon will not place an added burden on the language learner since predictable mappings between modules will allow for underspecified lexical entries. For example, in Manipuri the default phonology for a morphological prefix will be L3 phonology; the default phonology for a morphological suffix will be L2 phonology, but the suffix -ldk, will be specified as also having LI phonology. As stems may undergo LI, L2, and L3 phonology, they will arbitrarily be listed as having LI phonology. Although the default phonology for stems is LI, in cases where a stem is the initial stem and head of the compound, it will have to be specified as undergoing L3 phonology. This can be accomplished by adding a specification in the morphological lexical entry of the stem indicating that it may be used as a stem or a modifying prefix. Thus if the morphological representation of the word presents the phonology with a prefix, the stem will undergo L3 phonology, the default phonology for prefixes. Second, the semantic representation of the word is given in terms of the scope of affixes, where an affix X has scope over a constituent Y if X and Y are members of the same constituent and Y is a member of a constituent that X is not a member of. For example, in [[grammatical (y)]adjective ity (x)]noun, -ity has scope over grammatical as both grammatical and -ity are members of the constituent noun and grammatical is a member of the nested constituent adjective, of which -ity is not a member. Finally, the phonological module is a level-ordered representation of the phonological rules of the language. A phonological rule belongs to a particular level on the basis of the type of morphology it applies to. Thus in Manipuri, LI phonological rules will be those that apply to special verb derivation, L2 phonological rules will be those that apply to regular derivation and inflection and L3 phonological rules will be those that apply to derivational and inflectional prefixation. Consider, now, how the representations produced by the autonomous phonological, morphological and semantic modules are coordinated to represent a word of the language. For the representation of the word £oi)thord?dg9 'jumping out', the lexical entries will be as listed in (15). (15) a. Lexical entry for cop ']ump' morphology: stem phonology: Level 1

Voicing assimilation in Manipuri b. Lexical entry morphology: phonology: c. Lexical entry morphology: phonology: d. Lexical entry morphology: phonology: e. Lexical entry morphology: phonology:

21

for -thok 'to do something completely' suffix Level 1 for -tek 'action comes from a distant source' suffix Level 1 for -ID 'perfect aspect' suffix Level 2 for -kd 'associative case maker' suffix Level 2

The semantic structure of the word in terms of the desired scope of the suffixes gives a structure as shown in (16). (16)

tor) -thok -I9k-l3-k9 The morphological representation of this word, derived through phrase structure rules, is shown in the top half of the representation given in (17). The bottom half of the figure shows the level-ordered phonology of the language. (17)

W

Morphology

stem stem

Interface Phonology

root toy

suffix -tho

I

ι

toi)

suffix -ldk

22

Shobhana L. Chelliah

A set of rules (from one level) of the phonology is mapped onto a corresponding morphological environment following the association convention given in (18). (18)

Associate (linearly adjacent) affixes of the same level, build tree from one level to the next.

In (17), then, the linearly adjacent LI affixes and the stem are associated to LI phonology; the linearly adjacent L2 affixes are associated to L2 phonology. The tree is then built from the lower level, LI, to the next highest level, L2. How should these phonological trees be read? A morpheme dominated by a level node has undergone the rules represented by that node. So for example in (17), the suffixes -tbokand -19k are dominated by the LI node reflecting that they do undergo LI phonology whereas the suffixes -/a and -go are dominated by the L2 phonology node thereby reflecting that they do undergo L2 but not LI phonology. The fact that a morphological environment that undergoes LI phonology will also undergo L2 and L3 phonology is reflected in the fact that the LI node is dominated by both the L2 and the L3 node. Consider the representation of words like ikagi which had to be represented as bracketing paradoxes in Lexical Phonology. In (19), the lexical entries for the lexemes of the word are given; (20) gives the semantic structure of the word in terms of the desired scope of the affixes. (19) a. Lexical entry for i 'water' morphology: prefix phonology: Level 3 b. Lexical entry for ka 'rise' morphology: prefix phonology: no specification necessary c. Lexical entry for -ki 'genitive case marker' morphology: inflectional suffix phonology: Level 2 (20)

Voicing assimilation in Manipuri

23

Interface

Phonology

The top half of the figure in (21) gives the morphological representation of the word; the bottom half shows the level-ordered phonological rules of the language. By the association convention given in (18), each affix is associated with the phonology that it undergoes: i to L3, ka to LI and -ki to L2. The tree is then built from LI to L2 and then from L2 to L3.

5. Bracketing paradoxes in English Two types of examples have been presented in the literature which counter the prediction made by Lexical Phonology that LI affixes will always precede L2 affixes. Kiparsky (1983) and Pesetsky (1985) discuss examples like ungrammaticality where, as described in section 3 of this paper, because of the subcategorization requirements of affix un-, the bracketing of the word must be [un [grammatical ]]ity ] so that the LI suffix -/iy occurs outside of the L2 affix un-, as in (22). Second, Aronoff (1976) presents examples like readability and computerization where suffixes that affect stress placement (the LI suffixes -ity and -tion ) occur outside of suffixes that are neutral with regards to stress (the L2 suffixes -able and -ize). Given the level ordering of phonological rules and the autonomous representation of phonology and morphology as

24

Shobhana L. Chelliah

presented for the Manipuri data, the ordering paradoxes cited in English can be accounted for. Ungrammatically will be handled as in (22).

(22) Noun Morphology

Adjective Affix I

Affix

Adjective I grammatical

Interface Phonology

(23) The domain of a cyclic rule must minimally contain a stem. Lexical listings will specify that LI phonology applies with the stress affecting LI suffix -ity and the stem grammatical·, and that L2 phonology applies with the stress neutral prefix un-. By the association convention given in (18), the same-level, linearly adjacent -ity and grammatical are associated with cyclic rules of LI phonology. The L2 affix un- is associated with L2 phonology. The tree is then built up from LI to L2. Each projection from the node of the cyclic level marks the boundary of one application of the rules of that level. Along with the condition on the domain of cyclic rules given in (23), we can account for stress placement in the word. The projection of the LI node which associates the stem to LI phonology ensures that grammatical will receive word stress. The stress rule which is triggered into applying by the association of -ity, will, by (23), correctly take grammaticality as its domain of application and assign stress to it. Given that the phonology of affixes is specified in the lexicon and by the association convention as stated in (18), words with a phonological structure like readability can be dealt with as in (24). The suffix -ity and the stem read are associated to LI phonology; the suffix -able is associated to L2 phonology. The tree is then drawn from one level to the next: from LI read to L2 -able and

Voicing assimilation in Manipuri

25

from LI -ity to L2 -able . By virtue of the fact that the cyclic stem read is associated to the cyclic rules of LI, it will receive word stress. (24) Noun Morphology

Adjective Verb

Affix

Affix

Interface Phonology

When the stress rule is triggered by the affixation of -ity, it will, not being able to find an acceptable domain in which to apply, treat the m o r p h e m e immediately left-adjacent to it as a stem. In this way, -ability will correctly be assigned stress as if it were an independent phonological word. The observation that L2-L1 affix s e q u e n c e s are treated as individual words with regards to stress was first made by Aronoff - Sridhar CI987). They build a system for dividing up morphological words into the type of "phonological word" (what they call pword) created by the phonology out of -ability in example (24). Their system hinges on the distinction that can be m a d e on purely morphological grounds in languages like English, between t h e existence of two types of affixes: those that attach only to stems (stem affixes) and those that attach only to words (word affixes). Aronoff - Sridhar show that on the basis of this distinction, a set of linear principles can be devised to divide English words into four types of phonological atoms (stem, stem + stem affix, word affix + stem affix, stem + word affixfes]) 7 each of which would be stressed individually. The entire morphological word would then undergo the noncyclic rules of the language. Since phonological boundaries are assigned on the basis of linear principles which refer to the morphological and not phonological characteristics of affixes, Aronoff - Sridhar's analysis does not predict that affixes

26

Shobhana L. Chelliah

will occur in a phonologically determined order. They can thus account for cases like compartmentalization. In languages like Manipuri where affixes cannot be distinguished on the basis of morphological characteristics, 8 an attempt at assigning phonological boundaries on the basis of linear principles will result in a complicated informal statement: (25) a.

List of affix types: Type 1: special derivational suffixes Type 2: suffixation Type 3: prefixation b. Informal system for creating pwords: assign a pword (1), pword (2), or pword (3) boundary to the edges of a type 1, type 2, or type 3 affix respectively. c. Domain of phonological rule application: (i) level 1 rules apply to pwords of type 1 (ii) level 2 rules apply to pwords of type 1, and 2 (iii) level 3 rules apply to pwords of type 1, 2, and 3. d. Constraints on rule application: phonological rules must be ordered. e. Derivation of a word a. assign pword boundaries b. apply phonological rules f. Sample derivations where the superscripts (1,2, and 3) refer to pword boundaries 1, 2, and 3 respectively. Derivation of toi)thor9?9gd cop coi) coi) coi)

'thok' 'lak' 2ld2 2ko2 'tho' 'lok' 2la2 2kd2 W 'ldk' V 2g92 'tho' 'rdk' V 2ga2

Derivation of mayumdd 'yum' ΠΊΘ 'yum' 3ma3 'yum' 3md3 'yum' 3

ΠΊΘ3

3

'jumping out'

3

td2 2td2 2dd2 2dd2 2

assign pword boundaries apply prule 1 apply prule 2 apply prule 3

'to his house' assign pword boundaries apply prule 1 apply prule 2 apply prule 3

In accounting for Manipuri, there are two reasons for preferring the system that I have argued for over the one illustrated in (25).

Voicing assimila tion in Manip uri

27

First, in (25), specific statements have to be made to restrict t h e domain of application of phonological rules. Opposed to this, in an analysis as r e p r e s e n t e d in (17), t h e domain of application of phonological rules is automatically derived from the way in which the morphology is associated to the phonology (by the association convention [18]) and is reflected in the formal representation in terms of dominance of nodes. Second, t h e assignment of pword boundaries in (25) is totally unnecessary for Manipuri, given t h e fact that type 1 suffixes (special derivational) are always going to be assigned a type 1 boundary (undergo LI phonology), type 2 suffixes (suffixation) are always going to be assigned a type 2 boundary (undergo L2 phonology), and so on. The same information can be retrieved from a redundancy statement in the lexicon. Instead of creating pwords through a linear segmentation of words, it would be more economical to map phonological rules directly o n t o morphological representations and allow the phonology access to lexically listed information. Thus, although for languages like English it is possible t o express t h e interaction of m o r p h o l o g y and phonology on the basis of linear principles, in a language like Manipuri, t h e a u t o n o m o u s r e p r e s e n t a t i o n of morphology and phonology along with a convention for mapping phonological rules onto morphological elements is more appropriate. Furthermore, such an analysis can be easily extended to apply to English.

6. Conclusion In this p a p e r , I have shown t h a t t h e interaction b e t w e e n morphological and phonological processes in Manipuri can best be represented in a system where level-ordered phonological rules are associated to an independently represented morphological component. The analysis was extended to account for facts in English that were problematic for Lexical Phonology. Finally, I have argued that an approach which assigns phonological boundaries to morphological words on the basis of linear principles does not yield an appropriate analysis for languages like Manipuri.

28

Shobhana L. Chelliah

Notes 1.

Manipuri is a Tibeto-Burman language of the Kuki-Chin Group spoken in the central valley of Manipur state which is located in Northeastern India. The data presented in this paper are taken from Bhat Ningomba (1986), Devi (1979), Pettigrew (1912) and from my own notes and tapes gathered during fieldwork carried out in New Delhi in 1984, and in Manipur State and New Delhi in 1986 and 1987. I would like to thank Hyunsook Kang, Willem de Reuse, and Curtis Rice for helpful comments and discussion on earlier versions of this paper. Ideas about the ways to represent the interaction of phonology and morphology in Autolexical Syntax come directly from suggestions made by Anthony Woodbury. Responsibility for any shortcomings in the analysis is mine. Abbreviations used in the morphological analyses are: agen ass att casern dat dsource ex gen

2. 3. 4.

5. 6.

agentive associative attributive case marker dative action takes place from a distant source experiential genitive

inf INFL perf pres prog seq vinfl 3PP

infinitive inflection perfect aspect present tense progressive aspect sequential verb inflection third person possessive

The head of a word is defined as that affix which determines the category of the word. In these cases note that the final k of yok, lak, and -lokdo not undergo the Voicing Assimilation rule as that rule applies only on syllable-initial voiceless stops. Here for the sake of space, I limit myself to the discussion of three phonological rules. A more exhaustive analysis of the lexical phonology and morphology of Manipuri can be found in Chelliah (1990), where I discuss the level ordering of eleven phonological rules in Manipuri. See also Woodbury (1995), where a formalization of the morphophonological component using the terminology of Autolexical Syntax is given. I refer to LI derivation as special because the suffixes that undergo LI phonology are derived from etymologically related stems that are still present in the language and used as stems as well; (e.g., the LI suffix -ldk 'action comes form a distant source' and lak 'to go'; the LI suffix -thok 'do something to completion', and thok 'out'; the LI suffix -kh9t 'begin to V' and khdt 'out'). This might explain the unique phonology of these suffixes.

Voicing assimilation in Manipuri 7.

29

Level 1 and Level 2 affixation do not constitute two types of morphological processes like stem affixation or word affixation in English: for instance, suffixes from either level can attach to bare stems.

References Aronoff, Mark 1976 Word formation in generative grammar. Cambridge, MA: M.I.T. Press. 1988 "Head operations and strata in reduplication: a linear treatment", in: Geert Booij — J. van Marie (eds.), Morphology yearbook. Dordrecht: Foris, 1-15. Aronoff, Mark - S. N. Sridhar 1987 "Morphological levels in English and Kannada", in: E. Gussmann (ed.), Rules and the lexicon. Lublin: Redakca Wydawnictw Katolickiego Uniwersytetu Lubelskiego, 9-22. Bhat, D. N. S. - M. S. Ningomba 1986 A Manual of Manipuri grammar. [Unpublished manuscript, Manipur University.] Chelliah, Shobhana L. 1990 Lexical phonology and morphology in Manipuri. Linguistics of the Tibeto-Burman Area. 13.2: 27-72. Cohn, Abigail 1989 "Stress in Indonesian and bracketing paradoxes", Natural Language and Linguistic Theory 7: 167-216. Devi, P. Madhubala 1979 Manipuri Grammar. [Dissertation, University of Poona (Deccan College).] Hoeksema, Jack 1985 Categorial morphology. New York: Garland Publishing. Kiparsky, Paul 1982 "Lexical phonology and morphology", in: The Linguistic Society of Korea, Linguistics in the Morning Calm. Seoul: Hanshin Publishing Company, 3-19. 1983 "Word-formation and the Lexicon", in: Frances Ingemann (ed.), Proceedings of the 1982 Mid-America Linguistics Conference. Lawrence: University of Kansas, 3-29. Marantz, Alec 1988 "Clitics, morphological merger, and the mapping to phonological structure", in: M. Hammond - M. Noonan (eds.), Theoretical morphology. New York: Academic Press, 253-270.

30

Shobhana L. Chelliah

Mohanan, K. P. 1986 The Theory of Lexical Phonology. Dordrecht: Reidel. Pesetsky, David 1985 "Morphology and logical form", Linguistic Inquiry 16: 193246. Pettigrew, W. 1912 Manipuri (Mitei) grammar. Allahabad: Pioneer Press. Primrose, A. J. 1887 Manipuri grammar. Manipur: Government Press. Sadock, Jerrold M. 1985. "Autolexical Syntax", Natural Language and Linguistic Theory 3: 379-439. 1988 Unpublished class notes on Autolexical syntax. Department of Linguistics, University of Chicago. Woodbury, Anthony C. 1995 "On restricting the role of morphology", in this volume.

Autostructural analysis of semantic roles*

Jan Ter je Faarlund

1. Introduction In this paper I will adopt the view of linguistic organization which forms the basis of Autolexical Syntax CSadock 1985, 1987, and 1991) and Autosegmental Phonology (Goldsmith 1976 and 1989). In bringing together some basic tenets from the two theories, I will attempt to provide an account of semantic roles and argument structure as part of the semantic module of the grammar. I will follow Autolexical Syntax in viewing the semantic component as a distinct, autonomous module. And I will follow Autosegmental Phonology in operating with separate tiers that are associated according to certain principles. I will draw the following definitional distinction between modules and tiers: a module is an autonomous component of grammar which operates according to rules and principles that are not necessarily found in other modules; a tier is a layer of representation within a module. Thus within the semantic module there is a tier of semantic roles. The main purpose of this paper is to offer a first approximation to an account of the rules and principles that govern the association of semantic role tiers with the syntactic surface structure of the sentence. I will do this by examining a variety of different construction types in different languages. To this end, I will establish a formalism for representing semantic roles in terms of tiers within the semantic module. I will furthermore assume that the level of syntactic representation with which these tiers associate is the shallowest possible surface structure, which may be taken to be the phonetic form equipped with labelled syntactic brackets. In other words, no linguistic material which is not present in the syntactic surface structure is visible to the semantic module; and no linguistic elements occur in a different position relative to other elements when they are associated with semantic roles than what they do in the phonetic form. The most important implication of this is that a semantic interpretation combines with a syntactic structure without need

32

Jan Terje Faarlund

for other "deeper" levels of syntactic representation, or "empty categories".

2. The autonomy thesis In accordance with the modular view of linguistic structure, it has been suggested that semantics be treated not only as a separate module, but as a module independent of syntax (Faarlund 1985, Lapointe 1987: 181-182). In Faarlund (1985) I put this into practice by treating subject properties as originating within separate, sharply distinct modules, such as syntax, semantics, pragmatics, etc. That way it was possible to explain both why certain properties are shared by the subjects of all sentence types, and why others are found in certain types only. It was also possible to simplify somewhat the enormously complex rules and constraints that seem to determine the use of the passive in Norwegian. Jackendoff (1983, 1987) and Culicover (1988) propose an organization of grammar where an autonomous conceptual component is generated independently of syntax and then mapped onto a syntactic representation. I will follow this view of the relationship between syntax and semantics, as I will Jackendoff's rather programmatic statement that "the fundamental point, from which all else proceeds, is that thematic relations are part of a level of semantic/conceptual structure, not part of syntax" (Jackendoff 1987: 372). I will entertain what I consider a rather extreme autonomous view in trying to distinguish as strictly as possible between semantic and syntactic phenomena. "A grammar consists of a set of components, each of which characterizes a set of well-formed representations at one or more levels. Each component of the grammar has its own primitives, rules of combination, and well-formedness conditions" (Culicover 1988: 38). The organization of the syntax and the semantics into two discrete modules first of all means that the semantic structure associated with a linguistic expression such as a sentence is not derived from the syntactic structure of that sentence, be it the deep structure, the surface structure, or any level in between. Rather, the semantic representation is generated independently as a representation of the meaning associated with the expression, and whether the semantic representation is well-formed is determined within the

Autostructural

analysis of semantic roles

33

semantic module alone. Hence, a sentence is well-formed if its syntactic structure is in accordance with the syntactic rules of the language, if the semantic structure is interpretable, i.e., if it "makes sense" within the universe of discourse established by the context of that sentence, and if the two representations can be associated by means of general principles of some interface system. As a consequence, certain phenomena traditionally treated as part of the syntax, can now be shown strictly to belong to semantics. An obvious phenomenon of that sort is semantic role, or "theta-role". Sentences such as (la) and (lb) are syntactically equivalent, any difference between them resides in the semantic structure. (1)

a. b.

The student reads q uickly. The book reads quickly.

Both sentences consist of a NP followed by a finite verb followed by a manner adverbial. They are thus syntactically well-formed sentences of English. The fact that the NP in (la) expresses an agent and that the NP in (lb) rather expresses a patient are semantic facts. Any differences between (la) and (lb) should therefore be expressed in the semantic representation of the sentences.

3. Syntax and semantics as separate modules Through the history of transformational grammar, part of the implicit motivation for positing different levels of syntactic representation is semantic. Consider for example a rule such as "Dative Shift". Sue sent a card to Max. Sue sent Max a card.

(2)

a. b.

(3)

a. VP -> V NP PP b. VP -> V NP NP

In traditional generative grammar these two sentences are considered to be derived from a common underlying structure, 1 although the grammar of English can be shown to include phrase structure rules that generate both independently. The phrase structure rules (3a) and (3b) generate the VPs in (2a) and (2b),

34

Jan Terje Faarlund

respectively, and they are both needed to generate the structures in (4). (4)

a. Sue sent a card to Madison. b. Sue called Max a fool.

In (4b) no "Dative Shift" has applied. The motivation for deriving (2b) from a structure corresponding to (2a) and thus positing a level of representation different from the surface structure is semantic: (2a) and (b) are presumably synonymous, and the NP Max has the same semantic role in the two sentences. The two sentences (2b) and (4b) are considered to have different deep structures because the NP Max has different semantic roles in the two sentences. Although the importance of transformations and thereby the contrast between different levels of syntactic representation have been minimized in the current Government and Binding framework, the same semantic motivation now underlies the existence of traces and other empty categories in syntactic surface structure; semantic interpretation still needs access to invisible features. A passive sentence such as (5) contains a trace in object position which is coindexed with a NP in subject position. (5)

The cardj was sent f, to Max.

The motivation for the trace is to secure that the NP in subject position is interpreted as the object of send. Syntactically, nothing is required in that position, since the grammar of English does allow participles or adjectives followed by a PP also in sentences where there can be no trace: (6)

a. Sue has talked to Max. b. Sue is fond of Max. c. Sue is worried about Max.

As an objection to the above it may be argued that the trace in (5) is required not only for semantic reasons, but also for syntactic reasons, since the verb send takes a direct object: (7)

*Sue sent to Max.

Autostructural

analysis of semantic roles

35

However, (7) is syntactically well-formed, as is any English sentence of the form NP V PP. The unacceptability of (7) is lexical. It is the lexical specification of the verb send which is violated in (7), not the syntax of English sentence structure. The only lexical information directly relevant to syntax is the information about syntactic category. Thus the fact that send is marked +V in the lexicon is sufficient for the syntax to accept (7) as well-formed. As will be demonstrated in more detail below, (7) is discarded by the semantic module, where there is an argument tier with three positions, each requiring a NP in the syntactic structure to associate with. The view of the relationship between sentence-level semantics and syntax proposed here corresponds in a way to the Saussurean view of the linguistic sign. The sign consists of a content side and an expression side. They are separate and autonomous in the sense that they are made up of different sorts of material, and the one is not derivable from the other; yet they are interdependent in the sense that the one cannot exist without the other. A sentence comes with an expression and a content. The content, that is, the semantic structure, is not derived from the expression, the syntactic structure, or vice versa. The semantic representation is the representation of the content of the sentence according to the competence of native speakers. The linguist's task is to describe and explain the relationship between the two representations. And this of course is where the analogy with the Saussurean sign ends. The relationship between the expression and the content of a sentence is not arbitrary, except at the lexical level. That is precisely why it makes sense to try to establish general principles for the combination of the two modules. A syntax freed of all semantic and lexical concerns may of course be very simple. All it needs to do is to account for the possible concatenations of syntactic categories and their hierarchical structure. Note that in this way we also avoid the undesired redundancy between the lexicon and the phrase structure component. The way these two are conceived of, for example, in the Standard Theory of generative grammar, the same collocation restrictions were stated twice, in the lexicon and in the phrase structure rules. In the model proposed here, the redundancy is avoided by stating all such restrictions once in the semantic module in the form of an argument tier (see section 6). The degree of complexity of the syntactic module and the possible need for movement rules and other transformations are

36

Jan Terje Faarlund

partly language-specific and empirical questions, partly a question of descriptive adequacy. I will not discuss those questions here, merely state my point as clearly as possible by claiming that this syntax does not need to account for the relationship between such sentence pairs as (8a-b) or (8c-d). (8)

a. b. c. d.

Sue sent a card to Max. Sue sent Max a card. Sue kissed Max. Max was kissed by Sue.

Such relationships are taken care of by the semantics, where all the relevant generalizations are stated.

4. Semantic roles and features The semantic representation associated with the syntactic structure of a sentence includes at least one proposition. The semantic roles of the arguments in a proposition follow from the lexical content of the predicate word. Hence the different semantic roles of the NP in Max is sleeping and Max is running follow from the contents of the respective verbs and need not be stated independently in the grammar.2 On the other hand, when there are two or more arguments to a single predicate, the distribution of the roles among those arguments presumably follows from some general-partly universal and partly language-specific-principles. The fact that we interpret (9a) and (9b) differently must follow from some general rule of the grammar of English, which again may follow from some universal principle. (9)

a. Sue loves Max. b. Max loves Sue.

These rules and principles are the topic of the following sections of this paper. In this section I will define the entities that these rules and principles operate on, that is, semantic roles. They need to be distinguished in those cases where more than one of them occur in a given proposition. The important factor for the grammar to capture, then, are the distinctive features of the roles,

Autostructural analysis of semantic roles

37

not their actual material content. I will therefore offer a description of semantic roles in terms of a minimal number of features. In two argument propositions, the argument denoting the source of an action, or the place, the owner, the recipient, the person harboring the emotions, etc., will be said to have the feature "locus". The sentences in (10) have two-place predicates, and one of the arguments (in roman print) contains the feature "locus", the other one does not. (10)

a. b. c. d.

e. f.

Sue killed the spider. Sue owns a Volkswagen. Max received a card. Max loves Sue.

Sue lives in Madison. There is a spider in the bathroom.

The argument which denotes the actor, instrument or source which brings about the action denoted by the proposition, as in (10a), contains the feature "agent" in addition to "locus". T h e term does not imply that it has to be animate or intentional. I will assume that the semantic module includes a tier with the semantic roles linearly represented. Each proposition in the semantic structure has its own role tier. The order of semantic roles on the tier depends on their feature values: a role with the value "plus" for both features precedes one with the value "plus" for only one of the features, which again precedes one with the value "minus" for both features: (11)

[+agent, +locus] > [-agent, +locus] > [-agent, -locus]

The three roles specified in (11) may conveniently be given the names of "Agent", "Locus", and "Theme", respectively. I will use those terms informally as abbreviations of the feature clusters. In the formal notation of associations I will adopt the following convention: A stands for a role with the value "plus" for both features, L stands for a role with the feature values [-agent, +locus], Τ stands for a role with the value "minus" for both features, and F stands for any role with either value for either feature. T h e s e notations and terms can be summarized as in (12):

38

(12)

Jan Terje Faarlund

'Agent'

A

[+a,+l]

'Locus'

L

[-a,+I]

'Theme'

Τ

[-a,-l]

F

The advantage of a feature analysis is that in most cases a distinction between all three semantic roles is irrelevant; it is in fact needed only in three-place propositions. In one-place propositions no distinction is necessary, since the only semantic role will follow from the lexical content of the predicate verb, and the notation F can be used in those cases. In two-place propositions, certain distinctions need not be made when we use a feature analysis. (13)

a. Sue hit Max. b. Sue saw Max.

For example, in (13a), Sue is [+agent] and Max is [-agent]. We do not need to determine whether Max is "Locus" or "Theme". The exact character of the role follows from the lexical content of hit. Similarly, in (13b), it is easy to determine that Max is [-agent, -locus], and that Sue is [+locus]. We do not need to determine whether Sue is also [+agent].

5. Association between phrases and roles 5.1. General principle In English and other languages where morphological case-marking plays a minor part in the interpretation of sentences, the relative order of the NPs in surface structure determines their association with semantic roles. The basic principle governing associations between semantic roles and NP arguments in non-case-marking languages can be formulated as follows:

Autostructural

analysis of semantic

roles

39

Principle of Linear Association: Associations are between the first role on the semantic tier and the first NP in the syntactic surface structure, then between the second role and the second NP, and so forth. This principle yields the associations shown in (14): (14)

a.

F Max b. +1 Sue c. +a Sue d. A Sue

is

sleeping. -1 owns a Volkswagen. -a hit Max. L T sent Max a card.

5.2. Prepositions Prepositions also take objects and correspond to predicates in the semantic structure. Prepositions may correspond to two different functions in a semantic structure. They may be lexical expressions of two-place predicates. The object of the preposition generally has a semantic role determined by the content of the preposition. The other argument is the object, state, or event being located, etc. by the preposition. For example, in (15), the arguments of the preposition are the proposition Sue studied linguistics and the name Chicago. (15)

Sue studied

linguistics

in

Chicago.

In this function, the prepositional phrase places the event expressed by the rest of the sentence in space or time, or makes it dependent on some other circumstance, such as cause, condition, etc. The other function that prepositions have is to combine with the verb to express one, possibly complex, predicate. Thus in (16), the PP in Chicago does not locate some state such as Sue lives in a certain place. Rather, live in combines to form one two-place predicate, of which Sue and Chicago are the arguments. (16)

Sue lives in

Chicago.

40

Jan Terje Faarlund

There are some well-known syntactic correlates of this semantic distinction. The PP exemplified in (16) is a sister of V within V'; I will refer to it as a V'-internal PP, as opposed to the type in (15), which is dominated by a V"-node. The V'-internal PP can be replaced by the VP pro-form do so, whereas the higher, V'-external PP is left behind by do so, since it is not a sister of V. (17)

a. Sue lives in Chicago, and so does Max. (=lives in Chicago) b. Sue studied linguistics in Chicago, and Max did so in Kalamazoo. c. *Sue lives in Chicago, and Max does so in Kalamazoo.

Also, the external PP can prepose to the beginning of the sentence much more easily than can the V'-internal one. (18)

a. In Chicago, Sue studied b. ?In Chicago, Sue lives.

linguistics.

The prepositions in the V'-external PPs are higher predicates, and their objects are associated with semantic roles at a different level than the other types of arguments in a sentence. I will not discuss this in detail here, basically because it seems to me that it is a rather trivial matter. The following discussion of prepositions is concerned with the V'-internal ones. Although V'-internal prepositions also correspond to (parts of) predicates in the semantics and thus contribute in an important way to the semantic structure of a sentence, they are considered "case markers" in the syntax. Hence, NPs governed by prepositions are not dependent on their linear position for their association. Principle of Association with Object of Preposition: The role associated with a NP governed by a preposition is fully determined by that preposition. A NP governed by a preposition is thus allowed to "jump the line" to associate with the semantic role demanded by its preposition. (19)

+1

-1

Sue lives in Chicago

Autostructural analysis of semantic roles

41

Therefore in (16), the preposition in also demands the role of Locus for its NP. (For a more or less equivalent analysis of verbpreposition combinations within a different framework, see Marantz 1984: 18-19.). This is not to say that the role necessarily is the one that corresponds to some concrete, lexical meaning of that preposition. For example, although to has directional movement as its concrete lexical meaning, it does not have that meaning in, for example, belong to; still to also in that context selects Locus for its object. Consider the synonymous pair in (20). (20)

a. Sue owns the VW. b. The VW belongs to Sue.

The two sentences have the same semantic role structure, and since they are synonymous, the same role should be associated with the same NP in both sentences. In (20a) Sue associates with the Locus because it is the first NP from the left and there is no Agent in the sentence, in (20b) Sue is Locus because it is governed by to. In both cases the other NP must be content with what is left over, the [-agent, -locus] role.3 As an illustration of how this treatment of prepositions works, let us take another look at indirect objects and "Dative Shift". (21)

a. Sue sent a card to Max. b. Sue sent Max a card.

It was argued above that (21a) and (21b) can be generated independently by the phrase structure rules of English, and that the motivation for a transformational relationship between them is largely semantic. We can now express the shared features of the two sentences where they should be expressed, that is to say, in the semantics. Their common semantic structure can be represented as in (22). (22)

Proposition I— Pred cause

Agent Sue Pred come-to

Proposition J Locus Max

Theme a card

42

Jan Terje Faarlund

The verb send when followed by two objects or by a fo-phrase has the content of 'cause + come-to'. The tier in (23) corresponds to the structure in (22). (23)

...A...L...T...

In (21a) the preposition to demands the role of Locus for its object, and the other two roles are associated with the remaining NPs according to their linear position. In (21b) no preposition interferes, and the roles are associated linearly. (24)

a.

a

L T

Sue sent a card to Max.

b.

A

L

T

Sue sent Max a card.

The preposition by used in passive sentences seems to have something in common with to in that it also participates in what is traditionally considered a syntactic process (the passive will be discussed more fully below), and it occurs in one member of a synonymous pair. (25)

a.

+a

-a

Sue kissed Max.

b.

+a

-a

Max was kissed by Sue.

In (25b), by picks up the Agent, leaving the Theme role to the remaining NP in subject position, as desired. However, by may not necessarily require an Agent for its NP; consider (26).

Sue is liked by everybody.

So we see that by does not pick one particular role for its object; it picks the first role, even if it has the feature value [-agent].

Autostructural

analysis of semantic roles

43

5.3. Locative alternation The term "locative alternation" is used by Rappaport - Levin (1988) to refer to constructions of the type exemplified in (27) and (28) (their examples [8] and [17]): (27) (28)

a. b. a. b.

Jack sprayed paint on the wall. Jack sprayed the wall with paint. Mary emptied water from the tub. Mary emptied the tub of water.

According to Rappaport - Levin (1988: 19), an adequate lexical semantic representation of the locative alternation verbs ought to meet the following requirements: a. The near-paraphrase relation between the two variants must be captured. b. The linking of the arguments should be predictable in terms of their θ-roles. c. The affected interpretation of the goal as direct argument must be accounted for. Then they go on to show that it is difficult to meet all of these requirements using a representation which consists of a list of theta-roles. In an autostructural analysis, we are primarily interested in the semantic representation of entire sentences, not just that of the verbs. In the following I will try to demonstrate how a representation of the sentences in (27) and (28) within the present model can meet all of Rappaport - Levin's three requirements. I will consider all the PPs involved V'-internal. Thus the prepositions join with the verbs to form the predicates of the sentences: spray on, spray with, empty from, and empty of. Let us first consider the spray-sentences, (27). Both of the predicates, spray on and spray with take the roles of Agent (Jack), Locus (the wall), and Theme (paint). This takes care of requirement (a) about the near-paraphrase relation. As for requirement (b), let us reformulate that in the present framework as a question of whether the association of semantic roles and NPs is predictable. The preposition on takes a Locus role by virtue of its lexical content. 4 With takes a [-locus] role, that is Theme. In (27a), then, on requires Locus for its object, the wall; Jack, being the first NP, associates

44

Jan Ter je Faarlund

with the Agent, which is the first role. This leaves paint to associate with the remaining role, the Theme. In (27b), with requires the Theme for its object, paint; again, Jack is the Agent, and now Locus and the wall are left to associate. Requirement (c), which I take to refer to the fact that (27a) and (27b) are only near-paraphrases of each other, needs some further consideration. In (27b), the wall is the direct object of the verb, and its referent is therefore interpreted as being more directly affected than in (27a). We might ascribe this interpretation to the fact that it bears a direct relation to a predicate which denotes an action, that is, a verb, and not only to a preposition, which can never denote an action. More generally, a predicate with an Agent argument transmits an additional meaning of "affected" to its Theme argument. This applies not only to the "goal", the wall in (27b), but also to paint in (27a). All of the above applies equally, mutatis mutandis, to the emptysentences in (28). The most important "mutanda" are the prepositions. A fairly easy case can be made that of is also a theme preposition, as we see for example in nominalizations: the destruction of the city. I will also argue that from is a Locus preposition, since Locus is the role that denotes a place, or a movement to or from a place.5 So the tub associates with the Locus role in both (28a) and (28b). 5.4. Existential sentences In many languages there is a special construction type used to introduce new elements into the discourse. These constructions are often characterized by two specific features: the newly introduced element is not in the normal subject position, which means that in languages that have an obligatory, positionally defined surface subject, such as English, another element occupies that subject position; and the sentences often contain a locative adverbial. Sentence (29) is a typical English existential sentence. Such sentences are associated with a tier with two [-agent] roles on it, one of them being [+locus]. (29)

There is a spider in the

bathroom.

The locational NP, the bathroom, is marked by the preposition for the Locus role. The other role associates with the remaining NP,

Autostructural analysis of semantic roles

45

which is the new element in the discourse, the spider. An Agent role is out of the question here, so there is no role left for the dummy element there. This is also what we want, since it does not belong to the category NP, and hence is without nominal content. Consequently, it cannot be the bearer of a role in the described state.

5.5. Free word order and case-marking The notions of "free" or "fixed" word order refer to the order of elements in terms of their grammatical relations, not, for example, in terms of their discourse function. A language in which new information always follows given information is usually not called a fixed word order language. In a language with fixed word order, therefore, it is precisely the semantic roles of NPs that are associated with their relative positions. The system developed here so far, of course, only works to the extent that the language has a fixed word order. The next task now is to extend the system to work also for languages with different degrees of "free" word order. If in a language two possible orders exist of, say, subject and object, then one of the following two conditions seems to apply: either the language has some morphological mechanism such as case-marking or cross referencing to distinguish the NPs, or there is a preferred reading which comes out unless the context or the content of the NPs forces the opposite reading. Cases of the latter sort will be dealt with in section 9. In many languages the roles of NPs are determined by casemarking rather than by position. We have already seen how prepositions pick out their own role for their NPs regardless of their position in the sentence, leaving the remaining roles to unmarked NPs. We can easily extend this system to case-marked NPs by assuming that NPs are associated with semantic roles on the role tiers according to their morphological case, rather than according to their position. As an illustration, consider the German sentence (30). The morphological cases of the NPs are lexically selected by each verb. It is thus a lexical property of the German verb schicken that it combines with the three case forms nominative, dative, and accusative in the syntax. Thus it is also part of the lexical specifications of the verbs treffen 'meet' and begegnen 'meet' that

46

Jan Terje Faarlund

they combine with nominative + accusative and nominative + dative, respectively. (30)

A

Meinem

L Bruder

hat

der

Professor

Τ I

einen Brief

my-D brother has the-N professor a-A 'The professor has sent my brother a letter.'

geschickt.

letter sent

There are two possible systems of associations between the semantic role tier and morphological cases. One is asemantically based system where a certain role always and only corresponds to a certain case. For example, the nominative would always and only be associated with the Agent, the dative with the Locus, and the accusative with the Theme. Semantically Based Association: Association is between an NP marked with the case Cj and a semantic role Rj. The other system is syntactically based, where the cases are linearized in some language-specific order, and associations would take place in the same order, much like the way it works in noncase-marking languages. Syntactically Based Association: Given a language specific ordering of cases, associations are between the first role on the semantic tier and the first case, then between the second role and the second case, and so forth. The language-specific ordering of cases is based on empirical facts about the actual interpretation of case-marked NPs in each language. In most Indo-European languages the cases are ordered as follows: Nominative > Dative > Accusative > ... In the German example above, we cannot tell which system applies. There are three roles and three NPs in that sentence, so both the semantic role tier and the syntactic structure are "complete".

Autostructural

analysis of semantic

roles

47

It is probably impossible to find "pure" systems in any language in the sense that all associations between morphological cases and semantic roles follow either one or the other system. However, there seem at least to be languages which lean heavily towards one or the other strategy. Cebuano, a Philippine language, has a strongly semantically based case system. There are three cases, sometimes referred as nominative, genitive, and oblique (Bell 1979). The verb has four voices: active, objective, locative, and instrumental. The voice of the verb determines the semantic role of the nominative NP. For example, if the verb is in the locative voice, the NP marked with the nominative prefix has the locative or recipient (i.e., Locus) role.6 (31)

Basahan si Ana ug istorya ni Josefa. read-BEN NOM Ana OBL story GEN Josefa 'Josefa will read a story to Ana.' (or: 'Ana will be read a story to by Josefa.')

If the verb is in the instrumental voice, the nominative NP has the instrumental role. (32) labre ang abrilata sa sardinas. INS-open NOM can-opener OBL sardines Open the sardines with the can-opener.' If we consider the nominative prefix a topic marker rather than a case marker, the semantic basis of this case system becomes quite visible. The nominative (topic) prefix does not mark a semantic role at all, but a discourse function, and it is thus not part of the case system proper. The other two cases, genitive and oblique, show up on NPs that are not topics. Generally, the genitive associates with the Agent role, and the oblique with the Locus or the Theme. The occurrence of one does not depend on the occurence of the other in the sentence. In (31) there is a Theme, istorya, preceded by the oblique prefix, and an Agent, Josefa, preceded by the genitive prefix. In (32) there is only a Theme marked oblique, besides the nominative with an instrumental role. The Cebuano nominative or topic marker is different from an IndoEuropean nominative in that it is more or less neutral between the roles; it is not primarily an Agent marker, in fact there is, as we

48

Jan Terje Faarlund

have seen, another Agent marker in the language, the genitive case. Standard literary Arabic has three cases, nominative, accusative, and genitive. The genitive is only used after prepositions and with NPs modifying nouns. Only the nominative and the accusative are used for major sentence constituents. Except for certain marked construction types, the nominative always associates with the first role on the tier, and the accusative with the others. (33)

a.

Yaqrifu al-ustaadu altaaliba. knows the-professor-N the student-A 'The professor knows the student.'

b. Saa^ada al-ustaadu altaaliba. helped the-professor-N the student-A T h e professor helped the student.' Arabic therefore seems to have a more or less purely syntactically based case system. Most Indo-European case-marking languages seem to employ a combination of the two systems. Consider the following examples from Old Norse. (34)

a. Hon skyldi bera gl vikingum she-N should bring beer-Α vikings-D 'She should bring beer to the vikings.' b. Olaf konungr pakkadi henni vel. Olaf-N king thanked her-D well 'King Olaf thanked her very much.' c. Hallfredr setti bil saman. Hallfred-N set home-Α together 'Hallfred set up a home.' d. Νύ tekr hann hestinn. now takes he-N the-horse-A 'Now he takes the horse.'

In all of these sentences the nominative is associated with the Agent role; in (34a-b) the dative is associated with the Locus role, and in (34c-d) the accusative is associated with the Theme role. It is easy to find an endless number of examples from Old Norse where the dative and the accusative are associated like in (34),

Autostructural

analysis of semantic roles

49

and hard to find real counter-examples, at least beyond singular idiosyncratic usages that must be marked in the lexicon. So far the associations seem to be semantically based in Old Norse. The nominative, however, presents serious challenges to this generalization. Although it is associated with the Agent role as expected in (34), there are numerous instances of the nominative associated with other roles as well, as for example with the Locus in (35a) and with the Theme in copula sentences such as (35b). (35)

a.

Porbjgrn attl fe Thorbjorn-N owed money-Α 'Thorbjorn owed some money.' b. Hann var gerviligr. he was well-accomplished 'He was very skilled.'

litit. little

Old Norse therefore seems to employ both systems of association. Whenever there is a nominative in the sentence, it associates with the first available role, i.e., the Agent when there is one, otherwise with the Locus if there is one, and if neither of them is present, with the Theme. This account also explains the associations in passive sentences which correspond to active sentences with indirect objects. In this case the nominative is associated with the Theme, and not with the Locus, as we would expect. (36)

a

Fjorirhleifar brauds eru honum foerdir hvern dag. four loaves-N bread-G are him-D brought each day 'Four loaves of bread are brought to him every day.' b. Var peim gefinn dagverdr. was them-D given lunch-N 'Lunch was given to them.'

But if we assume that the semantically based associations take precedence over the syntactically based association of the nominative, we will get the desired interpretation of (36). Remember that in this theory the syntactic string including morphological suffixes such as case endings is generated quite independently of the semantic module with its semantic role tier. The string containing a nominative NP and a dative NP is a fait accompli, as it were, which the semantics has to deal with. If there is a way of associating the role tiers of the semantic module with the string of

50

Jan Terje Faarlund

morphemes, then the sentences are well-formed. In these examples, the dative associates first with the Locus role, since it is a semantically based association. Then what is left for the nominative is the Theme. This is the same thing that happens in the case of prepositions, which also pick "their" role first (cf. 5.2.). If the grammar of this language only allowed semantically based associations, then (36a) and (36b) would be ungrammatical, since there would be no Agent for the nominative to associate with. Unless semantic association took precedence over syntactic association, the nominative would associate with the Locus role; then there would be nothing for the dative to associate with, and the sentence would also be ungrammatical. If we adopt this approach, then other sentences without a nominative pose no problem either. Consider the following sentences without a nominative NP: (37)

a. Syfjar mik. becomes-sleepy me-A Ί am getting sleepy.' b. Minnir mik hinnar remembers me-A that Ί remember that woman.'

konunnar. woman-G

Again, the distribution of cases in the syntactic surface structure is not the responsibility of the semantic module. The rule about the association of the nominative of course only applies when there is a nominative available!7 5.6. NPs marked by cross-referencing In some languages, argument NPs may be identified by means of morphemes added to the verb. Such verbal morphemes agree with certain NP arguments in terms of class (gender), number or person. The English third-person singular ending -5 represents an impoverished version of this system. The -s is also highly redundant, since the subject NP in English can always be identified by its position. A more elaborate cross-referencing system is found in Swahili, where two argument NPs may be marked on the verb.8

Autostructural (38)

Msichana

analysis

of semantic

aliufungua

roles

51

mlango.

girl SUBJ-PAST-OBJ-open-IND 'The girl opened the door.'

door

In the verbal complex aliufungua, the initial a- indicates that the subject is singular and of the m/wa class of nouns (which is the class of msichana), -li- is the past tense morpheme, and the following -u- indicates that the object is singular and of the m/mi class (which is the class to which mlango belongs). The associations here are reminiscent of the syntactically based associations of some languages with case marking systems. The first role associates with the NP referred to by the subject-marker, and the second role with the NP referred to by the object-marker. That this must be the correct generalization for Swahili, becomes evident when we consider sentences with three arguments. (39)

Johni

alimpa

mkunga

John SUBJ-PAST-OBJ-give-IND nurse 'John gave the nurse a present.'

zawadi.

present

In this example, the object marker on the verb agrees with the noun mkunga, and not with zawadi. Mkunga is [+locus], and hence higher on the role hierarchy than zawadi, which is [-agent,-locus]. And as expected, it is marked as the second argument by the verbal morphology. This, then, is a parallel to the fact that indirect objects in English precede direct objects when they are not marked by prepositions: they are put in second position for the purpose of role assignment.

6. The argument tier We have seen that in many cases the most important property of a semantic role is not its label or feature composition, but its position relative to the other roles. If a sentence has no Agent, a [-agent] role in many respects plays the same part as an Agent would: it is left unassociated in the passive, or, as we have seen, picked up by the preposition by, as in (40a); it may be omitted with infinitives and gerunds, (b) and (c), and it may be left out in null-subject languages, as shown in the Spanish examples in (d). In the first

52

Jan Ter je Faarlund

example of each pair the proposition has an Agent, in the second a Locus as the first role. (40)

a. Sue was kissed (by Max). Sue was loved (by Max). b. It is hard to beat Max. It is hard to like Max. c. Bea ting Max is hard. Liking Max is hard. d. Compre un libro. bought-lS a book Ί bought a book.' Tengo un libro. have-IS a book Ί have a book.'

In addition, there is often a discrepancy between the number of roles in the semantic tiers and the number of NPs in the syntactic structure. Passive sentences, and infinitival and gerundive clauses are examples of structures that presuppose a role in the semantic representation without that role being associated with an overt phrase in the surface clause. Furthermore, the omission of the first roles in passives and in infinitives/gerunds seem to be two different sorts of phenomena, since there are such things as passive infinitives and passive gerunds, where a first role is lost in the transition from active to passive and a second role in the transition from a finite tense to a non-finite form. Thus in (41) the first and then the second roles are lost. (41)

a. +a -a Sue hit Max. b. +a -a Max was hit. c. +a -a Being hit (can be quite annoying).

In order to capture these and similar facts, we need to establish a separate tier between the semantic role tier and the syntactic structure. This tier, which in many ways is reminiscent of the skeletal tier in Autosegmental Phonology (Goldsmith 1989), I will call the "argument tier". The argument tier has exactly one position

Autostructural analysis of semantic roles

53

for each argument subcategorized by the lexical expression of the predicate. A sentence with a transitive verb is therefore associated with an argument tier with two positions, and so forth. Before we discuss the argument tier any further, some terminological groundwork is in order. In the semantics the basic unit is the "proposition", which consists of a predicate and a certain number of "roles". The roles are sets of feature values, and are referred to by the labels "Agent", "Locus", and "Theme". They occur on a semantic role tier in that order. The terms "first", "second", and "third" role are used to refer to roles in certain positions counting from the left on the semantic role tier, regardless of their labels. The semantic module also includes the argument tier which contains a certain number of "arguments". The arguments will be referred to by ordinal numbers, counting from the left. When there is a preposition present as part of the predicate, the argument governed by that preposition will be given the highest number; in other words it will count as the last argument. In the syntactic module the basic unit is the clause, which consists of (among other things) a predicate and a certain number of noun phrases. In languages which basically depend on case-marking, the NPs will be referred to by the name of their case (nominative, dative, accusative, etc.); in languages that depend mainly on word order, the NPs will be referred to by ordinal numbers counting from the left in the clause. When I want to refer to more than one module, or to these structures in general, I will use "sentence" in reference to the basic unit. The sentence has a predicate and a certain number of positions, which again are distinguished by means of ordinal numbers counting from the left. In order to get an idea of the general function of the argument tier, let us look at the sentences in (41) in more detail. The semantic role tier, the argument tier and the syntactic structure of (41a) are given in (42). (42)

Sue hit Max. Semantic roles: Arguments: Syntactic phrases:

+agent 1 1 Sue

-agent 2 2 Max

Sue is the Agent, the first argument, and the first NP in the sentence; Max is the second argument, and the second NP in the sentence. The associations are therefore straightforward, non-crossing, and

54

Jan Ter je Faarlund

one-to-one, between the semantic role tier, the argument tier, and the syntactic structure. Next, let us look at (41b). The full representation is given in (43). (43)

Max was hit. Semantic roles: Arguments: Syntactic phrases:

+agent

-agent 1 1 Max

This is a passive sentence, and an Agent is therefore presupposed in the semantic structure, but since the lexical predicate is a past participle, and past participles are adjectives, and thus intransitive, only the second role associates with the argument tier where it is then the first (and only) argument. The association between the argument tier and the syntactic structure is straightforward: the first NP is the first argument. Finally, the passive gerund in (41c): (44)

Being hit. Semantic roles: Arguments: Syntactic phrases:

+agent

-agent 1

Here there are no NPs in the syntactic structure, but the action described by the verb implies both an Agent and a non-agent, which are therefore represented in the semantics. Since the verb is in the passive, the first semantic role does not associate with the argument tier. The non-agent is the first argument of this verb in the passive. And since it is in the infinitive, the first argument may be left unassociated in the syntactic structure. (See section 8 on infinitives.) 6.1. Oblique arguments The content of the argument tier is determined by the lexical properties of the word that appears as the predicate in the syntactic clause. If there is a preposition in the clause, that preposition also determines which argument it associates with. Consider for example the verb-preposition predicate talk to, where part of the lexical specification of to specifies that it associates with a [-agent, +locus] role (45).

Autostructural analysis of semantic roles

55

The difference between a passive sentence like, for example, (41b) and the same sentence with a by-phrase, (46), is that in the latter the Agent role associates with an argument on the argument tier and a phrase in the syntactic structure. (45)

Sue talked to Max. Semantic roles: Arguments: Syntactic phrases:

+agent 1 1 Sue

-agent 2 to Max

As we know, by requires the first role for its NP, but this role does not associate with the first argument on the argument tier. The argument tier represents the argument frame of a passive verb plus a preposition, in this case (be) hit by, not that of an active verb. (46)

Max was hit by Sue. Semantic roles: Arguments: Syntactic phrases:

1 1 Max

2 by Sue

6.2. Double association There are verbs that generally seem to merge two roles into one phrase. One such verb is buy, where the first argument, besides being the Agent, is also the recipient of the object bought, and therefore associates with the Locus role. (47)

Sue bought a car. Semantic roles: Arguments: Syntactic phrases:

Agent^^Locus

Theme

1 1 Sue

2 2 a car

This verb may also combine with three phrases in the syntactic structure:

56

Jan Terje Faarlund

(48)

Sue bought her son a car.

Now there is only one argument position for each NP, and the Locus associates with her son according to the general rule. This analysis is supported by certain facts about subject control in infinitival relatives, discussed in a similar context by Ladusaw Dowty (1988). The first role of infinitival clauses associates with a NP outside the infinitival clause (see section 8.1). Generally, it associates with a NP that also associates with a Locus role in the matrix proposition. (49)

a. Sue gave Max a book to read. b. Max received a book to read.

In both sentences, Max has the feature [+locus] in the higher clause, being the first role of receive or the second role of give; and in both sentences, Maxis also the Agent of read, that is, the first role of read, [+agent], associates with the NP Max in the higher clause. In the case of buy, we see how the association of the lower [+agent] accompanies the association of the higher [+locus]. (50)

a. A L Τ +a Sue bought a book to read. b. A L T +a Sue bought Max a book to read.

Double association of a somewhat different kind is found with verbs of motion, such as run, where the first argument is both the Agent causing the event to take place, [+agent], and the object moving, [-agent]. (51)

+a-a Sue is running.

Some normally intransitive verbs of motion may also be used in sentences with two arguments. (52)

a. Sue is running the show. b. Max walked his dog. c. Sergeant Jones marched the soldiers back to the camp.

Autostructural

analysis of semantic

roles

57

Besides the obvious cases of transitive run, as in (52a), there are transitive uses of such verbs as walk and march with a clear causative element. In (52a) the associations are quite simple: Sue is [+agent] and the show is [-agent]. The other two examples seem to represent somewhat more complex structures. In (52b) both Max and the dog do the walking, and both are thus [+agent] and [-agent]. In addition, Max is the instigator or the causer of the event (given a normal man-dog relationship). The semantic structure of (52b) thus has a higher predicate "cause" with Max as the Agent. In addition, Max and the dog are both Agents and Themes of the two individual walking events. The associations are then as follows: (53)

Max walked the dog.

Semantic roles:

+agent [ent +agent

Arguments: Syntactic phrases:

1 Max

-agent

2 the dog

Of course we need a principled way of restricting this kind of association to the structures where they in fact do occur. Obviously, most verbs do not allow associations with such semantic structures. For example, Max ate the dog cannot mean that Max took the dog out for dinner. Furthermore, such associations are not generally found across languages. The Norwegian counterpart would, for example, be quite unacceptable. It seems, therefore, that this is lexically determined. Typical verbs of motion are generally intransitive and have double associations, as in (51). But in English they may also be transitive, as in (52a). The verb walk used as in (52b) combines these two properties, and the higher Agent associates with the same argument as the first Agent-Theme dyad.

6.3. Causatives A causative sentence differs from its non-causative counterpart in having one argument more. That extra argument is always an Agent. In the semantics, causative sentences are complex propositions, where the higher proposition has two arguments, an Agent and the lower proposition, which is the proposition expressed by the non-causative counterpart. For convenience, I will refer to the

58

Jan Ter je Faarlund

Agent of the lower propostion as the lower Agent, without thereby implying that the causative sentence has a complex structure at any level but the semantic. Languages differ as to whether or not their causative sentences are complex in the syntax, and whether they have one or two argument tiers. In English, causative sentences may be formed by means of the verb make and the infinitive of the basic verb. The causative counterpart of (54a) is (54b). (54)

a. Max ate the vegetables. b. Sue made Max eat the vegetables.

Make is a two-place verb whose arguments associate with a [+agent] and a [-agent]. In the case of causatives the latter argument is the proposition expressed by the embedded sentence. Sentences such as (54b) may therefore be analyzed as containing an embedded clause corresponding to (54a). In that case we also have two argument tiers. (55)

a. I. II. b.

l[+a]: l[+aj:

Sue Max

2[-a]: Max eat the vegetables 2[-a]: the vegetables

+a

-a9 +a -a Sue made [Max eat the vegetables].

Superficially, French causative constructions look like the English ones in that they are formed by means of the verb faire 'do'. Yet they are different in a number of respects; for one thing, they may involve the prepositions ä 'to' and par 'by'. (56)

a. Nicole Nicole 'Nicole b. Nicole Nicole 'Nicole c. Nicole Nicole Jean Jean 'Nicole

a fait partir Maurice. has made leave Maurice made Maurice leave.' a fait manger les legumes a Maurice. has made eat the vegetables to Maurice made Maurice eat the vegetables.' a fait ecrire une lettre ä has made write a letter to par Maurice. by Maurice made Maurice write a letter to Jean.'

Autostructural analysis of semantic roles

59

In (56a) the two NPs are the higher and the lower Agent, respectively. In ( b ) the a-phrase is the lower Agent, but in (56c) it is the lower Locus, whereas the par-phrase in ( c ) is the lower Agent. When these prepositions are used in non-causative sentences, a is used for the Locus, and par for the Agent. (57)

La lettre a ete ecrite a Marie par Nicole. T h e letter has been written to Mary by Nicole.'

If we look specifically at the surface marking of the causee, or the lower agent, we find that in (56a) it is a direct object, in (56b) it is marked as a dative or indirect object by means of the preposition a, and in (c) it is preceded by par. Turning now to a genetically and typologically unrelated language such as Turkish, we find the same pattern, (58). In (58a) the lower agent is in the accusative, in (58b) it is in the dative, and in (58c) it is marked with the postposition tarafmdan, which is also used for the agent in passive sentences. (58)

a. Mehmed gocugu aglatti. Mehmed-N child-Α cry-caused 'Mehmed made the child cry.' b. Mehmed gocuga yemegi yedirdi. Mehmed-N child-D food-Α eat-caused 'Mehmed made the child eat the food.' c. Hasan Mehmede kitabi Fatma tarafmdan Hasan-N Mehmed-D book-Α Fatma by göndertti. send-caused 'Hasan made Fatma send the book to Mehmed.'

This pattern is found in many languages throughout the world, as demonstrated in Comrie (1976). On the basis of facts from a great number of languages, Comrie establishes a general principle for the encoding of the lower agent. The lower agent is demoted down along a hierarchy to the first vacant role. The hierarchy is the same as the more general NP Accessability Hierarchy of Keenan - Comrie (1977): Subject -> direct object constituent

-> indirect object

-> other oblique

60

Jan Terje Faarlund

This principle and the NP Accessability Hierarchy account for the encoding of t h e lower agents in (56) and (58). In the embedded intransitive sentences there are no direct objects, so the lower agents become direct objects; in the embedded transitive sentences the agents are demoted to indirect objects; and in the embedded s e n t e n c e s with indirect objects, only the "other oblique constituents" are available. What remains unclear, however, despite Comrie's detailed and thorough analyses of a vast amount of data, is exactly what the nature of this demotion process is. What kind of grammatical process is it, how should it be explained, and why is it always the lower subject which is demoted? In an attempt to answer these questions, let us return to the French /a/recausatives. It has been demonstrated by many linguists who have written about the French /aire-constructions (e.g., Aissen 1974) that they are mono-clausal on the surface. 10 One type of data in support of this is the fact that clitic pronouns that are semantically part of the lower clause precede faire. They cannot intervene between faire and the main verb: (59)

a.

Nicole

le

fera

partir.

Nicole him make-FUT-3S 'Nicole will make him leave.' b. * Nicole fera le

leave

partir.

As a consequence, the French causative sentences have only one argument tier, with all the arguments represented at the same level, although in the semantics t h e r e are two separate, hierarchically ordered role tiers. In cases like this, the highest Agent is (universally?) the first argument on the argument tier, followed by the arguments of the lower sentence. The lower Agent is therefore the second argument. The representation of the semantic role tiers and the argument tiers of (56a-c) is as in (60a-c). (60)

a. Semantic role tiers: +a Argument tier:

+a Nicole a fait partir

-a F F Maurice.

Autostructural

analysis of semantic

b. Semantic role tiers: +a Argument tier:

roles

61

-a +a +a

+a

-a -a

Nicolea fait manger les legumes a Maurice.

c. Semantic role tiers: +a

-a A

Argument tier: A

A

L

L

T

T

Nicole a fait ecrire une lettre ä Jean par

Maurice

In order to account for the associations in French causative constructions in the correct way, we need to make certain assumptions about the prepositions involved. Recall the two principles of associations with case-marked NPs outlined in section 5.5, the semantically based association whereby a specified role associates with a specified surface case, and the syntactically based association, whereby the first roles associates with the first NP, etc., by a language-specific ordering of surface cases. One of the prepositions, a, associates semantically, that is, with Locus; if there is no Locus, it associates syntactically, that is, with the second argument. The other preposition, par, associates semantically with the second Agent (presumably by looking for an Agent starting from the right). In (56a) there are two Agents, which associate with the two NPs in the expected order. In (b) the preposition ä occurs, but there is no Locus in the semantic representation. Nevertheless, (56b) is a well-formed sentence, because ä may also have its object associate with the second argument (which in many cases is the Locus). However, in (56b) the second argument is the lower Agent. In (c) there is again a Locus for ä to pick for its object. The preposition par in this case picks the second Agent for its NP. This leaves the two bare NPs to associate with the remaining Agent and the Theme. Again, since French syntax has rules that generate the syntactic structures of (56) anyway, as in (61), no transformations are needed to move the lower subject after the verb or to insert a.

62

Jan Terje Faarlund

(61)

a. Nicole a voulu rencontrer Maurice. 'Nicole has wanted to meet Maurice.' b. Nicole a voulu donner les legumes ä Maurice. 'Nicole has wanted to give the vegetables to Maurice.' c. Nicole a voulu envoyer une lettre ä Jean par avion. 'Nicole has wanted to send a letter to Jean by air mail.'

As far as the Turkish causative constructions in (58) are concerned, the reader may now verify for herself that these rules also yield the following associations: (62)

a.

+a F Mehmed gocugu aglatti. Mehmed-N child-Α cry-caused 'Mehmed made the child cry.'

b.

+a +a -a Mehmed gocuga yemegi yedirdi. Mehmed-N child-D food-Α eat-caused 'Mehmed made the child eat the food.'

c.

A

A

L

Τ

Hasan Mehmede kitabi Fatma tarafmdangöndertti. Hasan-N Mehmed-D book-Α Fatma by send-caused 'Hasan made Fatma send the book to Mehmed.' In Turkish, as in many other languages of course, some verbs take their direct object in the dative rather than in the accusative. One such verb is ba§malak 'begin, start'. When these verbs are causativized, the lower Agent appears in the accusative. (63)

a.

Qocuk okula ba§ladi. child-N school-D started 'The child started school.' b. Fatma gocugu okula balatti. Fatma-N child-Α school-D start-caused 'Fatma had the child start school.'

This means that the verb 'begin' is understood as combining with a [+locus] role, which in fact is not all that far-fetched when we

Autostructural

analysis of semantic roles

63

consider the semantic closeness of 'begin, start' and for example 'go to'. The Locus role therefore associates semantically with the dative case. This leaves the accusative case and the lower Agent to associate with each other. Comrie (1976) discusses instances of doubling on a certain grammatical relation. In his framework, that means that the lower agent is not demoted all the way to the first vacant relation, but is expressed in the same case as some other relation in the embedded sentence. An example of such doubling is found in Turkish, where the lower agent of a three-place verb may appear in the dative, and not with the postposition tarafmdan, which is the most common type. Instead of (58c) we may therefore find (64)

Hasan

Fatmaya kitabi

Mehmede

göndertti.

Hasan-N Fatma-D book-Α Mehmed-D send-caused 'Hasan had Fatma send the book to Mehmed.' This can also be captured by the present approach; of the two datives, one associates semantically with the Locus (Mehmede), the other assoicates syntactically with the second argument, which is the lower Agent (Fatmaya). Another kind of double association is found in modern Hebrew, where causative sentences may exhibit two accusative phrases. 11 (65)

a.

hatalmidim

rakdu

et

harikud

haxadas.

the-students danced A the-dance the-new 'The students danced the new dance.' b.

hirkadeti

et hatalmidim

et harikud

haxadas.

caused-dance-lS A the-students A the-dance the-new Ί made the students dance the new dance.' (The accusative is marked by the preposed particle et.) The Agent associates with the nominative (or—modern Hebrew being a null subject language-is left unassociated as in the present example), and the remaining (lower) Agent and the Theme associate with the two accusatives. Cole (1976) argues that only the first of the two accusatives, the lower Agent, is a real direct object, since it alone can become the subject of a passive sentence. The other accusative NP is a chömeur in the terminology of Relational Grammar.

64

Jan Terje Faarlund

(66)

a. hatalmidim hurkedu et harikud haxadas. the-students PASS-cause-dance A the-dance the-new 'The students were made to dance the new dance.' b. * harikud haxadas hurkad et hatalmidim. the-dance the-new PASS-cause-dance A the-students

Since (66a) and (66b) are syntactically identical, the only difference being the lexical items in the argument positions, one of them cannot be syntactically unacceptable while the other one is syntactically well-formed. The difference between them must lie somewhere outside the syntax. The problem with (66b) is of course a semantic one. It is unacceptable paired with the same meaning as (66a), in other words as a passive version of (65b). In a nonderivational model, the fact that (66b) cannot be associated with the same semantic structure as (66a) must be expressed in another way. In the present framework that can be done quite straightforwardly. In passive sentences, the first NP (in modern Hebrew the one that is morphologically unmarked) associates with the second role or the first argument, in this case the lower Agent. The accusative NP associates with the remaining argument, the Theme. That gives the desired reading for (66a) and a nonsensical reading for (66b).

7. Unassociated roles: Passive An unassociated role is one that is represented on one of the semantic role tiers but not on the argument tier. This is the situation where some participant is understood or implied, but where the predicate in the syntactic structure does not allow that participant to be expressed. The most obvious case of such non-association is the Agent or first role in passives without a by-phrase. In a passive sentence, a past participle expresses the predicate, and the sentence has one argument NP less than the active counterpart. Still, there is often a clear meaning difference between a passive sentence and some corresponding active sentence with a more or less synonymous predicate. Consider for instance the following pairs:

Autostructural

(67)

a. b. c.

analysis of semantic roles

65

The dog died. The dog was killed. Max fell off. Max was pushed off. Sue heard about the incident. Sue was informed about the incident.

We can capture this difference if we assume that the passive clauses associate with a semantic role tier where there is also an Agent role present, and that the Agent does not associate with any argument on the argument tier. The participle is a predicate with one argument less than the corresponding active verb. Consider the role tiers and argument tiers of the following sentences: (68) a. Semantic role tier: Argument tier:

-a -a The dog died.

b. Semantic role tier: Argument tier:

+a

-a -a The dog was killed.

c. Semantic role tier: Argument tier:

+a +a

-a -a

Max killed the dog.

However, not only Agents are blocked from associating with the argument tier in passive sentences. Sentences without Agents also occur in the passive. (69)

Max was seen shortly

afterwards.

Here the missing NP is the Locus, the "seer". So the rule applies to the first role, not necessarily to the Agent. Let us see how other types of passives behave with regard to this association rule. Consider first the passive of double-object verbs. The two arguments that are expressed bear the Locus and the Theme roles. In (70a) and (70b), Locus associates with the first argument. This gives us the desired interpretation in (70a) but not in (70b). Therefore (70b) is unacceptable on the most natural reading, but of course if we imagine some Alice in Wonderland reality where the meaning corresponds to how the roles associate in (70b), then it is perfectly acceptable.

66

Jan Terje Faarlund

(70)

a. +1

-1

Max was shown the document. b.

+1 document +1 -1

?77ϊβ

c.

-1 -1 was shown Max.

The document was shown to Max. d.

+1

-1

* Max was shown the document to. There is nothing syntactically wrong with (70b)-it has exactly the same syntactic form as ( 7 0 a ) - t h e problem is one of selectional restrictions, and therefore semantic. In (70c) one of the NPs is governed by a preposition which demands the Locus role for its object, Max. The other role, Theme, therefore must associate with the remaining NP, the document. The fourth possibility, (70d) is unacceptable for syntactic reasons. Presumably the syntax of English does not allow a stranded preposition following a NP.12 If, however, the preposition follows the verb immediately, the sentence is acceptable.

(71)

a. Max was never listened to. b. His words cannot be relied upon. c. This bed was slept in (by George Washington).

Again, my claim is that no empty category, such as an invisible trace, is needed in the syntax or semantics of such sentences. As usual with passives, the first role does not associate with any argument; there is one additional role, which associates with the only argument on the argument tier, which again associates with the only NP in the syntactic structure, regardless of its actual position.

8. Unassociated arguments Unassociated arguments are positions on the argument tier that do not associate with a NP in the corresponding clause. The reason is usually that the NP with which it should associate is missing from the clause for some (syntactic) reason. However, the

Autostructural

analysis of semantic roles

67

lack of association between some argument on the argument tier and some NP is precisely the kind of situation that would filter out sentences as ill-formed, or that would call for special measures allowing associations with NPs in non-argument positions (see section 9). In order to keep the mismatch between the argument tier and the syntactic structure as a powerful constraint on wellformedness, we need to specify for each language precisely those instances where such a mismatch is permitted. In English and many other languages, the typical case of an unassociated argument is the first argument of a non-finite verb. The justification for claiming that the argument is present on the argument tier is that it is precisely the typical subject position in the syntactic structure that is left empty in infinitival and other non-finite clauses. Furthermore, there are passive infinitives, where the second r o l e - w h i c h has become first a r g u m e n t - i s left unassociated. This was illustrated in section 6. 8.1. Infinitives and gerunds In generative grammar the non-association of the first argument (the subject) is expressed by means of the abstract element PRO, which may be coindexed with some other element in the sentence. (72)

Maxi wants PROj to sing.

Again, the co-indexing is necessary for semantic reasons. There is nothing in the syntax per se that requires an element between want and the infinitive in (72). The indexed PRO is there to ensure that the role of subject of the infinitive is filled by the NP Max. With an autostructural approach, this is unnecessary, since the role of subject (that is, first argument) is on a separate tier, and therefore has no independent representation in the syntax; it associates directly from the argument tier to the desired NP. (72) is a complex proposition associated with two argument tiers: (73)

Want: Sing:

...+1 -1... ...F...

The highest argument tier contains two arguments, with a Locus role and the Theme role. The Locus associates with the first NP and the Theme with the infinitival phrase to sing. The lower

68

Jan Terje Faarlund

proposition has a one-place predicate, sing, but there is no NP in the infinitival clause for the argument to associate with. On the other hand, the first argument of the argument tier of the infinitival clause can associate with NPs that also associate with arguments from the other tier. Thus in (72), Max associates with the Locus argument of want and with the argument of sing. The result of these associations is shown in (74). (74)

+1 F

-1

Max wants [ to sing].

In English an infinitive may be preceded by a /or-phrase. In that case t h e object of for is available for association with the infinitive argument. (75)

+1

-1 F

I would hate [for her to fall off].

This applies also to other for-phrases other prepositions as well. (76)

a.

and indeed to objects of

+1 -1 +1

-1

It is easy for Sue [to understand

b. A

L T +a

the

problem].

-a

Sue left it to Max [ t o feed the

fish].

Easy is a two-place predicate, where the Locus denotes the person

for whom something is easy. So in (76a), Sue is both the Locus of easy and the first role, also Locus, of understand.13 In (76b), Max is the Locus of leave to and the Agent of feed. In general, the first argument of a nonfinite verb form associates freely outside its clause, or it does not associate at all, as in: (77)

a. b.

It is dangerous to swim here. Swimming here is dangerous.

The question of how first arguments of infinitives associate with NPs in the matrix sentence is a complicated one which has been

Autostructural analysis of semantic roles

69

the subject of much discussion and controversy. I am not going to discuss this question in detail here, since it has little bearing on the overall model that is being presented in this paper. In any syntactic framework or theory, it seems, some general principles are needed independently of the formalism of that theory, to account for the various control possibilities of first arguments of infinitives. One of the most insightful and balanced treatments of the question to my knowledge is Comrie (1985). Comrie's discussion is theoryneutral, and his conclusion is that the control is partly syntactically and partly pragmatically determined. The basic syntactic principle is that of "Minimal Distance", proposed by Rosenbaum (1967), and the basic pragmatic principle derives from speech act theory (Searle 1969). In conventional generative grammar, a class of verbs is characterized by the property of allowing "Raising to Subject". In the present framework, the equivalent property is "Role Spreading". Consider first sentences with seem, which allows three different syntactic constructions. (78)

a. It seems that Max is sick. b. Max seems to be sick. c. Mcix seems sick.

Assuming that the sentences in (78) are synonymous, they are all associated with the same semantic structure, with a one-place predicate seem which combines with one argument, the proposition Max is sick, where sick is the predicate and Max is the argument. There are two arguments, then, at two separate argument tiers. Since all the sentences in (78) also have the same lexical predicate at the two levels, seem and sick, the argument tiers are the same for all three. The differences, then, are in the associations with the different syntactic structures. In (78a) both clauses are complete sentences, and the two roles associate within their respective clauses, one with the r/iaf-clause and one with Max. In ( b ) the argument of seem associates first with the first NP in the syntactic structure, Max, and spreads across the verb to the infinitival clause. The predicate be sick is in the infinitive, and the first argument has to look outside the clause for something to associate with. This it finds in the first NP Max. In ( c ) both arguments have to associate with Max, since it is the only NP in the sentence. This can be illustrated as in (79):

70

Jan Ter je Faarlund

(79)

a.

F14 It seems that Max is sick.

b.

F F

F

Max seems [ to be sick].

c.

F F Max seems sick.

The structure in (79b) is also found with the passive of certain "Raising-to-Object" verbs, such as believe, say, claim, etc. (80)

Max is believed

to have sung "White

Christmas".

The passive of believe has one position on the argument tier, and that is the embedded sentence. In (80) the argument associates with the first NP, Max, and spreads to the infinitival phrase, just like in (79b). Consider finally the so-called "Tough Constructions", as exemplified in (81). (81)

a. b. c.

Honest people are hard to find. The car is comfortable to ride in. The box is heavy to lift.

There is a real question as to what the semantic representation and the argument tier of such sentences would look like. Starting with the semantics, it is clear from the few examples given in (81) that the degree to which the adjective can be said to be predicated of the first NP varies considerably (82). On the other hand, there is also variation as to whether the adjective can be predicated of a gerundive phrase (83). (82)

(83)

a. b. c. a. b. c.

? H o n e s t people are hard. The car is comfortable. The box is heavy. Finding honest people is hard. Riding in the car is comfortable. ? L i f t i n g the box is heavy.

We see, therefore, that the constructions known as "Tough Constructions" vary from those where the adjective is predicated

Autostructural

analysis of semantic roles

71

of the object of the embedded infinitive to those where it is predicated of the whole infinitival phrase, with a grey area in between where both interpretations are possible. What they do have in common, however, is that the infinitival phrases, to find, to ride in, to lift, etc., are incomplete. There is an object missing. Find is a verb that always requires an object, compare *We went out to find. Furthermore, the infinitive acts like a modification to the adjective: the car is comfortable as far as riding in it is concerned (but not necessarily when it comes to making love in the back seat of it). Hence I will assume that the infinitives in (81) are not real infinitival phrases, so they cannot receive semantic roles. They are modifiers to the predicate adjective. This analysis is supported by phrases like an easy to open jar, where an adjective plus infinitive together function as an adjectival phrase. If we take the predicates in (81) to be hard to find, comfortable to ride in, and heavy to lift, then there is no problem assigning the one and only argument to the first NP. (84)

F Honest people are hard to find.

The first argument of the infinitive is left unassociated, but note that as soon as there is something for it to associate with, it does. (85)

The car is comfortable for tall people to ride in.

Whenever a /or-phrase is added to the adjective, the object of for associates with the first argument of the lower tier. This is a formalization of the implication carried by (85) that cars are comfortable to ride in for whoever does the riding. Chomsky's well-known minimal pair in (86) below is based on the fact that please is optionally transitive when used in the infinitive, so that both to please and to please John may be considered complete infinitival phrases, and therefore to please may or may not be considered a complete infinitival phrase. (86)

a. John is eager to please. b. John is easy to please.

In eager to please, the infinitive is considered complete, and the first role of please associates with John. Easy to please, on the

72

Jan Terje Faarlund

other hand, is a complex predicate, where the infinitive modifies the adjective, cf. John is an easy to please guy. Whether or not a sentence is to be interpreted as (86a) or as a "Tough Construction" as in (86b), depends on two factors: whether the infinitival phrase is complete or not, and whether the predicate adjective is one that can possibly be predicated of an open proposition. 15 Since please, as we have seen, may be used transitively or intransitively, the interpretation of (86a) and (86b) really depends on the meaning of the predicate adjective. In German the verb lassen 'let', past tense liess, takes infinitival complements with one or two accusative NPs:16 (87)

a.

Grethe

liess

mich

den

Hund

Grethe let me-A the-A 'Grethe let me pet the dog.' b.

Grethe

liess

den

Hund

Grethe let the-A dog 'Grethe let the dog run' c.

Grethe

liess

den

Hund

dog

streicheln

pet

laufen

run streicheln

Grethe let the-A dog pet 'Grethe let the dog be petted.' In (87a) the first accusative, mich, associates with the second argument of lassen. The second accusative, den Hund, associates with the second argument of streicheln. The first argument of streicheln looks outside its clause for a NP to associate with, and the first NP it finds is mich. This gives us the same interpretation as the syntactic equivalent in English, Grethe let me pet the dog. In (87b) the associations are the same, except that laufen, representing a one-place predicate, has no second argument to associate with an accusative phrase. The first argument of laufen associates with the nearest NP to the left, den Hund. The interesting case, however, is (87c), where there is again a transitive verb in the lower clause. There is a second argument that needs to associate with something, and since argument NPs regularly precede nonfinite verbs in German, the NP it would be expected to associate with is the one immediately preceding the verb, den Hund. This construction leaves nothing for the lower first argument to associate with, and thus it is left unassociated. The English translation of (87c) is 'Grethe let the dog be petted'. (For a comparison of constructions like [87c] and passives in German, see Santorini 1985.) It must be considered a lexical property of lassen that the

Autostructural

analysis of semantic

roles

73

lower first argument cannot associate with the first NP, Grethe, as is the case with, for example, want. The syntactic structures in C87) thus associate as follows: (88)

a.

b.

c.

A

L T +a -a Grethe Hess mich [den Hund streicheln], A L T +a -a Grethe Hess den Hund [laufen]. +a -a -a Grethe liess [ den Hund streicheln].

Of course the c-type sentence is potentially ambiguous, which is also expressed through the two possible ways of association. (89)

a.

A

L T F Grethe liess den Hund [fressen]. 'Grethe let the dog eat.' b. +a -a F Grethe liess [ den Hund fressen]. 'Grethe let the dog be eaten.'

8.2. Null subjects In many languages the first argument may also be left unassociated with finite verbs. Spanish is such a language: (90)

a.

Compre una casa. bought-lS a house Ί bought a house.' b. Vendre a las siete. come-FUT-lS at seven Ί will come at seven o'clock.' c. Traigo una carta para bring-lS a letter for Ί am bringing you a letter.'

ti. you

74

Jan Ter je Faarlund

All these sentences have verbs which combine with an Agent role. Hence the Agent associates with the first argument on the argument tier. If, however, we associate the first argument with any of the NPs present in the syntactic structures in (90), we get unacceptable interpretations. In each of the sentences, the first argument has to be left unassociated. 17 Una casa in (90a) and una carta in (90c) associate with the second arguments. A general rule that we have to adopt for this type of language, then, is that the first role may remain unassociated. This may of course also lead to ambiguities. In (90a) we do not strictly know formally whether the Agent is left unassociated or whether it associates with una casa. But since the verb comprar usually requires an expressed Theme, and since the Agent of comprar generally has to be animate, the null-subject reading is strongly preferred. The above procedure also extends to imperative sentences in other languages. There the rule is that the first argument is left unassociated when the verb is in the imperative, thus securing the correct second argument association of the NP in Eat the fish!

9. NPs in non-argument positions There is an obvious and fundamental problem attached to a purely linear association model: even in non-case-marking languages there is a certain degree of freedom of word order, which does not imply the same degree of ambiguity. Consider the following examples; (91b) is from Spanish and (91c) from Norwegian. (91)

a. Who did Sue meet? b. Compro una casa bought-3S a house 'Jose bought a house.' c. Iskrem it Per ice-cream ate Per 'Per ate ice-cream.'

Jose Jose

In each of these sentences, a linear association of arguments to NPs would yield ungrammatical sentences or at best very unlikely interpretations. In order to deal with such cases without recourse to empty categories, we need to introduce some more apparatuses,

Autostructural

analysis of semantic roles

75

which presumably will be simpler and more "visible" than traces and the like.

9.1. Topicalized NPs in English In English, a sentence may under certain circumstances start with two argument NPs. (92)

That man Sue would never even talk to.

The second NP, Sue, associates with the first argument, and the first NP, that man, with the second argument. We can resolve this and similar cases by assuming that, in English, association starts with the last preverbal NP. "Preverbal" is here taken to mean "preceding the finite verb". (I will consider auxiliaries finite verbs.) In most cases that NP is also the first NP, since typically only one NP precedes the verb in English main sentences (on subordinate sentences, see below). Now there is no NP left for the second argument to associate with. It looks in vain further to the right in the sentence. In frustration it goes back to start from the far left end of the sentence, where it does find a free NP to associate with. (93)

-a That man Sue would never even talk to

9.2. wh-movement Consider first English main clauses starting with a vW>word. In sentences without an auxiliary, or in sentences with no NP occurring between the auxiliary and the main verb, the wh-word associates with the first argument, and no special measures need be taken. (94)

a. +a

-a

Who wrote the

b. +a

letter?

-a

Who will write the

c.

+1

letter?

-1

Who does not like Sue?

76

Jan Terje Faarlund

The problem arises when an NP intervenes between the auxiliary and the main verb. (95)

a. b. c.

What will Max write? Who did Sue not like? Who didn't Sue like?

In these sentences we need to associate the first argument with the second NP in order to get the correct interpretation and a grammatical sentence. In order to secure that kind of association, we will consider the auxiliary to be a mark of first argument, so that any NP immediately to the right of an auxiliary associates with the first argument. As in the c a s e of topicalized NPs, the second argument again goes to the beginning of the sentence, where it this time finds the wh-word to associate with. (96)

+a

-a

What will Max write

This procedure also yields the right results in sentences with more than two arguments. (97)

a. b.

A

L T

What did Sue send to Max

A

T

L

Who did Sue send the letter to

Sentence (97a) is quite straightforward; the second argument is left at the end with nothing to associate with, and goes to the front. Regarding (97b), we have to keep in mind that the preposition to demands a Locus role. Since it is not followed by anything, it goes to the front to find a NP to associate with. This leaves the letter for the Theme role. The question now remains how the second arguments of verbs with optional objects "know" that they are missing something. How does the wh-word in (98a) get associated when the one in (98b) does not, and cannot?

Autostructural

(98)

a. b.

analysis of semantic roles

77

What did Max eat? When did Max eat?

Since the semantic and the syntactic structures are independent, without the one being derived from the other, association is bidirectional or neutral as to direction. This means that not only do roles need to find a phrase to associate with; phrases have the same need to find a role. The what in (98a) belongs to the category NP, hence it has to be associated with some argument. Since the first argument is already taken up by the NP following the auxiliary, it is now the first NP which gets frustrated, and goes out to look for an argument to associate with. This it finds at the end of the argument tier, and it is therefore associated with the second argument. This principle also applies to extracted subjects of infinitives: (99)

Who do you want to come?

Here there is no dangling argument at the end of the sentence which can go and find the vW>word. On the contrary, the iWi-word is left unassociated, and must itself go out to look for an argument. It does not find one at the end of the sentence, and must continue towards the left. The first position on the argument tier counting from the right is the first argument of come, which is precisely what who associates with. Consider next indirect questions. The initial w/i-word may be followed by a finite verb or by a NP. (100) a. (/ know) who wrote the letter. b. who should write the letter. c. what Max said. d. what Max will say.

These are subordinate sentences starting with w/7-words, which are NPs associated with arguments. In (100a) and (100b), the association is linear, but not in (100c) and (lOOd). Those sentences have two preverbal NPs (what Max), so they are like the topicalization sentences discussed in 9.1. Association starts with the preverbal NP, and the dangling argument at the end of the sentence finds the initial w/i-word to associate with. The same principle applies in the case of extractions. If the dangling argument in an embedded sentence does not find anything

78

Jan Terje Faarlund

to associate with within the subordinate clause, it will go to the beginning of the sentence to look for a possible associate. (101)

+1

-1 +a

-a

What do you think [ Max will do]

9.3. Verb-second languages In Germanic languages other than Modern English, the finite verb is always in second position in main clauses. The element preceding the verb may or may not be a NP. This offers a certain freedom of word order, which seems to pose no problem in case-marking languages like Icelandic and German. However, not all Germanic verb-second languages use case-marking as a major strategy. The examples in this section are all from Norwegian, a non-case-marking verb-second language. (102) a. Kari sende Per eit hrev. Kari sent Per a letter, b. F0rre veka sende Kari Per eit brev. last week sent Kari Per a letter 'Last week Kari sent Per a letter.' In these sentences associations take place from left to right in linear order, regardless of whether the first NP precedes or follows the verb. The three NPs are in the same relative order in both sentences. However, the first NP does not always associate with the first argument. (103) Per skal Kari m0ta imorgon. Per will Kari meet tomorrow 'Kari will meet Per tomorrow.' As in the English ννΛ-questions (95), there is a NP between the auxiliary and the main verb. This position is a unique subject position, so that the NP occurring there always associates with the first argument. In (103) there are no more NPs to the right of the subject, but the argument tier also has a second argument. Again this argument has to look to the beginning of the sentence

Autostructural analysis of semantic roles

79

for a NP to associate with. The difference between English and Norwegian is, however, that in the latter, it being a true verb-second language, not only auxiliaries, but also finite main verbs may be followed by the subject. (104) Per m0tte Kari Per met Kari 'Kari never met Per.'

aldri. never

In this case the sentence adverbial aldri 'never', whose position is also between the finite and the non-finite verb, marks the subject position. A sentence with aldri following the main verb and thus allowing for a possible NP position between it and the main verb, is syntactically ill-formed. (105) a. Per har Kari aldri Per has Kari never 'Kari has never met Per.' b. * Per har Kari m0tt aldri.

mptt. met

Hence, any NP occurring between the auxiliary and an adverb like aldri necessarily associates with the first argument. In those cases where there is neither an auxiliary nor a sentence adverb, two different associations are in principle possible. (106) Per mptte Kari. Per met Kari 'Kari met Per.' or 'Per met Kari.' The second NP, Kari, follows immediately after the finite verb. As a consequence, it can associate with the first argument. On the other hand, it is not followed by a non-finite verb or a sentence adverbial, and may therefore associate with the second argument, in which case the first NP associates with the first argument. The actual interpretation of course depends on the context. In other cases it depends on the lexical content of the nouns. (107) Is krem liker Per Ice-cream likes Per 'Per likes ice-cream.'

80

Jan Terje Faarlund

Sentence (107) is in principle just as available to two different sets of associations as (106) is, but the unexpectedness of one of them naturally makes the other one strongly preferred. 9.4. Free word order in a null-subject language Another language without case-marking but with a considerable freedom of word order is Spanish. Consider for example (108), with a strong preference for associating the second NP with the Agent. (108) Compro

una casa

mi

padre

bought a house my father 'My father bought a house' In order to explain this association, we should keep in mind that Spanish is a null-subject language, that is, a language where the first argument may be left unassociated. Assume that this is what happens in (108). The second argument then associates with the second NP, and we get the same association as we would in the perfectly normal (109). (109)

-a

Compro una casa

In (108) a NP, mi padre, is left dangling at the end of the sentence, and it has to go to the beginning to find a role to associate with. There it finds the Agent role. This is then a clear parallel to what happens in the case of object-questions in English, (95), except that here the direction is reversed. 9.5. "Orphaned" arguments and phrases, a summary A phrase that is left at the end of a sentence without having an argument there to associate with I will call an "orphaned" NP. Similarly, an argument at the end of the sentence without a NP to associate with, is an "orphaned" argument. The proposals made above about associations in sentences with deviant word order are based on the general assumption that all arguments need a NP to associate with, and all NPs need an argument to associate with,

Autostructural

analysis of semantic

roles

81

except in specifiable and language-specific instances, such as with infinitives, in null-subject languages, etc. It is this principle that motivates and justifies crossing association lines in these sentences. The relevant instances can be summarized as follows: a) An orphaned argument goes to the beginning of the clause to look for a NP. b) An orphaned first NP goes to the end of the argument frame to look for an argument and a role. c) An orphaned last NP goes to the beginning of the argument frame to look for an argument and a role.

10. Conclusion By discussing a variety of construction types from different languages, I have shown that it is possible to account for the association of semantic role with syntactic position or morphological category without recourse to abstract elements or constructs such as underlying syntactic structures, movement rules, empty categories, or co-indexing. The emphasis has been on demonstrating that this can be done using an apparatus which includes only such elements that are accessible to the speaker, such as surface string of morphemes, meaning of sentences, and syntactic properties of words. To the extent that this endeavor can be said to have been successful, it is also an argument in support of the Autolexical and Autosegmental approaches to linguistic representation. One consequence of this is that a separate level of grammatical relations-as posited by, for example, Marantz (1984: 4)—is not needed, at least not for the purpose of the interpretation of semantic roles. The relative position of argument NPs, combined with certain general principles, suffices to determine their semantic roles. Likewise, since semantic roles can be associated directly with surface structure, a separate level of s-structure is also unnecessary. It goes without saying that several important questions remain to be discussed. Apart from data from various languages that I have not touched upon due to a number of limitations, there is of course the question of how the syntactic module should be viewed. I have been very non-committal in my references to syntax in this

82

Jan Terje Faarlund

paper. The reader will have noticed that I have not used the term "tier" when talking about syntax. I have argued that the syntax does not need a transformational component for semantic reasons. That does of course not preclude the need for transformations or some similar mechanism for purely syntactic reasons. A natural (and necessary) extension of the present approach will be to explore the possibilities of other modules and/or tiers. There might be a discourse-functional module with tiers of definiteness and givenness, etc., as discussed in Faarlund (1989). And there might be a tier of syntactic properties, such as the ability to govern verb agreement, to be the antecedent of reflexives, to undergo deletion in conjoined sentences, etc. On the basis of association with such tiers, it would be possible to characterize syntactic functions such as subject and object in terms of the diverse types of features usually involved in the determination of such categories. Ultimately we might be able to arrive at a fairly precise notion of prototypical subjects and objects. Notes *

1.

2.

3.

I wish to thank Alexander Caskey, John Goldsmith, Jerry Sadock, and Elisa Steinberg for valuable comments to earlier drafts of this paper. But, as always, the author assumes the full responsibility for whatever nobody else would want to be credited with. I am grateful to the following persons for their help with linguistic data: Elisa Steinberg (Spanish), Carolyn Killean (Arabic), and Richard Chambers (Turkish). This paper was written in the inspiring environment of the Linguistics Department at the University of Chicago, and members of the department, students as well as faculty, have contributed to the ideas here more than they probably like to admit. I also want to thank the Norwegian Research Council for Science and the Humanities for its generous support. Numerous references could be cited. Among the best known are Fillmore (1965) and Emonds (1972), in addition to practically every early introductory text to transformational grammar. For an early alternative analysis, see Goldsmith (1980). This is analogous to the c a s e of a language which has no vowel opposition in certain environments, for example in unstressed syllables. For any vowel segment in that position it would be unnecessary to state other feature values than, e.g., [+vocalic]. The fact that this leads to crossing association lines need not worry us. In other modular or tier-based frameworks, the possibility of a universal prohibition of crossing association lines has been discussed

Autostructural analysis of semantic roles

4.

83

(Sadock 1987; Lapointe 1987; Goldsmith 1989). It would be unreasonable and futile to assume a similar constraint in the semantico-syntactic interface, given the abstract nature of the modules involved and the purely graphic nature of linearity in the semantics. This may be less immediately obvious, but consider near synonymous pairs such as The man has a big hat. the man with a big hat

5.

where with and have express the same relation, namely one between a [+locus] Cthe man) and a [-locus] (a bighatX The question then arises of how to distinguish between the roles of the two NPs in sentences like Ci) Sue walked from the library to the pub. In sentences like this, only one of the PPs, usually the fo-phrase, is V'-internal, as shown by the tests used in this section (ii and iii). (Admittedly, iia is not a very good sentence either, but it must be infinitely better than iib.) (ii)

a. b.

(iii) a. b.

6. 7.

8. 9.

Sue walked from the library to the pub, and Max did so from his office. *Sue walked from the library to the pub, and Max did so to his office. From the library, Sue walked to the pub. ?To the pub, Sue walked from the library.

Walk-to is therefore a predicate which takes a Locus role, whereas there is no predicate walk-from or walk-to-from in (i). This sentence has a complex propositional structure, with from as a higher predicate, and the walking action as another proposition embedded under it. The semantic role of the object of from depends on the lexical content of that word, and not on the roles connected with walk. The Cebuano data are from Bunye - Yap (1971). Note, by the way, that accusatives in nominative-less sentences never express an Agent. Furthermore, sentences like the ones in (37) may be interpreted as having implied Agents. So in (37a), for example, there is an understood Agent which makes me sleepy; the verb sofa, where no Agent is possible, always combines with a nominative NP; in (37b) "somebody" reminds me of the woman, etc. The Swahili data are from Comrie (1976). The symbol directly over the " [ " indicates the semantic role

84

Jan Terje Faarlund

associated with the following clause. 10. Also in Japanese the syntactic structure of causative sentences can be shown to be mono-clausal (Mihara 1987). 11. The Hebrew data are from Cole (1976). 12. There are exceptions to this generalization, as in the case of idioms: Max was never paid much attention to. 13. Sentence (76a) is structurally ambiguous; another possible analysis is F +1

-1

It is easy [for Sue to understand the problem]

14. 15.

16. 17.

where the predicate easy has only one argument, the infinitival phrase, and where Sue is the first argument of the lower argument frame only. The role symbol placed over the complementizer indicates that the role associates with the clause. An "open proposition" corresponds to a generic statement in the sense that its first argument is not specified. "Pleasing John" may be said to be easy for whoever tries to do it. The statement that John is easy to please is therefore not a statement about any particular event. This explains why John is likely to please is not a "ToughConstruction"; likely cannot be predicated of a generic state or event, only about a specific state or event. Equivalent constructions exist in other Germanic languages, such as Dutch and Danish. It is sometimes suggested that the inflection on the verb is the subject of the sentence in cases like this. In this model, that would mean that the first argument associates with the inflectional ending. This cannot be so, however, since semantic roles have to associate with some lexical material. The inflectional ending represents or refers to an entity in the context or speech situation, but, if that entity is not represented lexically in the sentence, the argument is left unassociated.

References Aissen, Judith 1974 "Verb raising", Linguistic Inquiry 5: 325-366. Bell, Sarah J. 1979 Cebuano subjects in two frameworks. Bloomington, IN: Indiana University Linguistics Club.

Autostructural

analysis of semantic

roles

85

Bunye, Maria Victoria R. - Elsa Paula Yap 1971 Cebuano grammar notes. Honolulu: University of Hawaii Press. Cole, Peter 1976 "A causative construction in modern Hebrew: theoretical implications" in: Peter Cole (ed.), Studies in modern Hebrew syntax and semantics. Amsterdam: North-Holland, 99-128. Comrie, Bernard 1976 "The syntax of causative constructions: cross-language similarities and divergences", in: M. Shibatani (ed.), Syntax and Semantics, Vol. 6: The grammar of causative constructions. New York: Academic Press, 261-312. 1985 "Reflections on subject and object control", Journal of Semantics 4: 47-65. Culicover, Peter W. 1988 "Autonomy, predication, and thematic relations", in Wendy Wilkins (ed.), 37-60). Emonds, Joseph 1972 "Evidence that indirect object movement is a structure preserving rule", Foundations of Language^. 546-561. Faarlund, Jan Terje 1985 "The making of a subject: a modular account of subject formation", in: William H. Eilfort - Paul D. Kroeber — Karen L. Peterson (eds.), CLS 21.1: Papers from the General Session at the Twenty-First Regional Meeting. Chicago: Chicago Linguistic Society, 75-89. 1989 "Discourse functions and syntax", in: Bradley Music — Randolph Graczyk - Caroline Wiltshire (eds.), CLS 25.2: Papers from the Twenty-Fifth Annual Regional Meeting of the Chicago Linguistic Soceity: Parasession on Language in Context. Chicago: Chicago Linguistic Society, 30-40. Fillmore, Charles N. 1965 Indirect object constructions in English and the ordering of transformations. The Hague: Mouton. Goldsmith, John 1976 "An overview of Autosegmental Phonology", Linguistic Analysis 2: 23-68. 1980 "Meaning and mechanism in grammar", in: Susumu Kuno (ed.), Harvard Studies in Syntax and Semantics, Vol. 3. Cambridge, MA: Harvard University, Department of Linguistics, 423-449. 1989 Autosegmental and metrical phonology. Oxford: Blackwell. Huck, Geoffrey J. - Almerindo E. Ojeda (eds.) 1987 Syntax and semantics, Vol. 20: Discontinuous constituency. San Diego: Academic Press.

86

Jan Terje Faarlund

Jackendoff, Ray S. 1983 Semantics and cognition Cambridge, MA: MIT Press. 1987 "The status of thematic relations in linguistic theory", Linguistic Inquiry 18: 369-411. Keenan, Edward L. — Bernard Comrie 1977 "Noun Phrase accessability and universal grammar", Linguistic Inquiry 8: 63-99. Ladusaw, William A. - David R. Dowty 1988 "Toward a nongrammatical account of thematic roles", in: Wendy Wilkins (ed.), 61-73. Lapointe, Steven G. 1987 "Some extensions of the Autolexical approach to structural mismatches", in: Geoffrey J. Huck - Almerindo E. Ojeda (eds.), 152-183. Marantz, Alec 1984 On the nature of grammatical relations. Cambridge, MA: MIT Press. Mihara, Ken-ichi 1987 "Argument structure and complex predicate in Japanese", Gengo Kenkyu 92: 56-75. Rappaport, Malka — Beth Levin 1988 "What to do with θ-roles", in: Wendy Wilkins (ed.), 7-36. Rosenbaum, Peter S. 1967 The Grammar of English predicate complement constructions. Cambridge, MA: MIT Press. Sadock, Jerrold M. 1985 "Autolexical syntax: A proposal for the treatment of noun incorporation and similar phenomena", Natural Language and Linguistic Theory 3: 379-439. 1987 "Discontinuity in Autolexical and autosemantic syntax", in: Geoffrey J. Huck — Almerindo E. Ojeda (eds.), 283-301. 1991 Autolexical Syntax: A theory of parallel grammatical representations. Chicago: University of Chicago Press. Santorini, Beatrice 1985 "Null subject complements of the German verb lassenin: J. T. Faarlund (ed.), Germanic Linguistics. Bloomington, IN: Indiana University Linguistics Club, 135-156. Searle, John R. 1969 Speech acts: An essay in the philosophy of language. Cambridge: Cambridge University Press. Wilkins, Wendy (ed.) 1988 Syntax and semantics, Vol. 21: Thematic relations. San Diego: Academic Press.

Word order in Romanian subjunctive clauses*

Donka F. Farkas

The problem I will address here concerns the interaction of complementizers and word order in Romanian subjunctive clauses. The account to be proposed relies on the Autolexical framework proposed in Sadock (1985), which allows distinct components of the grammar to assign partially discrepant structures to the same string.

1. T h e data Romanian has two main complementizers, cä, which introduces indicative clauses, and ca, which introduces subjunctive clauses. Adverbial clauses are introduced by one of these complementizers preceded by various prepositions which determine the interpretation of the clause. The subjunctive mood in Romanian is formed with the verbal particle sä followed by the verb, which is inflected for the subjunctive. The subjunctive inflection is identical to the indicative one except in the third person. The present subjunctive paradigm is exemplified in (1). (1)

Present Singular lp: sämerg 2p: sä merg-i 3p: sä mearg-ä

Plural sämerg-em sä merg-efi sä mearg-ä

The use of a particle in the subjunctive is a Balkan trait that Romanian shares with Greek and Bulgarian. In Romanian, the particle sä may not be separated from its V except by a limited number of other particles and clitics occurring in a fixed order. They are exemplified in (2):

88

Donka F. Farkas

(2) Sä nici nu-mi mai vorbe§ti. SUBJ even not-me again speak.2p.sg 'Don't even talk to me anymore.' Note that the particle sä is the leftmost element in this sequence of verbal bound morphemes. The word order of major syntactic constituents is relatively free in Romanian. In particular, Romanian, like its Romance sisters and its Balkan neighbors, has "free" Su(bject)-V inversion and is a Subject-pro drop language. I turn now to data concerning the interaction of complementizers and word order.

1.1. Complements In indicative complements cä is obligatory and the clause may be V-initial or not: 3)

i*cä) Petru Peter 'Peter

(Χ) V (Y) §tie (*cä) Ana e aid/e aid Ana. knows that Anna is here/is here Anna knows that Anna is here.'

Matters are more complex in subjunctive clauses. The first generalization is stated in (4). The evidence for (4) is given in (5). The second generalization is stated in (6), and the relevant examples are given in (7). In non-standard Romanian (6) does not hold, i.e., both variants in (7) are acceptable in this dialect. (4) Ca is optional; if not present, the clause must be sä initial. (5) a. Petru vrea ca Ana sä piece. Peter wants that Anna SUBJ leave. b. Petru vrea sä piece Ana. Peter wants SUBJ. leave Anna 'Peter wants Anna to leave.' c. *Petru vrea Ana sä piece. Peter wants Anna SUBJ leave (6) If ca is used the complement may not be sä initial in standard Romanian. (7) Petru vrea (%ca) sä piece Ana. Peter wants that SUBJ leave Anna 'Peter wants Anna to leave.'

Word order in Romanian subjunctive clauses

89

I will now turn to two types of adverbial clauses, purpose clauses and färä 'without' clauses, both requiring subjunctive. Adverbial clauses in the indicative behave like complement clauses with respect to word order possibilities. 1.2. Purpose clauses Purpose clauses are introduced by the complex complementizer pentru ca, whose first element, the preposition pentru 'for', is optional: (8) a.

Petru a ad us flori (pentru) ca mama sä Peter has brought flowers (for) that mother SUBJ fie multumitä. be pleased b. Petru a ad us flori (pentru) ca sä Peter has brought flowers (for) that SUBJ fie multumitä mama. be pleased mother 'Peter brought flowers in order for mother to be pleased.'

Note that the ca sä sequence is acceptable in purpose clauses in the standard, as well as the non-standard dialects. T h e complementizer ca may be absent in purpose clauses as well. The generalization summarizing the word order constraints in purpose clauses is given in (9), and the relevant data are given in (10). (9)

In purpose clauses, if ca is absent, pentru may not be used, and the clause must be sä initial.

(10) a.

Petru a adus flori (*pentru) sä Peter has brought flowers for SUBJ fie mulfumitä mama. be pleased mother b. Petru a adus flori (*pentru) *mama Peter has brought flowers for mother SUBJ be pleased sä fie multumitä. 'Peter brought flowers in order for mother to be pleased.'

90

Donka F. Farkas

1.3. Färä ('without') clauses These clauses are introduced by the preposition färä 'without'. The relevant generalization with respect to the interaction of the preposition, the complementizer ca and the word order of the clause is given in (11) and exemplified in (12). (11) In färä Without' clauses, ca is optional. If ca is absent, färä must be used and the clause must be sä initial. (12) Petru a plecat färä (ca) sä-1 vadä Peter has left without that SUBJ-him see * Petru a plecat färä Ana sä-1 Peter has left without Anna SUBJ-him 'Peter left without Anna seeing him.'

Ana. Anna vadä. sees

Common to all clauses we have looked at so far is that they conform to the generalization in (4), independently of whether they are complements or adverbials. Let us note here that subjunctive relatives, unlike subjunctive complements or adverbial clauses, are not subject to any special word order restrictions. Relative clauses in Romanian are introduced by an obligatory wh- word, which must be the leftmost constituent of the clause. There is no constraint requiring the relative clause to be sä initial, nor is there any prohibition against sä immediately following the relative pronoun:1 (13) a.

Caut ο carte I-look for a book Mariei. Maria.DAT b. Caut ο carte I-look for a book Mariei. Maria.DAT Ί am looking for a

pe care Ana sä ο ACC which Anna SUBJ it

dea give

pe care sä ο dea Ana ACC which SUBJ it give Anna book for Anna to give Maria.'

Word order in Romanian subjunctive

clauses

91

2. Analysis I will now propose an account of these facts in the Autolexical framework. Crucial to this framework, and to my proposal, is that it allows for the possibility of a particular linguistic unit to function in the syntax as a constituent of type P, and, as such, obey whatever syntactic restrictions such constituents obey, and in t h e morphology as an affix of type A, and, as such, obey whatever restrictions affixes of this type obey. Thus, recall the analysis of's in (14) given in Sadock (1985). John's

(14)

(15) a.

here.

S NP I Ν

John

W VP V

Ν I ADV

CL

John

here

In Sadock's analysis, '5 is a V in the syntax, while in the morphology, it is a clitic that forms a unit with John. The syntactic structure of (14) is (15a); the relevant bit of morphological structure is given in (15b). Now, turning back to Romanian, the essence of my proposal is that we should view the particle sä as leading such a double life as well, i.e., I propose that sä is a verbal particle forming a unit with the verb in the morphology, while in the syntax it may optionally function as a complementizer. The proposal is summarized in A. A.

The morphological category of sä is [PART]; in the syntax sä may optionally function as a [COMP] element.

I assume that [PART] is a morphological feature borne by subsyntactic elements, such as clitics and adverbial particles, and that such elements are invisible to the syntax, i.e., they have no role in the syntactic structure of the expressions in which they occur. The feature [COMP], on the other hand, is a syntactic feature characterizing complementizers. Any element bearing it must occur in the syntactic position called COMP. The morphological categorization of sä, together with the immediate dominance (ID) rule that introduces it and the relevant

92

Donka F. Farkas

linear precedence (LP) statement are given in (16). The syntactic categorization of sä, together with the relevant immediate dominance rule and linear precedence statement, are given in (17). (16) morphology: sä [PART ] ID rule: V' - > PART, negj, neg2, (cl), adv, V LP statement: sä pentru, Η where Η is the head. The fact that Η may not be sä follows from the fact that sä is morphologically a particle, and the assumption that heads and mothers must share major category features. What has been said so far allows both cä and ca in (22) and (24). In the case of färä clauses, cä must be excluded; this can be done by requiring the clause to be in the subjunctive. In the case of pentru clauses, both complementizers are possible: if the complementizer is ca (followed by the subjunctive) the clause receives a purpose interpretation; if the complementizer is cä (followed by the indicative) the clause receives a "reason" interpretation, as exemplified in (25). (25) Ion a plecat pentru cä Maria a John has left for that Maria has 'John left because Maria came.'

venit. come

Now the fact that in purpose clauses both pentru and ca may be absent can be accounted for by allowing a subjunctive clause introduced by a bare ca or sä to be optionally interpreted as a purpose clause. Assuming that both ca and sä, when a complementizer, bear the feature [SUBJ(unctive)], which is

96

Donka F. Farkas

responsible for the subjunctive mood of the clause introduced by them, the facts pertaining to the interpretation of adverbial clauses can be accounted for by the rules in (26), which interpret clauses according to the nature of their complementizers. (26) Interpretation of complementizers: a. c[färä X[SUBJ]] : "without" clause b. c[pentruX/X[SUBJ]] : reason/purpose clause c. c [X[SUBJ]]: (purpose) So far, then, I have accounted for the possible absence of ca in all subjunctive clauses, for the fact that if ca is absent, the clause must be sä initial, as well as for the peculiarities of purpose and färä clauses. I will now turn to the prohibition against the ca sä sequence in complements in standard Romanian. First, I propose that in standard Romanian the ca sä sequence in COMP is necessarily interpreted as a purpose clause, by (27). (27) Standard Romanian c[ca s5]: purpose The non-standard dialect lacks this rule. The existence of (27) amounts to saying that in standard Romanian, complement clauses may not be introduced by complex complementizers, a constraint consistent with the cross-linguistic tendency of having complex complementizers in adverbial clauses and simple ones in complements. This tendency is functionally motivated, since the thematic role of complement clauses is determined by the predicate of which they are an argument, whereas in the case of adverbial clauses the complementizer has the semantic function of specifying the thematic role of the clause. The second part of the account of the ungrammaticality of the ca sä sequence in complements has to ensure that if the position of sä is compatible with it being in COMP position, it will be taken as occurring there, i.e., that if a clause is sä initial, sä will be analyzed as occurring in COMP. I suggest that this follows from a universal tendency, which I will call the Maximum Visibility Principle (28). According to this principle, linguistic expressions are assigned analyses in which their constituents are maximally visible to the various components of the grammar. Note that the existence of such a principle is desirable if optional categorization such as that

Word order in Romanian subjunctive

clauses

97

in (17) is allowed, since by it we reduce the potential for ambiguity created by such optional categorization. (28) Maximum Visibility Principle In case of potential structural ambiguity due to optional categorization in a component the ambiguity is resolved in favor of the structure in which the optional categorization is met. Note now that in our case the Maximum Visibility Principle predicts that if sä is the leftmost element of its clause, i.e., if one can analyze it as occurring in COMP, it will be so analyzed since then sä will be visible to both the syntax and the morphology. Therefore, if ca is immediately followed by sä, sä will be taken as occurring in COMP, but because of (27) such clauses may only be purpose clauses in standard Romanian, which explains the unacceptability of the ca sä sequence in complement clauses in this dialect. In the non-standard dialect ca sä will be acceptable in complements because of the absence of (27); w e also predict that the ca sä sequence will be acceptable in adverbial clauses in both dialects. Under the present account, then, the prohibition against the ca sä sequence in complements in standard Romanian follows from the interaction of the Maximum Visibility Principle and the dialect-specific rule in (27). Crucial for the present account is to find independent motivation for the Maximum Visibility Principle, a problem that I leave for further research. It is interesting to note here that the subjunctive particle da in Bulgarian has been analyzed by Rudin (1986) as an ordinary particle rather than as a complementizer (although proposals treating it as a complementizer have also been made), while the Serbo-Croatian subjunctive da appears to be best analyzed as a complementizer and not as a particle. Under the proposal made here, Romanian is a mixture of these two cases.

98

Donka F. Farkas

3. Further problems 3.1. Focus We will now turn to the interaction of the present analysis with certain facts concerning the syntax of focused constituents. Contrasted constituents in Romanian receive special stress (indicated by capitals) and they may occur in a position immediately to the left of the V and to the right of COMP, which I will call FOCUS. A constituent in FOCUS position may be separated from the V only by sä and the clitics and particles listed in (2): (29) a.

Petru vrea ca pe ANA sä n-o vadä. Peter wants that ACCAnnaSUBJ not-her see Ion (§i nu pe Maria). John (and not ACC Maria) 'It is Anna that Peter wants John not to see, (and not Mary).' b. *Petru vrea ca pe ANA Ion sä n-o Peter wants that ACC Anna John SUBJ not-her vadä (§i nu pe Maria). see (and not ACC Maria) 'It is Anna that Peter wants John not to see, (and not Mary).' c. *Petru vrea pe ANA ca Gheorghe sä ο vadä. Peter wants ACC Anna that George SUBJ her see "It is Anna that Peter wants George to see.'

I assume in what follows that FOCUS is adjoined to either VP or to S in the syntax, but that a constituent in this position is subject to a prosodic requirement forcing it to be immediately preverbal, and therefore the VPs and Ss that have a FOCUS adjoined to them must be V-initial. The problem to be addressed now is the interaction of FOCUS with the non-null complementizer constraint and the optionality of ca. The analysis presented above predicts that if there is an element in FOCUS, ca will be obligatory since the focussed element will be to the left of sä, preventing it from being analyzed as occurring in the COMP position. This prediction is borne out in adverbial clauses, as illustrated in (30), and in most complement clauses, as illustrated in (31).

Word order in Romanian subjunctive clauses

99

(30) a. Petru a adus flori (pentru) ca MAMA Peter has brought flowers for that mother sä fie mulpumitä. SUBJ be pleased 'Peter has brought flowers so that Mama will be pleased.' b. *Petru a adus flori MAMA sä fie Peter has brought flowers mother SUBJ be mulpumitä §i nu Maria. pleased and not Maria 'Peter has brought flowers so that it is Mama who is pleased and not Maria.' c. Petru a plecat färä *(ca) ANA sä-1 vadä. Peter has left without that Anna SUBJ-him sees 'Peter has left without Anna seeing him.' (31) Petru a hotarlt/spus *(ca) ANA sä vinä cu Peter has decided/said that AnnaSUBJ comes with noi nu Maria. us and not Maria 'Peter has decided/said that it is Anna who is to come with us, not Maria.' The problem is posed by the verb a vrea 'to want', which allows the omission of ca in the presence of a preverbal constituent in FOCUS: (32)

Petru vrea (ca) ION sä piece (§i nu Maria). Peter wants that John SUBJ leave (and not Maria) 'It is John that Peter wants to leave, and not Mary.'

In order to account for the version of (32) without ca we have to suspend the non-null complementizer constraint in the case of complements of a vrea 'to want', but only in the presence of a preverbal focused constituent. I suggest that what is going on here has to do with the deeper motivation of the non-null complementizer constraint, which is that Romanian requires clauseboundaries to be marked. The strongest clause boundary mark is a complementizer since complementizers can only adjoin to clauses. I suggest now that a FOCUS position adjoined to S is a clause boundary mark as well, albeit a weak one. It is a clause-boundary because it provides the clause with superstructure; it is a weak boundary since it may adjoin to VPs as well as to clauses. The special property of the verb a vrea 'to want' is that it allows its

100

Donka F. Farkas

complement to be separated from the main clause by a weak clause boundary and not only by a strong one. The fact that it is this particular verb that exhibits this particular property is not surprising in view of the exceptional behavior of volitional verbs cross-linguistically, behavior that suggests a weaker clause boundary here than elswhere. In English, for instance, volitional verbs are Exceptional Case Marking verbs in Government and Binding terminology, i.e., they may Case mark the subject of their complement. Rivero (1987) reports that a vrea 'to want' is such an Exceptional Case Marking verb in certain dialects of Romanian as well. Note also that for French, Pica (1985) observes that a matrix negative licenses negative polarity items such as rien 'nothing' and aucun 'none' in complements of volitional verbs but not elsewhere. In order to make these suggestions more precise one would have to provide a more refined theory of clause boundaries than I am able to at present.

3.2. Relative clauses Recall that in relative clauses the relative pronoun must be the leftmost constituent of the clause and that there are no particular restrictions on the word order within a subjunctive clause. These data are accounted for by the above analysis under the assumption that the relative pronoun occurs in COMP position. If, however, the position in which the relative pronoun is assumed to occur is other than COMP, namely Specifier of CP or some other position adjoined to S, we must modify the non-null complementizer constraint so as to take this position, when filled, as marking a clause boundary as well. Given the observations made above, such a modification would not be counterintuitive.

Notes * 1. 2.

I am indebted to Jerry Sadock for useful comments on an earlier draft of this paper. In relative clauses there is a preference for V-initial order independently of the mood of the clause. See Farkas (1988) for a detailed comparison between an analysis along the lines to be presented here and previous analyses of these facts, proposed in Farkas (1984) and Grosu - Horvath (1984).

Word order in Romanian subjunctive clauses

101

References Farkas, Donka F. 1984 "Subjunctive complements in Romanian", in: P. Baldi (ed.), Papers from the Xllth Linguistic Symposium on Romance Languages. Philadelphia, PA: John Benjamins, 355-372. 1988 "On the morpho-syntax of subjunctive clauses in Rumanian", International Journal of Rumanian Studies 6: 7-18. Farkas, Donka F. - Jerrold M. Sadock 1989 "Preverb climbing in Hungarian", Language 65: 318-338. Grosu, Alexander - Julia Horvath 1984 "The GB theory and raising in Rumanian", Linguistic Inquiry 15: 348-353. Pica, Pierre 1985 "Subject, tense and truth", in: Jacqueline Gueron-HansGeorg Obenauer-Jean-Yves Pollock (eds.), Grammatical representation. Dordrecht: Foris, 260-2 91. Rivero, Maria-Luisa 1987 "Barriers and Romanian", in: Janet DeCesaris-Carl Krischner (eds.), Proceedings from the Linguistic Symposium on Romance Languages XVII. Amsterdam: John Benjamins. Rudin, Catherine 1986 Aspects of Bulgarian syntax: Complementizers and WH constructions. Columbus , OH : Slavica. Sadock, Jerrold M. 1985 "Autolexical syntax: A theory of noun incorporation and similar phenomena", Natural Language and Lingusitic Theory 3: 379-440.

Control in Autolexical Syntax David Kathman

1. Introduction In this paper I will examine, using the framework of Autolexical Syntax (Sadock 1991a), what are commonly known as "control" structures-i.e., sentences which involved Equi-NP Deletion in classical transformational grammar, and which contain PRO in current Government and Binding work. Sentences (1) and (2) are examples: (1) (2)

John tried to leave. John persuaded Mary to leave.

In sentence (1), John is both the overt subject of try and the understood subject of leave, and in (2), Mary is both the object of persuade and the understood subject of leave. Such sentences are superficially similar to "raising" sentences such as (3) and (4), but they differ in a number of ways. (3) (4)

John seemed to leave. John believes Mary to smoke.

For example, meaningless dummy NPs can occur as the subject of (3) and the object of (4), but not in the corresponding positions in (1) and (2): (5) (6) (7) (8)

*It tried to rain. It seemed to rain. *John persuaded it to be the case that Mary smokes. John believed it to be the case that Mary smokes.

Such observations led (in 1960's-style transformational grammar) to an analysis where John tried to leave had a deep structure corresponding to John tried John to leave, with the second instance of John being deleted by Equi-NP Deletion, while John seemed to

104

David Kathman

leave came from a deep structure corresponding to Seemed John to leave, with John being raised to subject position. Eventually Equi-NP deletion fell out of favor with most syntacticians, and a variety of devices have taken its place-most notably the empty category PRO in Government and Binding work, though a number of linguists have concluded that the phenomena in question involve not syntax, but semantics (Bach 1979, Brame 1976) or pragmatics (Dowty 1985). These disagreements over the analysis of control structures are symptomatic of broader disagreements over the boundaries of both syntax and semantics and the relation between them; how a given theory handles control structures can tell us a lot about how it views syntax and semantics in general. Autolexical Syntax radically separates syntax and semantics (as well as morphology and other subsystems) into maximally simple modules interacting in potentially complex ways; thus, it should have something to say about how to treat control phenomena. So far, though, Autolexical work on the syntaxsemantics interface has tended to be sketchy and programmaticunderstandably so, given the newness of the theory and the fact that its focus was originally the morphology-syntax interface. In the following pages I hope to give a coherent account of control structures in Autolexical Syntax. In order to do so, I will have to venture into other areas of syntax and semantics, sometimes suggesting additions to, and modifications of, the theory. The conclusions will be tentative, but I hope I can contribute something both to the discussion of control structures and to the development of the Autolexical model.

2. Basics 2.1 The Autolexical model Before we start looking at control in particular, it will be useful to summarize the relevant features of Autolexical Syntax.1 The most obvious such feature is that the various subsystems of language, such as syntax, semantics, and morphology, consist of independent modules, each with its own primitives and principles of combination. Of course, these subsystems are separated to some extent in

Control in Autolexical Syntax

105

virtually every model of language, but usually they are tightly bound together, with one system seen as more primary than the others. For example, for most of the history of generative grammar, syntax has been seen as basic, the source of the infiniteness of language, with semantics and phonology being derived from, or somehow dependent on, syntax. 2 Autolexical Syntax is generative in the broad sense (it explicitly specifies the structure of a language), but the relationship between the modules is looser than in most other models, with mismatches of various kinds being possible. The earliest Autolexical literature (e.g., Sadock 1985) dealt mostly with morphology-syntax mismatches, but I will suggest below that control structures involve a type of syntax-semantics mismatch. The syntax in an Autolexical model is a context-free phrase structure grammar, of the general type described in Gazdar et al. (1985); however, all the descriptive apparatus of that theory may not be necessary when the contributions of other modules are taken into account. This syntax is surface-oriented; there are no empty categories. 3 Syntactic categories are bundles of features, though in this paper I will generally only show features which are relevant to the discussion. Similarly, in principle, syntactic rules will be decomposed into immediate dominance and linear precedence rules, but for ease of exposition I will generally collapse these. The semantics in an Autolexical model shows the logical relations among the elements of a sentence, including such things as argument structure and scope relations. It does not relate sentences to states of affairs in the world; if such a mapping is to be done, it will have to be done elsewhere. Thus Autolexical semantics is similar in broad outline to LF (i.e., Logical Form) in Government and Binding theory, in that it handles scope and variable binding; however, it differs from LF in that it is autonomous and its primitives are different from those of syntax. Like the syntax, Autolexical semantics is a context-free phrase structure grammar, which by stipulation is limited to binary branching. A predicate taking χ arguments is symbolized F"x, and a proposition (i.e., a predicate which has been "saturated") is symbolized F. A given predicate will combine with a variable to give a predicate of the next lowest valence, as in the following rules of the semantic "grammar": (9) a. F = F"1 (x) b. r ' = F"2(x)

106

David

Kathman

A semantic operator (Ο) combines not with a variable but with a proposition to give a predicate of the next lowest valence: (10) a. F = 0~'(F) b. F"1 = O"2 (F) Such operators will be important in the discussion of control structures below. The various arguments in a semantic tree are represented by variables, but these variables must be bound by quantifier expressions CQ), which represent the meaning of noun phrases; one quantifier expression is needed for each distinct variable which occurs in the main semantic tree, though we will see shortly that a single Q-expression can bind more than one occurrence of the same variable. The internal structure of Q-expressions will not be particularly relevant for us, so I will mostly ignore it, using proper names for arguments and abbreviating the content of Q-expressions as John(x) or Mary(y) or whatever is appropriate. A Q-expression must occur higher in the semantic tree than the variable(s) it binds, and relative hierarchy of Q-expressions is relevant for scope phenomena, but we will not be very concerned with that here. There are other modules for such things as morphology and discourse, but we need not worry about them for our purposes. What ties all these modules together is the lexicon, which contains information about a given element's behavior in all the modules where it has some function. An element need not be present in all modules; for example, the English past tense morpheme plays roles in the morphology and the semantics, but not in the syntax. 4

2.2 Autolexical control structures in a nutshell How can the syntactic and semantic systems described above interact to describe control structures? As an example, take the sentence John tried to leave. The syntactic structure of this sentence is straightforward-try is a verb which takes an infinitival VP complement, and the rules in (11) give the structure in (12) (to simplify the tree I have omitted the V node under VP[BSE]):5 (11) a. S -> NP VP b. VP -> V VP [-FIN] c. VP[-FIN] -> V[to] VP

Control in Autolexical Syntax

107

(12) NP John

VP V

VP [-FIN]

tried V

to

^ VP[BSE]

leave

Semantically, though, try takes a proposition as a complement. The above sentence might be paraphrased (awkwardly) as John

tried to make it the case that he (John) left, similar to the

(semantically-based) deep structure of classical transformational grammar. Thus, in our terms, try is a two-place semantic operator, which combines with a proposition to yield a one-place predicate. The semantic tree for the above sentence would be roughly as in (13).6 (13)

Q

John(x)

Here the two-place operator try combines with the proposition χ leaves to form a one-place predicate; this in turn combines with the variable χ to form a proposition. Both instances of χ are bound by the Q-expression John(x), meaning that the "subject" of the matrix proposition and the "subject" of the embedded proposition are the same individual, who has the property of being John.7 An "object control" verb such as persuade can be handled in a similar manner. Syntactically it will subcategorize for an NP object and an infinitival VP, giving us the syntactic tree in (14). Semantically, persuade will be a three-place operator (O 3 ), which combines with a proposition to give the semantic equivalent of a transitive verb as in (15). Here it is y, the variable bound by Mary, which occurs in the embedded proposition, since Mary is the understood subject of leave.

108

David Kathman

(14) VP

NP

John

V

persuaded

NP

VP [-FIN]

Mary V

VP

to

leave

(15)

Q

John(x)

Q

Mary(y)

Ο' persuade

Both of the verbs we have seen-f/yand persuade-will have lexical entries which indicate the relevant syntactic and semantic subcategorization facts (16) and (17).

(16) try syntax: [ VP VP [to]] semantics: O"2

(17)

persuade syntax: [ VP NP VP [ ί ο ] ] 3 semantics: O"

(These entries omit information about any modules other than syntax and semantics, which are the only ones relevant for our purposes.) What makes these verbs control verbs is the fact that they have a proposition in the semantics where there is only a VP in the syntax; the syntax contains one too few NPs, so the extra semantic variable in the embedded proposition ends up being bound by (the meaning o f ) one of the NPs which occurs overtly in the sentence. The specification of the understood subject of the

Con trol in A utolexical

Syn tax

109

embedded verb is thus handled not in the syntax (as in Government and Binding), but in the semantics. Another way of looking at the situation is that the Autolexical semantics corresponds to the deep structure of classical transformational grammar and the syntax corresponds to the surface structure, but instead of one being derived from the other, they each exist independently; instead of being linked by transformations, they are linked by the lexicon.

3. Complications 3.1 The syntax-semantics connection The above sketch leaves a number of questions unanswered, however, and we will have to address these if we want our account to be taken seriously. One issue which will have to be addressed is the linking between the syntax and the semantics. Although the modules of the theory are autonomous in that none is derived from any of the others, there are limits as to how different from each other the various representations of a given sentence can be; a significant part of Autolexical research has been defining these limits. As a start, Sadock (1991a) proposes the Linearity Constraint, which states that lexemes projected in two different modules occur in the same linear order in both, and the Constructional Integrity Constraint, which says (restating somewhat) that lexemes projected in two different modules have the same c-command relations in both.8 These constraints can be violated (there would be no point in having separate modules if they couldn't), but there are limits as to how severe the violations can be, the details of which we need not go into here. The important idea is: the fewer violations, the better. However, the discussion of these constraints in Sadock (1991a) and elsewhere is almost exclusively in terms of the morphologysyntax interface, where it is fairly clear what corresponds to what. If we try to compare semantic and syntactic representations, the correspondences are less clear. For example, consider the syntax and semantics of a simple sentence like John kissed Mary (18). Since the semantics is unspecified for linear order, the Linearity Constraint is not relevant here, but the Constructional Integrity

110

David Kathman

Constraint is. The versions of the Constructional Integrity Constraint given in Sadock (1991a) are oriented toward the morphologysyntax interface, so in order to make this constraint more general I will restate it as in (19) (based on Sadock's [1991b] Association Condition). b.

(18) a. VP

NP

John

V

kissed

NP

Mary

Q

John(x)

Q

Mary(y)

(19) Generalized Constructional Integrity Constraint If a c-commands b in module M, then a' c-commands b ' in module M' (where a and b are linked to a'and b' respectively). I assume here the standard definition of c-command, where a ccommands b iff the first branching node dominating a also dominates b. With this definition in mind, let us look at figures (18a) and (18b). T o start with, it is uncontroversial that the syntactic V kissed will be linked to the semantic F 2 kiss, since they are parts of the same lexical entry; what we are concerned with is the linking of the NPs. Suppose we link the NP John to the Q expression John(x) and the NP Mary to the Q expression Mary(y). John c-commands everything else in the syntax, and John(x) does likewise in the semantics. But Mary and kissed c-command each other in the syntax, while in the semantics, kiss does not c-command Mary(y), resulting in a Constructional Integrity Constraint violation. If, however, we link John and Mary to the variables χ and y respectively, then everything works out fine; the trees in question are structurally identical. It will be true in general that a much better syntaxsemantics matching is possible if we link NPs to variables rather than Q expressions, at least for the purposes of the Constructional Integrity Constraint, so I will assume such a linking from now on.9 In order to show such linking explicitly, we can use diagrams such

Con trol in Autolexical Syn tax

111

as (20), with syntactic structure on the top and semantic structure (upside down) on the bottom: (20) VP

NP John

NP Mary

V kissed ι F"2

y F-1

If syntactic NPs correspond to semantic variables, what do we do with control structures, where there are two semantic variables corresponding to a single syntactic NP? The most obvious answer is that the NP in question is linked to the variable which will cause the fewest Constructional Integrity Constraint violations, with the other variable not being linked to the syntax at all. For a sentence like John tried to leave, the decision is straightforward: John c-commands everything else in the syntax, so it should be linked to the higher10 of the two AS, which c-commands everything else in the semantics (ignoring here the Q expressions). The other χ is "free" (i.e., not linked to the syntax), a fact I will indicate by circling it:11 (21) VP

NP V John

tried

VP [-FIN] V to

VP leave

112

David Kathman

For an object-control verb like persuade, things get a little more complicated. In (22), it is clear that John corresponds to x, but it is not immediately obvious which of the ys Mary should be linked to. A table of c-command relations may make the decision easier. Table (23) shows the c-command relations holding in the syntax, assuming the structure shown in (22). Since John and χ c-command everything else in the syntax and semantics respectively, they are irrelevant here and are omitted from the folowing tables; the infinitival particle to is also irrelevant for the Constructional Integrity Constraint since it is not connected to the semantics. (22)

John

persuaded

Mary

to

?y 0 -3 7-2

7-1

(23) a. b. c. d.

persuaded persuaded Mary Mary

c-commands c-commands c-commands c-commands

Mary leave leave persuaded

Suppose we link Mary with the lower y in the semantic tree. Then the semantics will have the following c-command relations, where persuaded' and leave' stand for the meanings of those words, and Mary' stands for the linked y. (24) a. b. c. d.

persuaded' persuaded' Mary' leave'

c-commands c-commands c-commands c-commands

Mary' leave' leave' Mary'

Con trol in A utolexical Syn tax

113

The first three of these correspond to the first three relations shown in (23) for the syntax, but the fourth has no syntactic correlate; leave does not c-command Mary in the syntax. On the other side of the coin, the syntactic relation shown in (23d) has no semantic correlate; the meaning of Mary (the linked y) does not c-command the meaning of persuaded. Thus we have two Constructional Integrity Constraint violations. Suppose, on the other hand, that we link Mary with the higher of the two ys. In that case, the following c-command relations will hold in the semantics, where Mary' now stands for the higher y. (25) a. persuaded' b. Mary' c. Mary'

c-commands leave' c-commands leave' c-commands persuaded'

All three of these correspond to syntactic c-command relations, as a comparison of (25) with (23b) through (23d) reveals. The only discrepancy is that (23a) has no semantic correlate; the meaning of persuaded does not c-command the meaning of Mary. So here we have only one Constructional Integrity Constraint violation, and we should thus prefer the linking of Mary with the higher y. This seems intuitively correct; the higher y is the semantic direct object of the main proposition, just as Mary is the direct object of the matrix verb in the syntax, and having the unlinked variable be the semantic subject of the embedded proposition parallels what we saw with try. Although the Constructional Integrity Constraint works well in the cases we have seen so far, with more complicated sentences it does not always give us an unambiguous linking. For example, in the sentence John persuaded Mary to kiss Bill, where the embedded clause has a direct object, the methods we have been using so far predict that Mary can be linked to either of the ys in (26), since both of these linkings involve only one Constructional Integrity Constraint violation. This is due to the fact that the lower y and the predicate of the embedded proposition no longer mutually c-command each other, as they do in (22), so there is no analogue here to (24d). Presumably we would like Mary to be linked to the higher y, by analogy with (22), but the Constructional Integrity Constraint by itself will not do it for us; we need something else. One possibility is to simply specify in the lexical entry for persuade that the semantic subject of its complement F must be unlinked to the syntax; this could be done by augmenting the

114

Da vid Kathman

semantic portion of the lexical entry as in (27), where (x) stands for any syntactically free variable:12 (26)

s

NP

VP

F"1

χ F (27) persuade syntax: [VP semantics:

NPVP[to]] [F (x) F"1]]

This way, the lower y will automatically be free, leaving the higher y as the only variable left for Mary to link to. (The Constructional Integrity Constraint will insure that John is linked to χ and Bill to z.) Such an augmentation is not limited to persuade, of course; we will need it for all object control verbs. And as long as we are modifying the lexical entries of object control verbs in this way, for the sake of consistency we may do it for subject control verbs like try as well, giving lexical entries like (28). (28) try

syntax: [VP VP[to]] semantics: [ F , [F (x) F 1 ]] On the one hand, this move seems vaguely undesirable; after all, we have seen that the Constructional Integrity Constraint alone can handle the linking for at least the subject control cases, and

Control in Autolexical Syntax

115

we should be wary of introducing extra machinery unless it is absolutely needed. On the other hand, though, we are capturing the obvious generalization that for all the verbs in question, the semantic subject of the complement F is syntactically free. We could state this generalization as a kind of cross-modular lexical rule: (29) If a verb subcategorizes for a VP in the syntax and an F in the semantics, the semantic subject of this F is syntactically free. If we state this rule as part of the grammar, there is no need to specify the augmentations in (27) and (28) for specific lexical entries; they will automatically be part of these v e r b s ' subcategorization frames. In many cases (29) will be redundant, since the Constructional Integrity Constraint will do the same work; but in cases where the Constructional Integrity Constraint breaks down, rule (29) will step into the breach and ensure the right linking. 3.2 Controlling variables Another question we will have to address is this: how do we guarantee that the right variables appear in the right places in the semantic tree? For example, in the sentence John tried to leave, the understood subject of leave must be John and nothing else. (30)

S VP

NP

VP [-FIN]

V John

tried

V to

VP leave

116

Da vid Ka thman

This means that in the semantic tree, the variable appearing as the subject of the embedded proposition must be the same as the variable appearing as the subject of the main proposition, or at least we must somehow specify that they refer to the same individual. But what is to prevent a representation like that in (30) for John tried to leave, thus causing the sentence to mean something like "John tried to make it the case that somebody left"? And similarly, what will prevent the free variable in John persuaded Mary to leave from being something other than the variable associated with Mary, thus causing the sentence to mean something like "John persuaded Mary that he (John) should leave", or "John persuaded Mary that someone should leave"? Nothing in the theory as sketched so far prevents these possibilities; clearly we are going to need some additional machinery. There is an obvious generalization to be found in the control sentences we have looked at: the free variable in the semantic tree is the same as the next variable up the tree. In a subject control sentence such as John tried to leave, this next variable is the semantic subject (as in [21]); in an object control sentence such as John persuaded Mary to leave, it is the semantic direct object (as in [22]).13 As a first approximation, we can state this generalization as in (31). (31) A semantic variable which is unassociated to the syntax must be the same as the next variable up the semantic tree. This is essentially the same as Rosenbaum's (1967) Minimal Distance Principle, only it is stated in semantic rather than syntactic terms. When combined with the semantic subcategorization frames we saw in (27) and (28), this rule will insure the right distribution of variables in all the cases we have seen so far (e.g., figures [21], [22], and [26]). Technically, we probably also need a rule stating that variables linked to distinct lexical NPs must themselves be distinct, to insure that, say, John and Mary are not both linked to xs. I will assume a rule of this sort, but will not try to specify it in detail. As we might expect, the rule in (31) does not tell the whole story. For one thing, when an infinitival VP appears in subject position, (31) does not work at all, and the understood subject has a much wider range of interpretations; it can refer to some overt NP in the rest of the sentence, as in (32a), or it can have a general

Control in Autolexical

Syntax

117

reference, as in (32b) and (32c) (the so-called "arbitrary PRO" of Government and Binding): (32) a. b. c.

To behave himself in public would help Bill. To behave oneself in public would help Bill. To criticize oneself is foolish.

The obvious difference between infinitival VPs in subject position and those that are complements is that the latter are subcategorized for; somehow we need to distinguish between free variables which are specified in subcategorization frames and those that are not. One way to do this is to modify the semantic subcategorization frames for control verbs so that the subject of the complement F is not only specified as syntactically free, but it is also specified as being quantified by a special quantifier (call it p, following Sadock 1991b) which has the property of ensuring that the variable in question refers to the same entity as the next highest variable up the tree; in short, this quantifier instantiates rule (31).14 Thus the revised lexical entry for fry would be as in (33): (33) try syntax: [VP VP[-FIN]] semantics: [ F , [p p(x) F"1]] Since ρ will always occur in the subcategorization frames of control verbs, we can abstract it out of the specific lexical entries just as we did with the free variables, and we can revise our lexical rule (29) as follows: (34) If a verb subcategorizes for a VP in the syntax and an F in the semantics, the semantic subject of this F is syntactically free and quantified by p. Let us assume that in the default case, a syntactically free variable is free to refer to anything, and that ρ is only introduced when it is subcategorized for. Then each of the sentences in (32) will have a free variable in the semantics of its subject VP, but this variable will not be quantified, so it will be free to corefer to anything; it can be coreferential with another variable in t h e proposition, as in (32a), but it need not. For John tried to leave, though, the free variable is quantified by ρ and is thus restricted to being coreferential with the next higher variable as in (35).

118

David Kathman

(35) VP V John

tried

VP [-FIN] V

to

VP

leave F

I

->

Ο"2

Sadock (1991b) would have two different variables in the tree in (35), with the quantifier specifying that the reference of the lower one is the same as that of the higher one; I prefer to use the same variable in both places, and for all intents and purposes these options appear to be notational variants. Also, I am leaving the exact nature of ρ open to question; Sadock (1991b) assumes that it is part of a regular Q expression such as those corresponding to NPs, but it seems possible that ρ should be seen more along the lines of the existential quantifier 3, which binds a variable all by itself. I will leave such speculations for a future paper.

3.3 Reflexives and control In looking at our control rule (31), one thing which is interesting is how similar this generalization is to that governing reflexives; it is usually the case that a reflexive pronoun must be coreferent with the next higher NP in the syntax. We can represent a reflexive sentence such as John shaved himself with the Autolexical diagram in (36). The variable linked to the reflexive here must be identical to the higher variable, just as the free variable in control sentences must be identical to a higher variable. It is worth asking whether we could combine these observations into a single rule; specifically, we might ask whether the semantic realization of reflexives is the quantifier ρ which we posited for the free variable in control structures.

Control in Autolexical Syntax

(36)

119

S VP

NP

John

V

NP

shaved

himself χ

χ F First of all, despite the similarity noted above, there are differences between reflexives and subjects of infinitives. In control sentences, the variable quantified by ρ is the semantic subject of a subordinate proposition, and the variable it takes its reference from is in the matrix proposition. With reflexives, the two variables are in the same proposition, with the ρ variable generally being a semantic direct object or object of a preposition. When it is a semantic direct object, there is no problem; the control rule (31) will correctly identify the variables in a sentence like (36). However, when there are two variables higher in the same proposition, either of them may serve as the antecedent of the reflexive; this is illustrated by the sentences in (37):

(37) a. Johrij told Mary} about himself'.. b. Johnf told Mary} about herselfr I will not give Autolexical diagrams for these sentences, since the Autolexical semantics of prepositional phrases is somewhat uncertain, but it is fairly clear that we are dealing with single propositions here. Thus our control rule will not work here. However, we can modify the rule so it will work: (38) A variable quantified by ρ must be coreferent with a variable higher in the same F; if there is no variable higher in the same F, it must be coreferent with the next higher variable up the semantic tree. This new definition works, but at the cost of splitting up the two things we are trying to combine into one generalization. The first half of (38) handles reflexives, and the second half handles control structures, but the two halves are patched together in an

120

David Kathman

unsatisfying way. It is usually the case that the two halves of (38) will give the same result (i.e., the next variable up the semantic tree), but as (37a) shows this is not always the case. However, reflexives and infinitive subjects are not completely disjoint in their semantic distribution; there is one place where a reflexive can be the semantic subject of a subordinate proposition, namely the so-called "raising to object" construction. A sentence like John believes Mary to be a genius is r e p r e s e n t e d autolexically

as in (39) (following Sadock 1991b). (39)

VP

be a genius

The syntax of such sentences is the same as that of object control sentences (cf. figure [22]), but in the semantics there is only one

y; 1 5 this sentence could be paraphrased as John believes that Mary

is a genius. When the direct object is reflexive, we get the representation in (40). (40)

s VP

NP V John

believes

VP [-FIN] himself

I

V

to

VP

be a genius

ι

Control in Autolexical Syntax

121

The variable linked to the reflexive here is the semantic subject of the subordinate proposition, just like the free variable in control structures; thus the two halves of the definition in (38) are not totally disjoint, and we have better reason to consider this an interesting generalization. However, a skeptic might take (40) to be evidence that reflexive binding should be handled in the syntax, rather than in the semantics as we have been trying to do: although semantically this sentence is an exception to the rule that variables linked to reflexives are bound in the same proposition, syntactically it is perfectly consistent with the rule that reflexives are bound in the same sentence (this rule is, of course, Principle A of the Government and Binding binding theory). Although there is some basis for considering reflexives and control structures to involve the same phenomenon, the evidence is not as compelling as it might seem at first. Nevertheless, I will tentatively consider the semantics of reflexives to involve the same quantifier we posited earlier for free variables in control sentences (with the revised definition given in [38]), while recognizing that it might turn out that separate quantifiers will be needed. John shaved himself would thus have the revised structure given in (41). (41) VP John

V shaved

NP himself I px

The lexical entry for himself would then look something like the following: (42) himself syntax: NP[+OBJECTIVE] semantics: px Syntactically this word is specified as +OBJECTIVE, meaning that it cannot appear as the syntactic subject of a sentence (note the

122

David Kathman

ungrammatically of *Himself loves John or *John thinks that himself is a genius). Semantically it is a variable quantified by our special quantifier p. This differs from the lexical entry given in Sadock (1991b); there the meaning of a reflexive is a complete Q expression containing a variable and p, along with a specification that the entity in question is male and singular. I have not done the semantics this way for the tentative reasons given at the end of section 3.2 with regard to control structures; also, it seems unnecessary to specify in the semantics that the referent of himself must be masculine and singular, since such agreement facts are probably better handled in the syntax and/or morphology. However, nothing crucial hinges on the specific way ρ is instantiated. Given the above analyses, we can account in a fairly straightforward way for the interaction of reflexives and control if we make some natural assumptions. For example, John tried to shave himself would have the representation in (43), with two instances of ρ in the semantics: (43) Ν

John

tried

to

shave I F2

NP[+OBJ] himself I px

The topmost χ will be bound by a Q expression (John(x)), the middle variable will get its reference from this χ by virtue of rule (38), and the lowest variable will also get its reference from the top χ as long as we assume that rule (38) is transitive. Thus, all three arguments in the semantic refer to the same person (John).

Control in Autolexical Syntax

123

A sentence like John persuaded Mary to shave herself will work in a similar way (44). Here the lowest two variables get their reference from the semantic direct object y, which in turn will be bound by the Q expression Mary(y). (44) VP

NP

NP

VP [-FIN] VP

John

persuaded

Mary

to

3.4 Residual problems Although the analysis sketched above will account for the most important facts of control structures, there are a number of details which still need to be worked out. Probably the most well-known of these is the verb promise, which is a subject control verb but takes a direct object, thus violating the Minimal Distance Principle (and its analogue in the present theory, rule [38]). In John promised Mary to leave, it is John, not Mary, who is the potential leaver, meaning that semantic ally we have something like (45). The problem is how to ensure that the semantic subject of the embedded proposition is coreferent not with y, but with x; clearly the mechanisms we have so far will not do so. One posssibility which suggests itself is that the semantic subcategorization frame for promise contains not p, as with other control verbs, but a different quantifier, call it γ. This quantifier

David Kathman

124

would then be defined so that the variable it quantifies must be coreferent not with the nearest variable in the next higher proposition, but with the highest variable in the next higher proposition. (45) NP

Jo m

VP V ι

promised

NP ι

Mary

VP [-FIN] V ' VP to

©

leave

I Γ'

O" F"2 Γ' Admittedly this solution is somewhat ad hoc, but no more so than many other solutions that have been proposed for this problem; also it does fit in neatly with the analysis we have seen so far, and it does work. Another possible solution would be in terms of thematic roles, as suggested in such works as Sag — Pollard (1991); such a solution would necessitate somehow augmenting the Autolexical semantics with thematic roles, as in Schiller (1989). This road may be promising, but I will not pursue it here. Another control phenomenon which I have not addressed is adjunct phrases with an understood subject, as in (46): (46) a. b. c.

Muttering to himself, John walked into the room. John walked into the room muttering to himself. Bill hired Mary to please his boss.

Adjuncts are not subcategorized, so according to what we said earlier (section 3.2) they cannot contain the quantifier p. Yet it is clear that the understood subject of muttering in (46a-b) must be John. We need some way to account for this; I will not offer one here, except to suggest tentatively that pragmatic factors may be at work somehow. Note, though, that in (46c) the understood subject of please can be either Bill or Mary, the former interpretation implies that Bill is trying to curry favor with his boss, while the

Con trol in Autolexical Syn tax

125

latter interpretation seems vaguely defamatory to Mary. This ambiguity is consistent with our theory; since the phrase to please his boss is not subcategorized for, it cannot contain ρ in its semantics, and so its understood subject can be coreferent with either Bill or Mary.16

4. Conclusion The account of control phenomena sketched in the previous pages is not without its gaps, but it does account for the major facts in a fairly neat way. I hope to have shown here that an automodular model of grammar such as Autolexical Syntax can deal with familiar phenomena without resorting to such devices as transformations, and I also hope I have helped clarify some of the issues which have to be addressed in order to make such a model rigorous enough to be taken seriously by other linguists. Obviously there is still much work to be done, both on the issues raised in this paper and on Autolexical theory in general, but hopefully this paper has made a contibution to our understanding of language.

Notes 1.

2.

3.

A clarification: in this paper I will capitalize both words (Autolexical Syntax) when referring to the theory as a whole, but only the first word (Autolexical syntax, Autolexical semantics) when referring to specific modules within the theory. Some linguists in the Chomskyan tradition have, however, proposed automodular models quite similar in spirit to Autolexical Syntax-most notably Ray Jackendoff in his recent work (Jackendoff 1990: 16-18). And numerous theories outside the Chomskyan tradition, such as Tagmemics (Pike - Pike 1982) and Stratificational Grammar (Lamb 1966) are explicitly automodular in nature. At the very least, there are no empty categories corresponding to PRO in Government and Binding work, which is all that is relevant for this paper; whether or not there are null NPs such as are found in Generalized Phrase Structure Grammar (Gazdar et al. 1985) is an open question, as virtually no Autolexical work has been done on vWi-phenomena.

126 4. 5. 6.

7.

8.

9.

David Kathman It is probably best to think of the lexicon not as a separate component per se, but as a collection of links between fragments of the various components, as described in Jackendoff (1990:18) and Sullivan (1977). I here follow Pullum's (1982) analysis of English infinitival to as a non-finite syntactic verb heading its own VP. In this semantic diagram, as in all others in this paper, I have ignored such things as tense and aspect, which should probably be included in a full semantic account. Also, note that Autolexical semantics is unordered; I will generally draw the semantic diagrams so as to make them as compatible with the syntax as possible in terms of linear order, but this is not the only way they could be arranged. Though "subject" is properly a syntactic term, I use it informally here and elsewhere to mean the topmost argument in a semantic proposition, essentially in the sense of Dowty (1982). Similarly, the semantic "direct object" is (in our terms) an argument that combines with an F"2 to give an F"1. "Lexeme" here means essentially "lexical entry"; the syntactic projection of love is a transitive verb, while its semantic projection is a two-place predicate (F"2). I will sometimes also use the term "lexeme" to refer to the terminal elements in a given module. I am sidestepping here the issue of what to do about NPs with internal structure, such as the old man\ in the semantics, this structure will be located within the Q expression, as in (i), since variables have no internal structure. It seems that NPs as a whole correspond to variables in the semantic tree, but in their internal structure they correspond to Q expressions. There are a number of possible ways out of this dilemma; one would be to simply define scope relations out of our semantics altogether and plug the Q expressions into argument positions in the main semantic tree. i)

Q

man Another possibility is to keep the scope information in the semantics, but to have two types of linking between the syntax and semantics: an NP node is linked to a variable, while the component words of this NP are linked to the component parts of the Q expression which binds this variable. This could be accomplished by conceiving the syntax-semantics linking in terms of subscripts rather than lines, as

Control in Autolexical Syntax

127

in Jackendoff (1990), and allowing nonterminal nodes to have a subscript and thus be linked to another module. ii) NP

VP N'

Det A I The,a old b

K

V Ν I man c „

ate

NP, Det I a

N' I sandwich

The syntax and semantics of The old man ate a sandwich would then be something like (ii) and (iii), where elements with the same subscript are considered linked. The Constructional Integrity Constraint applies in the main semantic tree, and within each Q expression: nodes x,y, and ζ are in the same structural configuration in both modules, as are a, b, and c, and d and e. Details would need to be worked out, but such a system looks like it would work.

10.

By "higher" I mean "closer to the initial symbol of the tree", even though in this diagram the "higher" variable is actually lower on the page due to the semantics being upside down. 11. Note that structurally, this second χ corresponds to the infinitival particle to, but it would make little sense to say that these elements are linked; neither one is linked to anything, and it is just a coincidence that they appear in the same structural position. 12. Actually, the semantic subcategorization frame shown here violates the usual convention that only sisters of the head can appear in such frames, since the crucial element ( χ ) is a "niece" of the head. We can get around this by the same means used in Generalized Phrase Structure Grammar (GPSG) to subcategorize for prepositional phrases with a specific head: a feature on the F in question would

128

13.

14.

15.

16.

David Kathman indicate that its subject variable is free, just as the feature [ ί ο ] on a prepositional phrase in GPSG indicates that its head is the preposition to. Or, alternately, we could simply say that the convention in question does not apply to semantic subcategorizaion. For the time being I am ignoring subject control verbs which take a direct object, of which the standard example is promise, as in John promised Mary to leave. These will be discussed later. Note that we still need to specify that the variable in question is not linked to the syntax; this new quantifier ensures the correct distribution of variables in the semantic tree, but the linking problem in John persuaded Mary to kiss Bill (figure 26) remains the same. Here again, Autolexical semantics is analogous to the deep structure of classical transformational grammar: the deep structure of John believes Mary to be a genius would contain only one instance of Mary; which is raised out of the embedded clause into the matrix clause. In the Autolexical account, though, there is no movement; the two representations exist side by side. W e still have to account for the fact that the understood subject cannot have arbitrary reference; this, too, may involve pragmatic factors.

References Bach, Emmon 1979 "Control in Montague grammar", Linguistic inquiry 10: 515531. Brame, Michael 1976 Conjectures and refutations in syntax. New York: NorthHolland. Dowty, David 1982 "Grammatical relations and Montague grammar", in: Pauline Jacobson - Geoffrey Pullum (eds.), The nature of syntactic representation. Dordrecht: Reidel, 79-130. 1985 "On recent analyses of the semantics of control", Linguistics and Philosophy 8: 291-331. Gazdar, Gerald - Ewan Klein - Geoffrey Pullum - Ivan Sag 1985 Generalized phrase structure grammar. Cambridge, MA: Harvard University Press. Jackendoff, Ray 1990. Semantic structures. Cambridge, MA: MIT Press. Lamb, Sydney 1966 Outline of stratificational grammar. Washington, DC: Georgetown University Press.

Control in Autolexical Syntax

129

Manzini, Maria Rita 1983 "On control and control theory", Linguistic Inquiry 14: 421446. Pike, Kenneth - Evelyn Pike 1982 Grammatical analysis. Arlington, TX: Summer Institute of Linguistics. Pullum, Geoffrey 1982 "Syncategorematicity and English infinitival to", Glossa 16: 181-215. Rosenbaum, Peter 1967 The grammar of English predicate complement constructions. Cambridge, MA: MIT Press. Sadock, Jerrold M. 1985 "Autolexical syntax: A theory of noun incorporation and similar phenomena", Natural Language and Linguistic Theory 3: 379-440. 1991a Autolexical syntax: A theory of parallel grammatical processes. Chicago: University of Chicago Press. 1991b Passive and the cycle in an automodular theory of grammar. [Unpublished manuscript, University of Chicago.] Sag, Ivan - Carl Pollard 1991 "An integrated theory of complement control", Language 67: 63-113. Schiller, Eric 1989 The Case for Autolexical case. [Unpublished manuscript, University of Chicago.] Sullivan, William 1977 "A stratificational view of the lexicon", Language Sciences 46: 11-22.

On deriving the government constraint for incorporation and inflection* Steven G. Lapointe

In recent work, two distinct analyses of the phenomenon of incorporation have been proposed. One, offered by Baker (1988), is set within the general terms of the version of the Government and Binding theory formulated in Chomsky (1986; henceforth, the Barriers version). This account assumes that the element to be incorporated starts from a D-structure position which it is required to occupy by the Theta Criterion and the Projection Principle. The actual incorporation is then carried out as a special case of Move α where α = X°, i.e., a lexical node. The resulting S-structure contains the moved item as a sister to the host element to which it is bound on the surface, coindexed with a trace in the position it started from in D-structure. Baker attempts to show that the conditions determining where the incorporated item can move follow from more general principles already established in the Government and Binding theory. The second account, proposed by Sadock (1985), forms the cornerstone of the "Autolexical Syntax" approach to crosscomponent structural interactions. It assumes a version of the monostratal theory of syntax Generalized Phrase Structure Grammar (GPSG) formulated by Gazdar et al. (1985) and asserts that the properties of incorporation follow from the way in which lexical items in the syntactic representations of sentences are linked to their morphological representations. The analysis thus claims that the host and incorporated items appear separately in the syntactic structure, much as in the D-structures posited by Baker, and that they form a single, morphologically complex unit in the morphological representation of the surface word. Sadock goes on to argue that the positions that the host and incorporated items are allowed to occupy in the syntactic structure follow from the general conditions governing the association of elements across the syntax/morphology component boundary. While one can point to a number of convergent and divergent areas in comparing these two approaches, I would like to focus

132

Steven G. Lapointe

here on the corresponding constraints in the two theories that require one of the elements involved in the incorporation relation to "govern" the other element in the relevant syntactic structure, and in the relevant sense of " govern". Although the two sets of constraints actually make somewhat different claims, it will be convenient to have a single label for both of them, and for this p u r p o s e I will use the term "government constraint". My aim in the present paper then will be to show that there are problems with b o t h t h e Move X° and t h e Autolexical versions of t h e government constraints and to argue that the generalizations that we need to c a p t u r e in this area actually follow from separate, independently required constraints on syntactic and morphological structures. In particular, in the next section I will briefly present Baker's approach to incorporation, including the government constraints on X°-movement which he adopts, and examine several problems with that approach and the proposals of Pollock (1989) and Chomsky (1991). In section 2, I will do the same for Sadock's account of the government constraints on Autolexical associations, in the process identifying what the relevant generalizations about government a p p e a r to be in cases of incorporation and inflection, and I will briefly r e t u r n to Baker's analysis to consider s o m e f u r t h e r difficulties that the generalizations pose for that treatment. In section 3, I will give an alternative account of the government constraint in Autolexical terms that involves two independently needed principles, one a condition relating the kinds of items to which f u n c t i o n elements can c o r r e s p o n d a c r o s s syntactic, morphological, and logical structures, the other a condition which is an extension of Lieber's (1983) constraint on argument structures in compound forms. These principles, along with others that are standardly assumed in Autolexical work, will be shown to derive the generalizations about the government constraint on association identified in the preceding section, without requiring that there be a s e p a r a t e principle stipulating the government constraint in Autolexical theory. In section 4,1 will examine two cases investigated by Sadock (1991) that represent potential counterexamples to the generalizations about the government constraint established in section 2 and argue that these cases do not in fact create problems for the approach proposed here. I will conclude by summarizing the results obtained in the previous sections. Before we continue, it is necessary to consider two issues that arise in comparing the government constraints in an approach

The government

constraint

133

like Baker's with those in one like Sadock's. First, there is no movement in Autolexical Syntax, and so grammatical processes which in transformational terms can be described as moving upward or downward cannot accurately be so described in Autolexical terms. The situation is somewhat complicated by the fact that while both authors employ a notion of "government", the two versions are not entirely equivalent, as will become clear in the discussion to follow. In an attempt to maintain as strict a parallelism between the two theories as possible in light of these differences, I will adopt the following terminology. If X governs Y, in Baker's sense of "govern", I will refer to the movement of Y to X as "upward" and that of X to Y as "downward", in the standard way. On the other hand, in the Autolexical analysis, if X governs Y in syntactic structure (or by extension in logical structure) in Sadock's sense of "govern", t h e r e are two cases depending on how t h e morphological elements X', Y', the lexically specified counterparts to the syntactic elements X, Y, are related to one another. If X' is the morphological head to which Y' is attached, I will refer to the associations linking Χ, Y with X', Y' as "upward"; likewise, if Y' is the morphological head to which X' is attached, the associations linking Χ, Y with X', Y' will be termed "downward". Here, I am following Sadock's (1985: 404) definition of "morphological head" as "that morpheme . . . which determines the inflectional properties of the unit"; the definition is set up this way so that inflections will not count as morphological heads, while derivational affixes and the appropriate elements in compounds will serve as morphological heads. 1 The second issue has to do with the problems that arise as a result of the two accounts attempting to generalize Ν incorporation to different classes of facts. As noted above, Baker uses Move X° to handle Ν incorporation and grammatical relation changes; as we will see, this rule is also used to handle certain inflectional p h e n o m e n a as well. In contrast, Sadock uses Autolexical associations to account for Ν incorporation, certain kinds of cliticization, and certain inflectional facts which overlap with those that Baker has analyzed. Because I believe that solid arguments can be presented that grammatical relation change is not to be handled via X°-movement (Lapointe 1989) and that cliticization phenomena are to be handled in the interface between phonology and syntax or phonology and morphology rather than in the syntaxmorphology interface (Lapointe 1987; Woodbury, this volume), in the discussion below I will focus primarily on Ν incorporation by

134

Steven G. Lapointe

Vs and Preps as well as the several inflectional cases that the two authors treat.

1. Baker's account of noun incorporation 1.1. The account The intent of Baker (1988) is to formulate a Move X° account of noun incorporation and then to show that the processes that are often discussed in terms of changes in grammatical relations (e.g., passive, antipassive, causative, etc.) can be insightfully accounted for by means of parallel sorts of X°-movements. That is, what Baker wishes to do is to show that noun incorporation and the classical grammatical relation changing rules are different aspects of basically the same phenomenon, all being instances of Move X°. Although grammatical relation changes and their relations to noun incorporation occupy much of Baker's work, in what follows I will be focusing primarily on his discussion of noun incorporation and will only briefly touch on his analysis of one type of grammatical relation change, passive. The account of noun incorporation which Baker gives takes the following form. In a typical case, the head Ν of the direct object NP can be incorporated into the V of its clause, yielding a morphologically complex V form, as can be seen in ( l a ) and ( l b ) . (1) Noun incorporation (Onondaga; Baker's examples [3a], [ l a ] from Chapter 3, cited in H. Woodbury 1975) a.

b.

Pet Pat 'Pat Pet Pat 'Pat

wa?-ha-htu-?t-a? ne? o-hwist-a? PAST-3mS/3N-lost-CAUS-ASP the PRE-money-SUF lost the money.' wa?-ha-hwist-ahtu-?t-a? PAST-3mS/3N-money-lost-CAUS-ASP lost the money.'

The kind of derivation that Baker has in mind for cases like ( l b ) can then be sketched as in (2). (For expository convenience, I

The government constraint

135

follow Baker here and below in using the English equivalents of non-English morphemes as the terminal elements of the tree structures.) (2) a. (Simplified) D-structure for ( l b ) b. S-Structure S

S VP

NP I Pat lost 4_

NP I money

The D-structure for ( l b ) parallels that for ( l a ) as a result of the Theta Criterion and the Projection Principle, since they both involve the same V with the same subcategorization/thematic role requirements. Move X° then applies to produce the structure in (2b), in which the head Ν of the direct object is attached to the V in a syntactically created compound structure dominated by V*; the Ν is coindexed with a trace that remains in the direct object head Ν position. The two conditions which jointly form the government constraints of Baker's account are the Head Movement Constraint (3), originally from Travis (1984), and the Government Transparency Corollary (4). (3) Head Movement Constraint An X° may only move into the Y° which properly governs it. (4) Government Transparency Corollary A lexical category which has an item incorporated into it governs everything which the incorporated item governed in its original structural position. The Government Transparency Corollary is of particular importance in determining the derived structural properties in cases where more than one Move X° process applies or where Move X° processes interact with other types of movement and anaphora, while the Head Movement Constraint limits where X° nodes can move in the first place. Because issues involving,how the X° nodes move will

136

Steven G. Lapointe

be of greater concern to us in what follows than issues regarding the interactions of X°-movement with other rules, we will be focusing more on the Head Movement Constraint than on the Government Transparency Corollary below. Together with the Move X° analysis of incorporation, the Head Movement Constraint makes a number of correct predictions about the behavior of noun incorporation, as Baker emphasizes. First, noun incorporation is highly productive in those languages which employ it, and the incorporated Ν continues to exhibit referential properties. Both of these facts follow on Baker's account from noun incorporation's being a syntactic process in Government and Binding terms; productivity is a hallmark of syntactic phenomena, while the referential properties of the incorporated noun are attributed to its being bound to a trace in a syntactic position where it can have such properties. Second, in transitive clauses, the direct object head Ν can incorporate into the V, but the subject head Ν cannot. The Move X° account predicts this asymmetry because a direct object head Ν is properly governed by the V of its clause, and so movement to the V is allowed by the Head Movement Constraint, as can be seen in the derivation in (2). On the other hand, movement of the subject head Ν to the V will violate the Head Movement Constraint, since the V (= Y° in [3]) does not properly govern the subject (= X° in [3]). A third prediction that this account makes involves unergative versus unaccusative clauses; the subject head Ns of unergative clauses generally cannot undergo incorporation into the V, while those of unaccusative clauses generally can. This fact follows from what has already been discussed if we adopt the widely accepted analysis of these types of intransitive clauses due to Perlmutter (1978) and Burzio (1981), as Baker does. In Government and Binding terms, the account assumes that the subjects of unergative clauses start out in D-structure subject position, whereas the subjects of unaccusative clauses start out in D-structure direct object position and move to S-structure subject position. Then, the subject head Ns of unergatives should behave exactly like those of transitive clauses, and the subject head Ns of unaccusatives should behave exactly like the direct object head Ns of transitive clauses; i.e., the former should not be able to incorporate while the latter should. The Move X° analysis makes several further correct predictions. Thus, it is generally not possible to incorporate the head Ν of an object NP of a PP into the V of its clause, nor is it possible to

The government

constraint

137

incorporate the head Ν of an adjunct NP. This fact follows f r o m PPs and adjunct NPs being barriers to government in this version of Government and Binding, and so the V of a clause d o e s not properly govern the contents of either of these kinds of phrases. Hence, by the Head Movement Constraint the movement of the head Ns in these phrases is disallowed. Finally, it is generally possible to incorporate the head Ν of the object of a PP into the Prep of that phrase. Again, this must be the case on Baker's account, since such Ns bear just the same structural properties to the Prep that the direct object head Ν bears to the V of its clause. Now, in addition to making all these correct claims about noun incorporation, the Head M o v e m e n t Constraint, along with the Government Transparency Corollary, has the following important property: apparently, neither it nor the Government Transparency Corollary has to be added as a separate principle to the Barriers version of Government and Binding. In particular, Baker sets out to prove formally that the Head Movement Constraint follows from the Empty C a t e g o r y P r i n c i p l e and that the G o v e r n m e n t Transparency Corollary follows from other properties of that framework. If these conclusions are correct, then t h e y add considerable f o r c e to Baker's analysis, for they show that the properties of incorporation and grammatical relational processes are a consequence of the Government and Binding theory without any additional machinery beyond the assumption that M o v e α applies to word level as well as phrase level syntactic categories. 1.2. Some problems One problem with the M o v e X° account is that technically, Baker does not prove either of the results just mentioned concerning the derivability of the Head Movement Constraint and the Government Transparency Corollary from more basic principles of the Barriers version of Government and Binding. Instead, what he does is to change several of the key definitions used to establish the notion "barrier to government"; these changes in effect lead t o a new version of the Empty Category Principle and the overall theory, which w e might call Empty Category Principle* and Barriers*, respectively. Next, Baker observes that for purposes of the proofs he is giving, the Head Movement Constraint needs to be modified in several ways, yielding what w e might term the Head Movement Constraint*. Finally, Baker proves formally that the Head Movement

138

Steven G. Lapointe

Constraint* follows from the Empty Category Principle*, and that the Government Transparency Corollary follows from Barriers*. Unfortunately, the claim that Baker makes is that the Head Movement Constraint and the Government Transparency Corollary, as given in (3) and (4), without the changes just mentioned, follow from the original Barriers framework, and strictly speaking he does not prove this claim. Although Baker admits to making the modifications in order to facilitate the formal proofs, it is easy to miss the admissions because they appear either in footnotes or in the middle of the technical discussion surrounding the proofs themselves. Nonetheless, it is important to emphasize the fact that Baker has not shown that the Head Movement Constraint and Government Transparency Corollary are derivable from the usual Barriers version of Government and Binding.2 We might, perhaps, grant Baker the benefit of the doubt here, since after all the exact formulation of theoretical principles is in a constant state of flux in linguistics, and the principles of Government and Binding which he modifies are no doubt in need of some reformulation on both empirical and theoretical grounds. On the other hand, some of the changes that he adopts have far-reaching consequences for the claims made by the rest of the framework. Baker does not attempt to evaluate those consequences, but without such an evaluation, and in particular without a demonstration that the implications lead to a more adequate account of the syntactic facts, we cannot be fully confident about claims that incorporation and grammatical relation change follow directly from the Government and Binding theory. Setting aside this issue and focusing more closely on the Head Movement Constraint, it is interesting to note that Baker has proposed an analysis of some facts in a different context which contradicts the spirit, if not the letter, of the Head Movement Constraint. In attempting to account for the peculiar set of properties exhibited by gerund phrases in English—that they have clausal properties inside the phrase and NP properties outside—Baker (1985) offers the following sort of analysis, (a) Gerunds are Ss at D-structure and NPs at S-structure. ( b ) The gerundive affix -ing starts out in INFL in D-structure and moves to the V and attaches to it by Move a. (c) The lexical entry for -ing contains the information in (5); thus, although the affix must attach to the V, it carries nominal category features, a situation which induces the derived form to be a Ν and forces the dominating phrase nodes to recategorize as N-type phrases by the conventions

The government

constraint

139

of the X-bar theory, resulting in a NP at S-structure. The kind of derivation that Baker envisions for a gerund phrase like John's singing the aria is given in (6). (5) Lexical information for -ing -ing. - base generated in INFL - carries Ν category features - attaches to Vs (6) D-structure:

S-structure:

INFL' (=S) NP I John

VP INFL I -ing V I sing

NP

NP I the

NP I John V I sing

NP Ν -ing

NP I the

It is crucial that the movement in such cases does not leave a trace because, as Baker observes, if it did, there would still be an INFL node in the resulting S-structure, and so we would still have an INFL', that is, an S, at S-structure, rather than the NP that we need if we are to account for the nominal properties exhibited by gerunds. Two points need to be made about this analysis. First, there is a further and theoretically more significant reason why the movement shown in (6) cannot involve the leaving of a trace, which Baker does not mention-doing so would violate the Head Movement Constraint and also the Empty Category Principle. Second, it is unclear how Move α is supposed to operate in these cases so as not to leave a trace, since it is a bedrock assumption of Government and Binding that the syntactic operation of Move α always leaves a trace. This is yet another place where Baker makes a potentially sweeping change in Government and Binding, in this case permitting some instances of Move α not to leave a trace while others must, without examining what even some of the immediate implications of such a move might be for the theory, e.g., exploring under what conditions a trace is or is not left.3 One fairly straightforward suggestion concerning when a trace is not left would be to say that this happens only when a function element (f-element) is moved; whenever a major lexical category,

140

Steven G. Lapointe

at t h e word or phrase level, is moved, the standard assumption that a trace must be left behind would continue to hold. However, it is not entirely clear that such a change alone will suffice. First, two distinct kinds of analyses have been proposed within the Government and Binding framework for handling cases where an f-element appears in a higher syntactic position than the surface item to which it is ultimately attached. On the one hand, there are accounts like Baker's analysis of gerunds above, as well as "Rule R" of Chomsky (1981) which is intended to carry out the movement of TNS from INFL to the main V, w h e r e t h e f-element moves downward to its host. On the other, there are accounts like that for t e n s e elements given in Chomsky (1986) where the main V moves to INFL to attach to the TNS element and that for passive in Baker (1988) where the main V also moves to INFL, this time to attach to the passive marker. In these analyses, a major lexical n o d e moves upward in the standard way, leaving a trace, rather than the f-element moving downward. However, there is nothing in the Government and Binding theory which determines which of t h e s e a p p r o a c h e s we should adopt; i.e., theoretically, t h e r e is nothing preventing an account of gerunds in which the V moves upward to INFL to attach to the gerund marker, nor is there anything preventing an analysis of passives in which the passive marker moves downward and attaches to the V without leaving a trace behind. Until it is determined exactly which approach is to be employed in which kind of case, simply saying that f-elements do not leave t r a c e s when they move leaves too many q u e s t i o n s unanswered. Second, and perhaps more importantly, positing a condition that permits f-element movements not to leave a trace is too weak a constraint. One of the chief motivations for proposing the trace theory in the first place, along with the attendant conditions which govern where traces may legitimately occur, was to capture the generalization that syntactic elements always appear to move upward in a structure. That generalization was, of course, based on movements involving major lexical categories, and in fact major lexical phrases; the a s s u m p t i o n that lexical categories, and specifically f-elements, can move immediately opens up the question of directionality of movement for such items. Permitting the movement of f-elements not to leave traces would allow such m o v e m e n t s downward in a structure, as in the case of Baker's analysis of gerunds, but it would not require them to do so. In effect then, such a condition would claim that f-elements are exempt

The government constraint

141

from the usual generalization about upward movement for major lexical categories, and hence they should be able to move either upward or downward. But this does not seem to be the c o r r e c t empirical generalization about such cases. It appears that f-elements in general only m o v e downward; in particular, this will happen whenever an f-element has to appear at some syntactic level in a higher position in the structure than the position that it finally occupies in the surface structure. Such cases abound in natural languages; Carlson (1983a, b ) examines a large number of these. Thus, tense, aspect, number, definiteness, and comparison markers regularly have to be analyzed as appearing at the maximal phrase level, or outside of the maximal phrase, of the lexical head that the marker must ultimately be attached to. Indeed, as Carlson points out, the classical transformational analysis of tense in English involving Affix Hopping Cor "Rule R" in Chomsky 1981 mentioned a b o v e ) is an attempt to account exactly for these kinds of effects. However, the reverse cases, where an f-element moves upward to attach to a host, appear not to exist.4 Some care must be taken here to distinguish two separate kinds of attachments. The sorts that we have been discussing fall under the traditional category of inflection; if they are analyzed in terms of syntactic movement rules, then somewhere in the analysis the f-element appears in a higher structural position than that of the item to which it is attached, regardless of which element actually moves in the analysis. There is, of course, a second type which falls under the category of cliticization, and here f-elements may indeed move either upward or downward. An example of upward movement would be English contracted auxiliary attachment, where either on the GPSG analysis of English auxiliaries (Gazdar - Pullum - Sag 1982) or the recent Government and Binding analysis (Chomsky 1986), the contracted auxiliary moves upward to attach to the rightmost item in the subject NP which governs it, as shown in (7). On the other hand, there are cases where the f-element clitic moves downward, as in an analysis like that offered by Sadock (1985) for the Latin bound form conjunction -que, illustrated in (8). As indicated here, the clitic element -que moves downward from its position under the CONJ node when it attaches to the first word of the NP to its right.5 In any case, it is clear that the kinds of f-element cases that Carlson is concerned with involve the inflectional rather than the clitic sort.

142

Steven G. Lapointe

(7) a.

The president's

arriving

in the

morning.

b. GPSG structure for (7a) (after Gazdar - Pullum - Sag 1982) S VP I V'

NP the

president

V [+AUX] I 's ι (8) a.

duo

VP V' I V I

PP in the

morning

arriving

puerl

fidelisque

canis

two boy, pi faithful, msg+que dog, sg 'two boys and a faithful dog' b. Structure for (8a) (modified from Sadock 1985) NP NP duo puerl

CONJ I que

NP I N' Ν I

canis

Since Baker is also concerned with this kind of f-element, Carlson's observations are relevant here. Hence, it seems reasonable to conclude that permitting f-elements to move without leaving a trace would not by itself account for the apparent fact that such elements do not move upward but instead move exclusively downward (or alternatively, stay in place and require a major lexical item to move upward to attach to the f-element).

The government

constraint

143

1.3. A note on more recent proposals Although this section has been devoted to an examination of Baker's account of Ν incorporation, it is appropriate to mention a few words about the proposals of Pollock (1989) and Chomsky (1991), since these bear directly on the issues that we have just been considering, i.e., the properties of major item movement vs. felement movement. Because Pollock's article is the more detailed of the two, I will concentrate on that work in my remarks here. In analyzing the placement of adverbs, quantifiers and negation in English and French VPs, Pollock assumes an expanded version of INFL in which both T(ense) and Agr(eement) head their own separate phrases (TP and AgrP, respectively) and in which TP takes over the function of the earlier IP as the phrase corresponding to traditional S and AgrP takes over the defective barrier properties of IP (Chomsky 1986). On these assumptions, a V in English or French moves from its base generated position to Agr and then to Τ via X°-movement; this constitutes the "normal" way in which Agr and Τ elements become attached to their Vs. Much of the paper is concerned with exploring cases in which this V-movement has to apply obligatorily or fails to apply, given that the rule is a subcase of Move X°, itself an instance of Move a, which is a completely general and optional rule. In particular, Pollock discusses two cases where V-movement does not apply: (a) The V cannot move without contravening some general principle of the theory or (b) a null "copy" V moves from Agr to T, leaving the Τ and Agr elements still morphologically unattached to an appropriate base V. In both of these cases, Affix Movement lowers the items onto the main V. An example of (a) involves French infinitives, which Pollock argues cannot move up to the higher Τ or Agr nodes, and so the infinitive endings must then move down to the V (e.g., ne pas partir aujourd'hui 'not to leave today'); an example of (b) is English simple present and past tense Vs, which Pollock contends involve null counterparts to the auxiliary do that occurs in corresponding negative sentences (The train arrived on time versus The train did not arrive on time/* The train not arrived on time/* The train arrived not on time). Pollock's analysis therefore permits both of the kinds of movements that previous Government and Binding accounts have allowed-movement of a major item upward to an f-element and movement of f-elements downward to major items. The paper does not, however, shed much light on the issues examined above,

144

Steven G. Lapointe

that is, why these appear to be the only two possibilities. Among the several major questions that he raises concerning the general V-movement approach, Pollock does ask (1989: 371), "Why does UG allow for Affix Movement, a lowering rule?" but this is the one question for which he does not explicitly summarize a response. At one point during the treatment of V-movement blocking in French infinitives, Pollock offers several ways in which a lowering rule like Affix Movement might be implemented within the Government and Binding framework he assumes: (a) Affix Movement involves an instance of Move a, followed by deletion of the traces left by the Τ and Agr elements, since these are essentially "empty", by which Pollock means recoverable; Cb) the rule involves movement at PF and does not leave a trace; (c) it involves PF movement and does leave traces, but because PF rules derive "flat" structures (Chomsky - Halle 1968), no c-command violations result; or (d) from some 1987 class lectures by Chomsky, the rule involves LF movement of V+T+Agr back through Agr to T, in effect undoing Affix Movement at LF but also involving the expected upward movement of a major item. However, Pollock does not attempt to evaluate these various possibilities. Each of t h e s e alternatives raises serious questions which cannot be considered in the detail they deserve here, since each presents complex issues for the analysis of the English and French auxiliary v e r b systems, the treatment of which lies well beyond the scope of the present paper. Nevertheless, two obvious questions arise with regard to possibility (a), namely, (i) how can Τ and Agr move past VP, an inherent barrier, to attach to the V, given that there are no major lexical heads present to void barrierhood in this case, and (ii) why are the traces of Τ and Agr more recoverable than those left by Move NP or Move WH, permitting the V-movement traces, but not the others, to be deleted? Question (i) might be answered by saying that these elements are adjoined to the VP node before moving into the VP, repeating in reverse the sort of adjunction that is involved in V-movement. On t h e other hand, question (ii) might find an answer in θ-theory: If f-elements like Τ and Agr do not assign lexically specified θ-roles in their b a s e positions, then this may be sufficient to consider their t r a c e s semantically "empty" and hence subject to recoverable deletion. 6 This latter answer comes very close to the possibility discussed above, wherein movement of f-elements does not involve the leaving of traces, but as we have already seen, such an approach appears not to go far enough. In section 3 below, I will formulate an

The government constraint

145

Autolexical account of the possible structures which is more like (d), in that it crucially involves a level of logical representation but, in concert with earlier Autolexical work, does not involve movement or traces. 1.4. Summary We have seen that while Baker's Move X° proposal correctly accounts for a wide range of facts about Ν incorporation, a number of problems are raised by that proposal. In particular, in light of further analyses presented by Baker and other Government and Binding researchers, the assumption that X° categories can move seems to lead to potentially major modifications in the Government and Binding framework, changes whose consequences for government constraints on moved elements have not been fully examined. In particular, under their present formulations those changes appear unable to account for known facts about the movement possibilities for inflectional f-elements.

2. Sadock's account of noun incorporation 2.1. The account Turning now to Sadock's (1985) Autolexical account, the example in (1) above would be analyzed in roughly the following way. First, the syntax, which employs a version of GPSG, generates a single structure for the sentence; for (1), that tree would be essentially the one given in (2a) above, repeated here in slightly modified form as (9a). Independently, the morphology generates a series of morphological trees, which in the case of (1) would look like (9b). Finally, the conditions regulating the linking of lexical material across the two types of structures would permit the associations in (9c). (It is important to note that these associations incorporate the changes in the Autolexical account suggested in Lapointe [1987, 1991/1988], which require that maximal lexical categories, and not the lexical terminal material itself, are the items that are associated across these structures.)

146

Steven G. Lapointe

(9) Autolexical analysis for (1) (Sadock 1991; Lapointe 1987, 1991/1988) b. Morphological structures a. Syntactic structure S V' Ν NP VP I N. V I Pat V NP Ν money lost I I I lost Pat Ν I c. Associations Syntax: Morphology:

money Ν I Ν

Ν V

V"

Here the single morphological node V* is allowed to associate with the two nodes V and Ν in the syntax because it contains the lexical items dominated by those two syntactic nodes. The major condition determining the syntactic relations that can hold between items that are autolexically linked in this theory is given in (10). The two kinds of structural configurations allowed under this government constraint are given in (11). (lO)Sadock's (1985) Condition VII' If a lexeme L combines syntactically with Xn to form a phrase, but morphologically with Y° to form a word, then Y° must be the syntactic head of a phrase that L governs. (11) a. D X (=YP)

YP

γο

w

w

What Sadock means by "govern" here is intended to include phrases subcategorized by a V and in addition subject NPs selected by Vs. Thus, in (11a), if C = VP, L = V, Xn = YP = NP, and Y° = N, then we will have the situation that we need when an object Ν incorporates

The government

constraint

147

with its V. In ( l i b ) , if C = VP, L = V, YP = NP, and Y° = N, then we have the situation where a subject Ν incorporates with its V. This sort of analysis can account for many of the same facts about Ν incorporation that the Move X° account can. Hence, the fact that Ν incorporation is highly productive and that the incorporated Ν continues to exhibit referential properties follows from the Ν having an independent status as the head Ν of a NP in the syntactic structure for the sentence. Furthermore, it should be possible to incorporate the object head Ν into its V, or the object head Ν of a Prep into the Prep, since this falls under case CI la). On the other hand, it should not be possible to incorporate the Ν from an adjunct NP or PP into a V, since these phrases are not subcategorized for by the V. There are some problems with the Autolexical analysis of Ν incorporation, but these would seem all to hinge on the exact definition of the term "govern" that one adopts in the government constraint in (10). For example, as formulated, that constraint would permit incorporation of the subject Ν with its V in general, whereas Baker has presented evidence that this only occurs with intransitive Vs which arguably fall in the unaccusative class. In addition, Sadock's account would seem to permit the incorporation of a Ν to a V if the Ν is in a NP inside a PP that is a subcategorized for complement of the V, but as Baker argues, the object head Ns of such NPs are in general not incorporable with the V, regardless of whether they are in adjunct or complement phrases. However, if we instead used a strictly structural notion of "govern" with something like the Barriers restrictions on it, then we would avoid the last problem, since on that assumption the contents of no PP are governed by an external V. Likewise, if a satisfactory Autolexical account of unaccusatives can be formulated in which the V selects for the subjects of unaccusatives, but not for the subjects of unergatives or the subjects of transitives, then the problems with these cases would be handled as well.7

2.2. Some problems Considerations like those just discussed have led Baker to suggest that the Head Movement Constraint in (3) is equivalent to Sadock's Autolexical condition in (10). If this is so, we might expect that parallel problems would arise for the Autolexical account in terms of the government constraint underspecifying the properties of

148

Steven G. Lapointe

association for f-elements versus major lexical items. This indeed appears to be the case. In particular, nothing in (10) says anything about the conditions under which L incorporates downward or Y° incorporates upward. Thus, Sadock (1985, 1987) and Lapointe (1987, 1991/1988) have suggested a number of analyses which conform to the generalizations in (12). (12) Correlation between category type and direction of association a. In (10), when L is a major lexical item, association is upward. b. In (10), when L is an f-element, association is downward. In terms of the definitions of upward and downward associations given in t h e i n t r o d u c t o r y section, (12a) holds when, in morphosyntactic representation, L is the head and Y° attaches to it, while (12b) holds when Y° is the head and L attaches to it. Two caveats about (12) are in order. First, it is not obvious that f-elements subcategorize for sister phrases that they operate over in syntactic structures, so that the government requirement in (10) is no longer met. To avoid this problem, we can relativize the notion of "govern" so that in the case of f-elements the straightforwardly structural notion holds (i.e., the f-element simply c-commands another phrase). Indeed, although Sadock (1985) does not explicitly comment on this difference between major lexical items and f-elements, it would appear from the analyses he presents that he tacitly assumes something along these lines. Second, the reference to (10) in (12) clearly must mean possibility (11a) and not ( l i b ) , since under any reasonable definition of "govern", if L is an f-element, then there is no way that it is going to govern Y° in ( l i b ) . Hence, in what follows I will focus primarily on structures like (11a). To take a few examples, the analysis of Ν incorporation given in (9) illustrates the generalization in (12a). An example of (12b) would be the Autolexical analysis of tense affixes outlined in Lapointe (1991/1988). The basic idea is that tense affixes have a representation in logical representation and in morphosyntactic r e p r e s e n t a t i o n but not in syntactic structure. In logical representation, the tense element is an operator taking propositions (PROPs) into propositions; in morphosyntactic representation, the tense element is a simple inflectional affix attached to the V. On this account, a simple sentence like The Cubs won would have an analysis like (13).8

The government

constraint

149

(13) a. Logical representation b. Morphosyntactic Representation PROP TNS Ν V PROP Det [+past] I PRED PAST ι R the Cubs won I THE CUBS WIN c. Association

_ ^ Det I Logical Representation TNS LSP

.. Ν I Τ R

V~

^ — I Morphosyntactic Representation ——I

However, among the numerous analyses involving incorporation or inflection presented in the works cited above, none deal with the opposite kinds of structural configurations and associations. This fact no doubt is a reflection of the generalizations about movements of the two types of lexical items discussed in the preceding section. We may formalize these generalizations as in (14) and present an example of each of the kinds of associations that do not seem to occur. (14) Two unwanted types of association: a. If L in (10) is a major lexical item, the association is downward. Morphosyntactic representation Ν Ν + head Associations Syntactic structure Morphosyntactic representation

Ν Ν

150

Steven G. Lapointe

b. If L in (10) is a function element, the association is upward. * Logical representation Morphosyntactic representation PROP Tns TNS PROP V [+tense] TP PRED R

"

:..

head

Associations Logical representation Morphosyntactic representation

R

TNS Tns

Intuitively speaking, the problem with an analysis like that sketched under (14a) is that the V seems to be taken as an argument of the N, and this seems backwards; moreover, the problem with the analysis under (14b) is that an element that should be an inflectional marker is acting as the morphological head of the form, which also seems incorrect. Unfortunately, neither of these possibilities is ruled out by the government constraint in (10) or by any other conditions that have been proposed in the Autolexical literature. 9 It is important to observe that (14b) is the closest possibility in the Autolexical system to the kind of V-movement permitted in Government and Binding theory which was discussed above in section 1.3. As we have seen, on that analysis, the V moves in the syntax from a position like the one occupied by R in the logical representation of (14b) up to a position like that occupied by TNS, in the process creating a derived syntactic constituent like the one given in the morphosyntactic representation of (14b) where the [+tense] element is the head. That is, the derived constituent after V-movement looks like (15a) rather than (15b). (15) a.

[TNSV + TNS]

b. [V

V + TNS]

However, in spite of the fact that V-movement yields a TNSheaded constituent like (15a) in the syntax, in the morphology the corresponding structure has to be the V-headed one in (15b), even

The government

constraint

151

on the Government and Binding theory, since morphologically, the TNS element is an inflection and inflections are not morphological heads-the base rather than the inflection typically determines the lexical category of the derived form, and the base generally determines properties of the inflection, rather than the other way around. It is unclear how the discrepancy between a syntactic constituent like (15a) and a morphological one like (15b), a discrepancy that has not received much attention in the literature, is to be resolved within Government and Binding.10 Nonetheless, it is clear that even the V-movement approach needs to have a morphological structure that is V-headed, rather than TNS-headed as in (14b). Furthermore, in Autolexical analysis, inflectional elements are generally treated as not having a representation in the syntax. So, the only place a TNS-headed constituent can arise is in logical representation or morphosyntactic representation. This option apparently must be allowed in logical representation, since at that level, the TNS element acts as a functor operating on PROPs, as illustrated in the analysis sketched in (13). However, we do not want this option open to us in morphosyntactic representation, essentially for the reasons noted in the preceding paragraph-TNS as an inflectional item is not the head of the constituent. Thus, what in movement terms would be the necessary TNS-headed derived structure (15a) is permitted in neither the syntax nor the morphology on the Autolexical approach. Moreover, the V-headed structure (15b) which is needed in the morphology cannot be involved in an analysis like that in (14b). The reason for this has to do with the way that directionality of association is defined in Autolexical terms. If the structure in morphosyntactic representation were V-headed, then the association would be downward, and so we would have a case like that described in (12b) and illustrated again in (13). Hence we find not only that the Autolexical analysis does not allow the kinds of TNS-headed structures in syntactic structure or morphosyntactic representation that are required on the Vmovement account, but also that it does not permit the requisite direction of association in such cases. It is therefore literally not possible to define association analyses that correspond to Vmovement derivations within the Autolexical framework as it is currently formulated. Other things being equal, this should count as an advantage for Autolexical analysis over movement approaches, since the former disallows a class of analyses which

152

Steven G. Lapointe

the latter permits. The matter cannot be settled at this point, however, since other things are not yet equal; that is, an Autolexical account of the English and French auxiliary V systems has yet to be formulated, and until one is, and it is shown to be able to cover much of the empirical ground that Pollock's (1989) account covers, this advantage of the Autolexical approach remains more potential than real. Nevertheless, these considerations suggest that the Autolexical account may well be on the right track in this area. To return for a moment to Baker's (1988) account of X° movements, it is possible that his analysis of passives sketched in (16) represents a further example of (14b), since it involves the movement of the main V up to INFL where it attaches to the -PASS morpheme. (16) Baker's (1988) analysis of passive S

The problem with saying that this definitely is an example of type (14b) is that Baker never makes completely clear what the status of -PASS is with respect to the major lexical item/f-element distinction. He argues at length that this morpheme must be regarded as a nominal argument, since on his account it receives the external theta role of the V, and in some languages it receives abstract Case as well, two properties that normally only correlate with nominals. However, he never explicitly identifies -PASS as a Ν or a NP; his example tree structures do not clarify the point, since they are drawn like the tree in (16) without any labelled nodes between INFL and -PASS. The more recent paper by Baker - Johnson - Roberts (1989) on passives claims that the morpheme is to be viewed as a clitic like the Romance clitic pronouns, suggesting that perhaps we should view -PASS as an NP, but they go no further in specifying the exact categorial nature of the -PASS morpheme within INFL itself. Thus, (16) may or may not be an illustration of (14b). In connection with the overall issue of the conditions under which different types of elements are allowed to move/attach

The government

constraint

153

upward or downward, it is interesting to note that the recent paper by Baker - Johnson - Roberts (1989) makes no definite claim about the issue of whether the -PASS morpheme (-en in their paper) moves downward to the V, or the V moves upward to INFL as in Baker's original account. In fact, at one point (1989: 243-244, footnote 21), the authors admit being unable to decide how to choose between these two alternatives. However, as we will see, the analysis of the above generalizations given in the next section provides a way to help decide this issue.

2.3. Some further possibilities Up to this point in the discussion we have considered various possibilities allowed by the general structure in (11a) that are determined by assuming that Y° is a major lexical item while varying whether L is a major lexical item or an f-element and whether association is upward or downward. This leads to the four possibilities that we have examined in (12) and (14). However, if we modify (11a) to the more general structure in (11a') below, a structure which we will be assuming in the remainder of our discussion, there are four additional logical possibilities that result from setting Y° as an f-element, holding that assumption constant, and then varying L and the direction of association in the same way as before. This yields the possibilities listed in (17).11 (11) a.'

C

W (17) Possibilities for Y a function element a. L is a major lexical item, and Y associates upward to L. b. L is a major item, and L associates downward to Y. c. L is an f-element, and Y associates upward to L. d. L is an f-element, and L associates downward to Y. However, these cases do not seem to offer grammatically plausible alternatives. (17a) would involve structures where an

154

Steven G. Lapointe

f-element operates in a lower phrase but must incorporate with an element in a higher phrase, as shown in (18). (18)

Logical representation ^ ^ PRED

Morphosyntactic representation V*

(Here and in the remaining structures in this section, I will forego explicitly associating the crucial elements and will instead simply indicate by means of arrows in phrasal structures the direction of the association.) If LP corresponds to (in symbols, S' and FN ~ COMP, then we would have a case in which inflectional material corresponding to functor properties of COMP in a lower clause appeared on the V in the higher clause. If LP ~ NP and FN ~ Det, we would have a case in which features of Det of an object NP occurred on the V. Such cases as these seem highly implausible, and I am not aware of any, again putting aside cases involving cliticization. Part of the problem here is that functors operating at a particular phrase level in logical structure can only correspond to elements that occur at that level or lower in the syntax and morphology. So, if a functor corresponds to an element that attaches to a higher item in the syntax and morphology, we expect the FN in logical representation to operate at a correspondingly higher point in the structure. 12 Case (17b) would involve structures like those in (19). (19) Syntactic structure

Morphosyntactic representation FN* FN (h^d)

If X° = Ν, Y = AP, and FN = Deg, then (19) would be a case where the Ν associates downward to the Deg item of a modifying AP.

The government

constraint

155

Again, such cases appear not to exist. Such instances are presumably to be ruled out for reasons parallel to those disallowing possibility (14a). The problem with (17c) and (17d) is that both the items are f-elements, and all the f-elements we have been discussing so far are affixes (or morphological operations). As a result, there will not be a stem for one or the other of the two f-element affixes to attach to (or for the morphological operations associated with the f-elements to operate on). So it appears, at first blush, that neither of these two possibilities are allowed. However, there is one case where (17d) might be grammatical, namely, where the lower felement is represented by a function word in syntactic structure which can be inflected, as shown in (20). (20) Logical representation

Morphosyntactic representation

LP FN,

LP .FN2 . .>

In such a case, there will be no difficulties concerning the scope of FN,, as there is with case (17a). Moreover, FN, is associating downward, as the f-element does in the legal case of inflections in (12b). Thus, taking FN! ~ COMP and FN2 ~ INFL, it should be possible to arrive at a legal case where features of the COMP appear on INFL; Vata (Koopman 1984) appears to be a language that exhibits this possibility.13 Another possible example of this type would involve taking FN, ~ Deg and FN2 ~ Quant. Then in English, if -er occurs in a higher Deg node and few occurs in a lower Quant position at some level of analysis, forms like fewer would show the desired association. Although inflected function words provide a potentially rich source of topics bearing on problems of fundamental concern to Autolexical theory, the status of the inflected function word in any given case with respect to the major item/f-element distinction is often difficult to settle. I will therefore set aside the treatment of such items in the discussion below. It is important to emphasize that there are a number of grammatical processes that superficially have the properties of

156

Steven G. Lapointe

(17a) and (17c) which we would also want to exclude from consideration here. Thus, potentially, quantifier raising to COMP position (May 1985) might be viewed as a case like (17c) in surface syntactic structure, where L = COMP and Y = the base position of the quantifier. However, the structural relations involved in this construction exhibit properties which are not like those typically found with incorporation, and as a result no one has suggested that they should be given an Autolexical treatment. Likewise, the occurrence of person/number/gender markers on embedded COMPs in West Flemish discussed in Sadock (1991) based on work by Smessaert (1988) would appear to be another case of (17c), with L = COMP and Y = INFL; however, the data in this case would seem to admit of an analysis in terms of agreement or morphological government of the relevant features that are realized as the person/number/gender markers without necessarily requiring the sorts of cross-component associations that we have been considering. In addition, the phenomenon of quantifier float in English might be viewed as a counterexample to the claim that (17a) is not possible, or perhaps even that (14a) is possible, but the issue hinges very strongly on the analysis that one adopts for auxiliary verbs and on the status of the positions of the autolexically linked items, and I will not go through the various possibilities here. In any case, it seems reasonable to exclude these sorts of examples from consideration here. 2.4. Summary The facts that need to be accounted for, then, can be summarized as in Table 1. The problem now is to devise a set of conditions which will account for the distribution summarized in the table. At this point, we could simply attempt to state conditions which directly stipulated which possibilities are allowed, but a far more interesting solution would be one in which we attempted to derive this distribution from more general conditions regulating the behavior of lexical items of the two types. In the next section I will propose such an account in terms of the Autolexical approach.

The government

constraint

157

Table 1. Possibilities for association between L and Y in CI la') Cases Category of L Category of Y Direction Possibility of association allowed? a. b. c. d. e. f. g· h.

(12a) (14a) (14b) (12b) (17a) (17b) (17c) (17d)

major item major item infl. f-element infl. f-element major item major item infl. f-element infl. f-element

major item major item major item major item infl. f-element infl. f-element infl. f-element infl. f-element

upward downward upward downward upward downward upward downward

yes no no yes no no no no

("infl. f-element" = "inflectional f-element", referring to an f-element represented as a nonhead affix in morphosyntactic representation or a set of features in morphosyntactic representation indicating processual morphophonological operations.)

3. Conditions on major items and function elements 3.1. Constraints on function elements The analysis of Table 1 which I wish to present involves two sets of general conditions. The first involves substantive universals which capture generalizations about the distribution of f-elements across structural levels which have tacitly been assumed in most generative studies but which have rarely been stated explicitly. 14 These conditions are presented in (21). To be a bit more precise about terminology here, by "function word in syntactic structure" I mean a member of a nonmajor lexical category which does not participate in forming X' projections (i.e., Det, Aux, Quant, Conj, etc.). 15 By "nonhead marker in morpho-syntactic representation" I mean that the element / can be expressed in morphology through affixation to some other element h serving as the derived unit's head, or t h r o u g h a set of f e a t u r e s indicating p r o c e s s u a l morphophonological rules operating on h (cf. Anderson [1977, 1988]

158

Steven G. Lapointe

and Zwicky [1985b, 1986] for discussions of processual morphological functions). In terms of the definition of morphological head given in the introduction above, this amounts to / being an inflectional element. (21) Conditions on the Distribution of Function Elements a. F-elements are represented as operators in logical representation. b. Let / be an f-element in logical representation. In the unmarked situation, / is represented either as a function word in syntactic structure, or as a nonhead marker in morphosyntactic representation. In the marked situation, /is represented by a major lexical item in syntactic structure and either by a major item or by a derivational head in morphosyntactic representation. c. If a function word represents an f-element in syntactic structure, it must be contained in the smallest maximal phrase containing the phrase corresponding to the logical representation constituent that the f-element operates over. d. Likewise, if a nonhead marker represents an f-element in morphosyntactic representation, it can only be attached to an item in morphosyntactic representation associated with an item in syntactic structure that meets the condition on function words in Cc) above. There are basically two types of conditions included in (21). First, (21a-b) regulate the kinds of items that can represent felements at the three structural levels that we have adopted; second, (21c-d) determine where the corresponding syntactic structure item can appear depending on the logical representation structure that the f-element occurs in. It is likely that some, if not all, of the conditions included in (21) actually follow from more general principles of Universal Grammar and hence need not be stated as separate conditions of the theory. This seems especially likely in the case of (21c-d), which specify, in rough terms, that the syntactic structure item in question cannot be any higher than the phrase corresponding to the logical representation constituent over which the f-element has scope. Such conditions undoubtedly fit into a larger theory of phrasal correspondences across the logical representation/syntactic structure interface, although it is simply not possible to explore the details of those correspondences here.

The government constraint

159

Examining the condition in (21b) a little closer, we note that it claims that f-elements are typically expressed by function words in the syntax or by nonhead markers in the morphology. It is important to emphasize, however, that (21b), being a markedness condition, does not require every logical representation operator to be expressed in this fashion but rather also permits, in the marked case, such elements to be expressed as major lexical items and/or derivational heads. Thus, in the unmarked case the conditions in (21b) allow tense operators, for instance, to be expressed either as function words (as with English auxiliary did) or as morphological markers (cf. English past tense -ed), and likewise for nearly all of the other examples that Carlson (1983a) cites. 16 On the other hand, (21b) allows for a situation in which the logical representation modal operator for possibility, for example, is realized in English as either the Aux verb can (a function word) in the unmarked case or the Adv possibly (a major item in both syntactic structure and morphosyntactic representation) in the marked case. Similarly, the conditions in (21b) permit an f-element / t o be a major lexical item in syntactic structure and a derivational head in morphosyntactic representation, parallel to the case of the incorporating West Greenlandic V affixes treated by Sadock (1985). The conditions in (21b) do not, however, specifically determine which category the morphological head will belong to, nor which syntactic category the function word will belong to; thus, (21b) does not explicitly require that the element to which English past tense -ed attaches is a V or that the category that did belongs to is Aux. The conditions in (21c, d) go some way toward determining this, although they simply state in which domain the corresponding syntactic item must appear but not which items in those domains are possible corresponding items. The latter relations presumably follow from a set of separate, but interacting principles. (Cf. in this regard the extensive discussion in Carlson 1983a.) Returning now to the central theme of the paper, (21) makes some very specific predictions about how associations involving inflectional f-elements can proceed. In particular, it correctly predicts all of the information contained in lines (c) - (h) of Table 1. First, in lines (c) and (d) we have L = an inflectional f-element and Y = a major item. In upward association (c), the f-element L would have to be the head in morphosyntactic representation with the major item Y attaching to it as nonhead material; this follows from the definition of upward association. However, this situation

160

Steven G. Lapointe

directly contravenes (21b) which requires the f-element to be a nonhead item in morphosyntactic representation in this case. Because we are considering only inflectional f-elements here, the marked option of (21b) is not open to us, since an inflectional item cannot be a major lexical item or a derivational head in morphosyntactic representation, and in the unmarked option, an inflection is not represented as an element of any sort, function word or otherwise, on the Autolexical account of inflections adopted here. Hence, (21) correctly rules out this possibility. 17 On the other hand, in downward association (line [d]), the major item Y would be the head in morphosyntactic representation, and the f-element L would be the nonhead, in complete accord with (21b). Further, (21a), (21c), and (21d) combine to guarantee that the f-element in logical representation will in fact be in a structurally superior position, as it is in (11a1). Thus, the conditions in (21) correctly predict this case to be possible. Second, in lines (e) and (f) we have L = major item and Y = inflectional f-element. In the case of upward association, the major item L is the head and the f-element Y is the nonhead, in agreement with (21b). However, such an association will violate (21d). To see this, consider (11a') once again, and let L', Y' designate the elements in morphosyntactic representation corresponding to L, Y in syntactic structure, respectively. When Y' attaches to L' in morphosyntactic representation (as it must in upward association), (2Id) is violated because L lies outside the smallest maximal phrase containing Y, which in (1 la') is X". The conditions in (21) correctly predict, then, that upward association is impossible here. The case of downward association is ruled out for the same reason that upward association in line (c) is disallowed; the f-element would have to be the head by the definition of downward association in this case, but (21b) requires the f-element to be a nonhead morphosyntactic representation item. Finally, in lines (g) and (h), both L and Y are inflectional felements, and both upward and downward association results in a violation of (21b). In the case of upward association, the higher f-element must be both a head by definition of upward association and a nonhead by (21b); in the case of downward association, it is the lower f-element that must meet these conflicting conditions. In addition, the upward association case is also disallowed because it involves attachment to an item whose corresponding syntactic structure element lies outside the appropriate smallest maximal phrase, and so it violates (21 d) as well.

The government

constraint

161

We therefore see that of all the possibilities in lines (c) - (h) of Table 1, the conditions on the distribution of f-elements correctly permit only one case, downward association when L = f-element and Y = major item. There is a further set of possibilities potentially allowed by the conditions in (21), though, which fall outside the generalizations summarized in Table 1. Specifically, since (21b) is a markedness condition, it permits the marked possibility in which the f-element is not inflectional but rather corresponds to a major item in syntactic structure and a stem or head affix in morphosyntactic representation, an alternative not listed in Table 1, which only includes information about inflectional f-elements. Thus, lines (c) and (e) of the table should yield two more possibilities in which L in (c) and Y in (e) are major items corresponding to the f-element in logical representation. Since these cases involve an f-element essentially acting like a major item, we would expect that they would effectively collapse to a special subcase of line (a) in the table, and hence that upward association would be possible in these cases. However, as we will see, the constraints on major lexical items to be introduced directly will have the effect of ruling out these cases as well.

3.2. Constraints on major lexical items Given that the distribution conditions (21) handle lines (c) - (h) of Table 1, we are left with lines (a) and (b), and the marked major item cases derived from lines (c) and (e), to deal with. The second set of general conditions will handle these possibilities. The first condition is actually a consequence of Sadock's (1985) conditions Γ and IV, and so it need not be stated as a separate condition in Autolexical theory. 18 (22) Consequence of Sadock's (1985) Conditions Γ and IV: If two elements a and b are combined by morphological rule to form element c, a is the head of c, and a has a syntactic representation, then a is the lexical head of a phrase in the syntax. The second condition is one that relates the argument structure of the head of a morphological unit to the nonhead and represents a strengthening of the condition on compounds formulated in Lieber (1983).

162

Steven G. Lapointe

(23) Condition on Argument Interpretation in Morphology (generalized from Lieber 1983) Let a and b be combined by morphological rule to form cwith a as the head. Then b is interpreted as (the lexical head of) an internal argument of a if and only if b is a major lexical item. Since this condition is stated quite generally, we would expect it to perform a good deal of work for us that goes beyond accounting for lines (a) and (b) in Table 1, and indeed this seems to be the case. Thus, when a and b do not have syntactic representations, the argument condition in (23) makes a number of correct predictions. First, when a and b are both major lexical items, we have a case of compounding, and the argument condition (23) reduces to a version of the condition on compounds in Lieber (1983) and Selkirk (1982). Second, when a is a major lexical item and b is not, according to (18) b cannot be interpreted as an argument, and this seems correct. In the first place, b might be a derivational affix which for one reason or another is not the head of c but, like other derivational affixes, operates on a to change its lexical meaning. An example of this sort would be English Vs prefixed with re- irestaple, repaint, reinterpret). Alternatively, b might be an f-element type of inflectional affix, in which case it has no direct role to play in determining the meaning of the morphological unit c but instead is correlated with meaning changes at the phrasal level in logical representation; a case like this would be the English past tense ending already discussed. Finally, b might be a concord-type inflectional affix, in which case it makes no contribution to meaning at all. Gender and number markers on articles in the Romance languages would seem to be examples of this type (e.g., Spanish el läpiz 'the [masculine singular] pencil'; las ninas 'the [feminine plural] little girls'). Here it is arguable that the gender and number features marked on the article are simply a formal reflection of the categories of the head N, where the properties of the N, and not the marking on the article, makes a contribution to logical representation. In each of these cases the affix is not an argument of the head to which it is attached. Since they appear to be representative of the type where a is a major item and b is not, it seems reasonable to conclude that the argument condition (23) correctly prevents b from being an argument of a here.

The government

constraint

163

The third possibility, when a is not a major lexical item and b is, involves cases like those in West Greenlandic examined by Sadock (1985), except that a strictly lexical analysis of the combination a + b is called for rather than an Autolexical account. (23) predicts here that if a is an argument-taking element and b is an element of a category that can be (the lexical head of) an argument of a, then the result should be fine. This prediction is indeed correct; the literature contains a number of cases which are more perspicuously treated lexically/morphologically rather than autolexically. Rosen (1989) presents several cases that fall into this type. The final logical possibility, neither a nor b is a major lexical item but a is the head of the unit, falls outside of the domain of the condition in (18), since at least one of the elements must be a major item in order for the condition to be able to apply at all; in any event, this case would at best involve an inflected function word, a type of item which we earlier set aside. Now, in cases where a and b do have syntactic representations, then we derive the results in lines (a) and (b) of Table 1. There, L and Y are both major items. Let L, Y again have morphosyntactic representations L', Y'. In the upward association case line (a), L' is the head and Y' is the nonhead. Thus, L' = a and Y' = b in (23). Moreover, by (22) L must be the head of a phrase in syntactic structure. Then by (23) because Y' (= b) is a major item, it must be interpreted as (the head of) an internal argument of L' (= a). But this is exactly what we have in the structure in (11a1) in the case where Y is the head of the phrase X" which is sister to L. Hence, (23) allows the sorts of cases of incorporation that we examined earlier. On the other hand, the reason why the downward association case in line (b) is disallowed is that here the relations are reversed; Y' = a is the head in morphosyntactic representation while L' = b is the nonhead. Then by (23) L ought to be (the head of) an internal argument of Y, but the configuration in the structure in (11a') does not allow this. Thus, the argument condition in (23) accounts for the intuitive problem with (14a) (=line [b]) noted earlier, that the argument item, the Ν in the syntactic structure there, is the morphosyntactic representation head, whereas we would expect the argument-taking item, t h e V, to be t h e morphosyntactic representation head. Finally, consider the cases derived from lines (c) and (e) of Table 1 that arise from the marked option of (21b), where an f-element in logical representation is represented as a major lexical item in syntactic structure and as an independent stem or a

164

Steven G. Lapointe

derivational affix/morphological operation in morphosyntactic representation. Since this in effect leaves us with cases involving two major lexical items, we might expect, as noted above, that upward association would then be possible, parallel to the upward association allowed in line (a). However, this is not the case. In line (c), the higher element L is the f-element, and its morphosyntactic representation involves the head of the derived form. However, since f-elements are operators in logical representation (21a), they are not the sorts of predicate items that have internal or external arguments in the sense intended here. Hence, even though the f-element might operate over Xn, that phrase is not its internal argument, and so Y does not meet the condition on heads of internal arguments in (23). A similar problem arises with line (e). Here, the lower element Y is the f-element; again, this means that it is an operator in logical representation, and so it cannot serve as (the head of) an internal argument for some other predicate item. Therefore, this case also contravenes (23), and we see that even in the marked options allowed by (21b), the associations are disallowed.

3.3. Some general considerations It therefore appears that the government constraint in Autolexical analysis can be derived from the two very general sets of conditions on the Distribution of Function Elements (21) and Argument Interpretation in Morphology (23), and so it need not be stipulated as a distinct condition in Autolexical theory. Because the conditions in (21) and (23) are stated in terms that are largely independent of the differences between the Government and Binding theory and Autolexical theory (though some caution needs to be taken here because of recent Government and Binding accounts of function categories in syntax; cf. note 15), it should be possible to formulate similar conditions as part of Government and Binding which would have the effect of requiring upward movement in the case of two major lexical items. This would represent a separate way to derive the Head Movement Constraint in (3) from more general principles of the theory. Those conditions would also require downward movement in the case of an f-element in a position governing that of a major lexical item. Assuming this to be the case, an interesting claim is made about the question raised earlier concerning whether the -PASS morpheme

The government

constraint

165

moves downward or the V moves upward in a passive structure in Government and Binding. The claim is that the -PASS morpheme must be an f-element of some sort, and the movement must therefore be downward. The other possibility, that it really is a N, would leave us with a situation in which the Ν in INFL governs the V in the VP that is a sister to INFL (cf. the structure in (16) again). This leaves the -PASS morpheme as an external argument of the V (indeed, this is one of the reasons that Baker 1988 and Baker - Johnson Roberts 1989 cite for permitting the passive morpheme to be assigned the external theta role of the V), but this is in direct conflict with the argument condition in (23); if -PASS is a N, it is a major lexical item, and so by (23) it must be interpreted as an internal argument of the V which is the head of the V+PASS morphological unit. Hence, we derive theoretical motivation for saying that if an analysis like the one offered by Baker is correct in that a -PASS morpheme starts from INFL in D-structure, then it cannot be a N, it must be an f-element of some kind, and its movement must be downward. To the extent that the arguments in favor of claiming that -PASS is a nominal argument presented in Baker (1988) and Baker et al. (1989) are cogent, some other explanation for them must be sought. 19 Finally, the account just outlined is intended to subsume only those cases of incorporation and inflection that arguably involve discontinuities in the positioning of the elements that are ultimately combined morphologically. In particular, it has nothing directly to say about cliticization phenomena which, as noted in section 2 above, permit both upward and downward movement/association. If we assume that such processes involve the interactions of syntactic structure and the phonology (as suggested by Lapointe 1987) or i n t e r a c t i o n s between m o r p h o s y n t a c t i c and morphophonological levels of representation (as argued by Woodbury, this volume), then it would be reasonable to assume that the conditions regulating the movements/associations in such cases are distinct from those involved in incorporation and inflection between logical representation, syntactic structure, and morphosyntactic representation. Indeed, Woodbury's analysis attempts to do just that. I will therefore leave the issue of how best to characterize those conditions for further study.

166

Steven G. Lapointe

4. Potential counterexamples In recent work seeking to simplify and extend his original Autolexical proposals, Sadock (1991) has discussed a number of incorporation and cliticization cases from a wide range of languages, some of which can be taken as counterexamples to the generalization in lines (b) and (c) of Table 1. Here I will briefly examine two of the more interesting of those cases to show that they are not in fact counterexamples to the claims made above. 4.1. Hungarian locational suffixes The first potential counterexample involves locational suffixes in Hungarian and their relations to free-standing locational postpositions. As Sadock notes, there are a number of parallels between the two classes of items. First, the suffixes only appear in one position in a NP, namely, attached to the head N; they are unlike regular case endings in this regard, since the latter induce concord on a Det and Adjs in the NP, whereas the locational suffixes do not. In this way the locational suffixes are more like regular postpositions, since these also can only appear once with a single NP. Sadock offers examples like those in (24). ( 2 4 ) a.

b.

c.

(25) a.

egy

feher

egy

feher

egy

feher

ez

a

a white 'a white house'

häz

house häz-rol

a white house-rol 'from a white house' häz

a white house 'beside a white house' häz

this the house 'this house' b. er-röl a häz-rol this-rol the house-rol 'from this house'

mellett

beside

The government constraint

167

An exception to this fact that Sadock observes is that in the demonstrative construction, which consists of a demonstrative followed by a definite NP, a locational suffix must appear both on the demonstrative and on the head Ν of the NP (25). However, regular postpositions behave in exactly the same way in demonstratives (26). (26) e mellett a this beside the 'beside this house'

häz house

mellett beside

Second, both locational suffixes and postpositions can contain two morphemes, which allows them to form several separate morphemes with related meanings from the same base. (27) a. -bAn 'in'; - M 'into'; -bÖl 'out of b. mellett 'beside'; melle 'to beside'; mellöl 'from beside' Finally, when the NP is a pronominal, the pronoun occurs as a suffix on the postposition, as shown in (28). Interestingly, the same thing happens with the suffixes; these occur as stems, with their own inherent vowel quality, and with the pronominal suffixes attached to them (29). (28) a. mellettem 'beside me' b. melletted 'beside you' c. mellette 'beside him'

(29) a. b. c.

bennem 'in me' benned 'in you' benne 'in him'

Sadock's analysis takes both the regular postpositions and the suffixes to be adpositions of category Ρ in syntactic structure. The difference between these two kinds of elements is that the regular postpositions have default morphology, which means that they have simple stems in morphosyntactic representation, while the locational suffixes are optionally specified as being affixes which attach to stems in morphosyntactic representation. When these elements are not affixes, they also receive default morphology and act as independent stems in morphosyntactic representation so that the pronominal forms can combine with them in the appropriate way to create forms like (29). Finally, the lexical entries for the pronominal forms indicate that they are NPs in syntactic structure but are affixes that attach to Ps to form words in morphosyntactic

Steven G. Lapointe

168

representation. The lexical information that Sadock gives for some of the example forms used above are listed in (30).20 ( 3 0 ) a.

mellett:

Syntactic structure - [N", ] Morphosyntactic representation - default b.

ben:

Syntactic structure - [N", Morphosyntactic representation - ([X°, c.

] ])

Am:

Syntactic structure - N"[l, SG] Morphosyntactic representation - [ w Ρ ] It might appear that this analysis poses problems for the account of upward and downward association presented above, since it seems to involve a major lexical item, the postposition, moving downward and attaching to the head Ν of the NP, contrary to line (b) of Table 1. However, the Autolexical analysis is more complicated than that, and closer inspection reveals that neither of the possible ways of viewing the attachments impugn the results of the previous sections. The difficulties here hinge on the categorial status of P, the category of the postpositions. If Ρ is a major lexical item, as the corresponding prepositional category apparently is in English (Jackendoff 1973), then we would have a syntactic structure, morphosyntactic representation, and associations like those in (31) for the NP in (24b). Such an analysis would be exactly parallel to that given in (9) for cases of Ν incorporation. The postposition would be in a syntactically superior position with respect to the NP, the head Ν of the NP would move upward to the Ρ position, (31) a. Syntactic structure for (24b), Ρ a major lexical item PP NP ρ Det I AP Ν egy

,

feher

I häz

-rol

The government constraint

169

b. Morphosyntactic representation Det A Ρ I I Ν Ρ ι egy feher ι haz -rol c. Associations Syntactic structure Det

A

Morphosyntactic representation

^

Ν

the Ρ would be the head of the morphosyntactic representation for häzrol in (31b), and the Ν would be (the head of) an internal argument of the P-all in accordance with the conditions stated in the preceding section. On the other hand, if Ρ is an f-element category, then the syntactic structure would look like (32b), assuming that there is an ID rule like that in (32a) that introduces this category in an adjoined structure. The morphosyntactic representation and associations would then be as in (32c) and (32d). Here, Ρ as an f-element associates downward, Ν is the stem and Ρ is the affix that attaches to it, and again, all of this accords with the earlier conditions. (32) a. NP[LOCAT] — NP P[LOCAT] b. Syntactic structure for (24b), Ρ an f-element NP [LOCAT] NP~ Det ' egy

^ P N'

AP

[LOCAT] ν ι

-rol

feher haz c. Morphosyntactic representation Det A Ν ' egy

' feher

Ν \ häz

Ρ ι -rol

Steven G. Lapointe

170

d. Associations Syntactic structure Morphosyntactic representation

Det

A

I

' A

Uet

Ν Ν

Therefore, it does not matter which of the two possibilities for the categorial status of the adposition Ρ here turns out to be correct; whether it belongs to a major lexical category or a functional category, there is a straightforward account of the associations which is consistent with the previous results. However, there is reason to believe that the facts presented above are to be handled by a purely lexical rather than an Autolexical account, an account which would render those facts irrelevant as far as the issue of upward/downward association is concerned. The alternative account takes the locational suffixes literally as a kind of inflectional ending which is required by the rules of Hungarian morphology to appear outside of the regular case endings of Ns.21 To ensure that the features associated with these elements do not induce concord, all we need to do within the GPSG-based syntax adopted in Autolexical analysis is to say that those features cannot serve as the value for the feature AGR. Indeed, it may not be necessary to specify for the grammar of Hungarian that this is the case if it is a general property of such "outer" inflections that they do not participate in agreement or morphological government; if that proves correct, we can simply stipulate in Universal Grammar that the features of "outer" inflections behave in this way. Finally, t h e suffix forms have separate entries from the forms that can o c c u r with p r o n o u n s ; t h e f o r m e r h a v e m o r p h o s y n t a c t i c representation information in their lexical entries indicating that they are affixes (33a), while the latter have entries indicating that they are stems which must be bound to pronominal forms (33b). -bArr. Syntactic structure - 0 Morphosyntactic representation - [X°, ] b. ben< Syntactic structure - N"[LOCAT] Morphosyntactic representation - [ N°[PRON]]

(33) a.

The entries for the pronouns can now be given as in (34); i.e., these items can be taken as simple N°s in the syntax, and either as

The government constraint

171

bound stems that must be attached to a Ρ or as simple stems in the morphology. 22 (34)

Am: Syntactic structure - N°[l, SG] Morphosyntactic representation - ([P

])

If we take the regular postpositions to have entries like that for mellett in (30a), and if we assume that the category Ρ is functional, and hence that a syntactic structure like (32b) is correct for these elements, we will have an account of these locational forms which predicts the same facts as the Autolexical one while handling certain additional facts which are a problem for the approach sketched by Sadock. The account just offered explicitly handles the fact that the locational suffixes are like the regular postpositions in not triggering agreement. Furthermore, it allows the pronominal forms to attach to both types of locational elements as desired, since the entry in (34) says that, on one option, the pronouns must attach to an item of category Ρ in morphosyntactic representation, without regard to the P's status as a separate word or a stem that must be bound. The present account does not say anything directly about t h e fact t h a t suffixes and p o s t p o s i t i o n s b o t h can have morphologically complex internal structure (27), but Sadock's analysis also does not directly accommodate these forms, since no explicit proposal is offered there that tells exactly how these forms are to be generated. Similar comments hold for the facts about demonstratives in (25) and (26).23 In addition, there are two problems which the present account solves. First, the inflectional account offers a simpler view of what is going on with the vowels in the stems that combine with pronouns (cf. note 20). The bound stem forms have separate lexical entries, and the inherent vowels of those stems are listed in their entries just as inherent vowels are listed in the entries for other stems in the language. It is not necessary on this account to refer to phonological rule overrides to account for the a p p a r e n t l y idiosyncratic nature of the vowel in the bound stem forms. Second, there is the question of why, in the account Sadock presents, if the suffix forms have entries like (30b), they cannot stand on their own as independent Ps in syntactic structures, since as Sadock himself points out, these Ps have the status of affixes only optionally. Sadock suggests two possibilities for overcoming this problem, neither of which is particularly inviting. The first is that there is a

172

Steven G. Lapointe

language-particular statement in the grammar of Hungarian that the suffixes must be treated as suffixes "except under duress"; the second would be to use a novel form of the Elsewhere Condition which involves information about representations and rules from several components at once. The present account avoids these alternatives entirely by recognizing that the suffixes have the entries in (33a) and (33b); in the first, they are affixes, in the second they are bound stems, and in neither type of entry are they allowed to be free-standing words.24 Finally, it would appear that the Autolexical approach offered by Sadock accounts in a more natural way for one fact than the purely lexical account presented here does. Sadock's proposal makes the connection between the locational suffixes (e.g., -bAri) and their corresponding pronominal forms ibennem, etc.) quite transparent, since on his account the locational element in the two cases is in fact the same item. On the present account, though, the suffixes and the pronominal stems have separate entries, and so the fact that there are any relations between them at all (for instance, that they describe exactly the same locational notions, and that they have largely the same phonological properties) would seem to be utterly accidental. There are two responses that need to be made here. First, there are several ways that we can arrange the lexicon of Hungarian to capture the generalizations about the suffixal versus pronominal items. One would be to say that there is a redundancy rule which connects entries like (33a) with those like (33b); such a rule would have to say that for any suffixal Ρ (33a) there is another entry for a form which must be bound to a pronoun (33b). An alternative would be to say that in a case like (33a-b) we do not really have two separate entries; instead, we have a single entry with constant semantics (in this case, meaning 'in') but with subentries distinguishing the suffix versus the pronominal form. The second response is to call into question whether it is appropriate to claim that the two forms should be taken as identical with the exception of a single morphological property in the way that Sadock does. While forms like -bAn and ben- are undoubtedly related to one another, synchronically as well as historically, it is unclear that we want to say that for each such pair the relation is one of formal identity. Thus, there are suffix/pronominal stem pairs like -bol/belöl- 'out of and -bA/bel- 'into' where there are extra consonants (and in the first case, a vowel) whose presence would appear to be a peculiarity of the individual stem.

The government

constraint

173

Furthermore, one form, -An/rajt- 'on' has an entirely suppletive pronominal stem form. It is not obvious how such idiosyncracies of form are to be stated in an account which takes the only difference between the suffix and pronominal stem forms to be an optional subcategorization in morphosyntactic representation.25 On the other hand, the alternatives suggested in the last paragraph allow for the separation of the properties of the two kinds of forms in such a way that these idiosyncrasies can be easily accommodated. In terms of the approach which takes the elements to have a single entry but distinct subentries, for example, a regular suffix/pronominal stem pair like -bAn/ben- would have an entry that would include the basic phonological and semantic information for the pair of forms, while the subentries would include information about the syntactic structure and morphosyntactic representation subcategorizations and the properties of the vowels, as shown in (35). (35) Lexical entry with subentries for -bAn/benPhononolgy -



C

V C

b η Logical representation - OP . . . Semantics - "in" Syntactic structure - 0 Morphosyntactic representation - [X°, ] Phonology - V = e Syntactic structure - N"[LOCAT] Morphosyntactic representation - [ N°[PRON]]

Here, the first line of the entry says that the CV-skeleton consists of a CVC, where the first consonant is b, the second is n, and the vowel is underspecified; the second line says that the item is an operator in logical representation, with further information specifying how it combines with other logical representation elements that has been omitted here; and the third line says that the semantic content of the operator is "in". Subentry contains the information about the suffix form; without any further information about the melody of the vowel in the form, other processes, in particular the vowel harmony rules of the language, will supply the missing feature specifications. The rest of subentry contains the information about syntactic structure and

174

Steven G. Lapointe

morphosyntactic representation that appears in (33a) above. Subentry indicates that, for this form, the vowel must be e, while the remaining information about syntactic structure and morphosyntactic representation is the same as that for (33b). In the case of a suffix/pronominal stem pair with idiosyncratic properties, like -An/rajt- we would have an entry like that in (36). Here, only the logical representation and semantic information about the forms are listed in the entry proper. The subentries contain not only information about the syntactic structure and morphosyntactic representation properties of the suffix and the pronominal stem but also information about the idiosyncratic phonological forms as well. (36) Lexical entry with subentries for -An/rajt-

Logical representation - OP . . . Semantics - "on" Phonology V C I

η

Syntactic structure - 0 Morphosyntactic representation - [X°, PhonologyC V C C Syntactic structure - N"[LOCAT] Morphosyntactic representation - [

]

N°[PRON]]

We thus see that there are a number of facts which suggest that a purely lexical account of the properties of the suffixal locational items in Hungarian is preferable to an Autolexical approach like the one offered by Sadock (1991). If such an analysis can be supported by further data, these sorts of forms will pose no difficulty for the account of directionality of movement/association formulated in the preceding section. 4.2. Finnish negative verbs The second potential counterexample involves negative Vs in complement clauses in Finnish, a set of facts which Sadock (1991) also examines. In Finnish, the negative is expressed by a distinct

The government constraint

175

V en which inflects for person and number just as other Vs do in the language; the V describing the event in the sentence then appears in a nonfinite complement form. (37) a. b. c. (38) a. b. c.

Hyppaan. jump, lsg Ί jump.' Hyppaatte. jump, 2pl 'You jump.' Kell ο on neljä. o'clock be, 3sg four 'It is four o'clock.' En hyppaa. NEG, lsg jump Ί do not jump.' Ette hyppaa. NEG, 2pl jump 'You do not jump.' Kello ei ole o'clock NEG, 3sg be 'It is not four o'clock.'

neljä. four

Interestingly, some complementizers, like effä'that', can optionally merge with a negative V in its clause. As Sadock observes, Karlsson (1983: 191) presents cases like the following in which the negative V is separate from (39a) or attached to (39b) the COMP element. (39) a.

Väitätkö, että kello ei ole neljä? 'Are you claiming that it is not four o'clock?' b. Väitätkö, ettei kello ole neljä? = (39a)

Sadock suggests that cases like (39b) be analyzed with the V incorporated into the COMP. However, such an analysis would run counter to the generalizations discussed earlier, since we would have a major lexical item moving upward to incorporate with an f-element, contrary to line (c) of Table 1. Hence, such an analysis presents potentially serious problems for the account given above. There is, however, an equally plausible approach that we can take to these facts that does not contravene the analysis of government constraints given above, namely, to claim that what is

176

Steven G. Lapointe

going on here is that the COMP and the negative V are cliticizing rather than incorporating. It turns out that there is an independent syntactic process in Finnish whereby yes/no questions are formed by fronting a focused element into initial position and adding the question suffix -ko; in nonfocused yes/no questions, the main V fronts. Thus, Karlsson (1983: 68) offers pairs like that in (40). (40) a. Pekka Pekka 'Pekka b. Saapui-ko arrived-Q 'Did Pekka

saapui Turkuun aamulla. arrived at Turku in the morning arrived at Turku in the morning.' Pekka Turkuun aamulla. Pekka at Turku in the morning arrive at Turku in the morning?'

The same fronting and -ko attachment occurs in embedded yes/no questions (examples again from Karlsson 1983: 70). (41) a. En tiedä, meni-kö Auli kotiin. NEG, lsg know go, 3sg-Q Auli home Ί don't know (whether) Auli went home.' (lit., Ί don't know, did Auli go home.') b. Kysy, onheillä Iämmintä ruokaa. ask be-Q at their place warm food 'Ask (if) they have warm food.' (lit., 'Ask, is there warm food at their place.') If this process has been generalized optionally to the negative V en when it occurs in embedded, affirmative clauses, then that item will appear immediately after the COMP että and hence will be in the correct position for it to be able to cliticize with the negative V. In that case, we need not claim that the V has been incorporated into the COMP, but rather that it has cliticized to it. As noted at the end of section 3, nothing said there rules out the possibility that a major lexical item might cliticize to an f-element; the conditions presented earlier were intended to handle certain cases involving incorporation and inflection, but not cliticization. Hence, if there are independent means for assuring that a morphologically combined major lexical item and f-element are adjacent in syntactic structure, then a clitic analysis, rather than an incorporation analysis, is open to us. Unfortunately, it is difficult to find compelling independent evidence for such an account on the basis of what Sadock (1991)

The government constraint

177

and Karlsson C1983) say. One piece of evidence that potentially could be cited involves the nature of the phonological merger of the two forms. As can be seen in the COMP + NEG V form in (39b), the COMP element loses its final vowel when the vowel initial negative V is attached. Indeed, this might even be taken at first glance as evidence that it is the COMP that a t t a c h e s to the V, rather than the other way around. However, one could counter that there are two facts about Finnish that suggest that this is not the case. First, the main morphological operation in the language involves suffixation, suggesting that the COMP is the base to which the negative V is added. Second, in inflectional affixation at least, short ä is dropped before certain vowel-initial suffixes; t h u s we have forms like vetä + i— veti 'pulled'; kylä + im-* kyliin 'into the villages'; syvä + in^syvin 'deepest' (Karlsson 1983: 42). But o n e can offer t h e following r e j o i n d e r t o t h e s e counterarguments. In the cases where ä drops, the following suffix begins only with i; since the negative V begins with e, it is not clear that t h e s a m e phonological p r o c e s s is at work here. Furthermore, while suffixation is far and away the most important morphological operation in the language, it is not the only one. Indeed, compounding is also quite productive, especially with Ns (cf. Karlsson 1983: 203), and since we are suggesting here that että + en involves the joining of two whole words, it is not unreasonable to view t h e morphological p r o c e s s that actually combines the forms as an instance of compounding. We thus see that the evidence in favor of the clitic analysis appears at the moment to be rather equivocal. Nonetheless, this analysis is in no w o r s e s h a p e than the incorporation account proposed for these forms by Sadock, since he offers no evidence for or against that analysis in his discussion. But this state of affairs is sufficient to establish the main point that needs to be made concerning these Finnish forms. While it may be possible to offer an incorporation account here which conflicts with t h e proposals about upward and downward association in section 3, it is by no means obvious that that account is correct. Indeed, as we have seen, there is an equally plausible cliticization account that can be given in this case which is completely consistent with the account given here for the directionality of association. To the extent that further work can show that t h e r e is evidence in favor of this latter analysis, the Finnish facts will not count as real counterexamples to this treatment of directionality.

178

Steven G. Lapointe

5. Conclusion In this work, we have reviewed the government constraints which restrict the movements of lexical items in Baker's (1988) Government and Binding analysis and the associations of items across structural representations in Sadock's (1985) Autolexical approach and found both accounts wanting. While Baker appropriately tries to derive the Head Movement Constraint (3) from more basic principles of the Barriers version of Government and Binding, we have seen that he in fact does not prove this result. Furthermore, while the Head Movement Constraint does a c c o u n t nicely for certain properties of Ν incorporation constructions, it leaves unspecified the question of exactly how movements involving inflectional elements are to be constrained. Likewise, Sadock's (1985) government condition VH' is also simply stipulated rather than being shown to be derivable from other principles of the Autolexical approach, and it, too, runs into problems concerning possible associations involving f-elements. To overcome these problems, two general sets of conditions, the Conditions on the Distribution of F-Elements (21) and the Condition on Argument Interpretation in Morphology (23), which are going to be required in grammatical theory in some form in any event, have been shown to imply the generalizations about possible upward/downward associations in Autolexical theory, and it has been suggested, although not explicitly demonstrated, that comparable conditions within Government and Binding will lead to similar results concerning the upward/downward movements of lexical items. We have also seen that several kinds of potential counterexamples to the generalizations summarized in lines (b) and (c) of Table 1 involve analyses that may well be incorrect. It is to be hoped that this last conclusion can be maintained by future research and that the general results of this paper can find further support through the investigation of a wider range of facts involving the interactions of major lexical items and f-elements.

Notes *

I wish to thank Jerry Sadock, Mark Baker, and Eric Schiller for their helpful c o m m e n t s on earlier versions of this paper. I would also like to thank Donka Farkas and Jeff Harlig for their help with various

The government

1. 2.

3.

4.

constraint

179

aspects of the Hungarian data discussed here. Remaining errors of fact or theory are, however, solely my own. For extensive discussion concerning the difficulties in specifying the properties of heads in both syntax and morphology, see Zwicky (1985a). I wish to thank Jerry Sadock for reminding me of the problems with the proofs that Baker presents. That the Head Movement Constraint and the Empty Category Principle are not completely equivalent is an issue that Chomsky (1991) returns to several times in his discussion. Pullum (1988) has noted two further problems with Baker's (1985) account of gerunds. The first has to do with the category features for the -ing affix; if they are simply [+N, -V], then INFL would have to be specified in the same way, but that would predict that Ss (=INFL") should in general be NPs, which they are not. Assuming a separate feature [±F] to distinguish the category specification of the affix versus that of INFL leads to further complications, as Pullum argues. The second difficulty involves the bar levels of the structure resulting from the syntactic affixation rule. In the version of X' theory that Baker is adopting, S = INFL" dominates the subject NP and INFL'; the latter node dominates INFL and VP. If the affix moves down from INFL and attaches to V, assuming some way around the problem just noted about the category features for -ing, the VP will recategorize as a NP, but nothing will force the INFL' or INFL" nodes to recategorize. Yet it is exactly the recategorization of those nodes that would turn an S into an NP as required on Baker's analysis. These technical problems, along with the more general ones pointed out in the text, suggest that the sort of approach to gerunds proposed by Baker (1985) is not tenable. Notice that I am focusing here on the movement of the f-element, not on other elements that the f-element governs and which may be analyzed as moving upward to attach to it. A number of recent Government and Binding analyses have treated movements involving f-elements by assuming that the major lexical item which serves as the morphological host for the f-element moves up and attaches to that element; cf., for instance, the analyses of the English auxiliary and V inflection systems proposed by Chomsky (1986), Pollock (1989), and Chomsky (1991). It is not clear that such analyses can in general be made to work. Thus, Radford (1988) points out that the Vmovement analysis of English TNS elements in Chomsky (1986) will derive the wrong surface constituent structures in many cases, although Pollock's (1989) account probably does not suffer from these same difficulties. Similarly, Baker (1988) himself notes a parallel problem with surface constituency that would arise if V-movement literally took place in English passives, forcing him to say that an

180

5. 6.

7. 8.

9.

Steven G. Lapointe "abstract" movement occurs in such cases, in which INFL and the V position are coindexed as if the V had moved to INFL and left a trace behind, but without any actual movement taking place. Although Baker employs such "abstract" movements at a number of key points in his analyses of Ν incorporation and grammatical relation changes, the exact status of such an operation remains quite open within the Government and Binding framework. Regardless of the problems with these sorts of analyses, the point I am making in the text is not that the recent types of Government and Binding analyses, involving major lexical items moving upward to attach to f-elements that govern them, are to be ruled out a priori, but rather that f-elements themselves do not move upward. Nevertheless, when I take up the Autolexical analysis in the next section, I will suggest that the stronger claim-that the movement of major items upward to attach to felements is prohibited-is in fact correct. For brief discussion of the internal structure of NP in Latin, see Lapointe (1985). Chomsky (1991) makes a similar suggestion: traces of f-elements that have no effects in LF are deleted, in the process creating structures which no longer violate the Empty Category Principle. He employs this proposal to delete the trace of Agr, since that element has no impact on LF, which, along with (d) in the text, then permits him to account for various grammatical and ungrammatical possibilities under his "least effort" approach. As with Pollock's analysis, a full evaluation of Chomsky's suggestions would extend well beyond the scope of this paper. However, it is important to point out that this approach to deletable material is potentially quite hazardous, since in principle it permits us to posit an abstract f-element with no LF effects heading an abstract phrase in the syntax and move it to a major lexical item nearby, with subsequent deletion of its trace, anytime we have a structure where we need X°-movement to avoid violating the Empty Category Principle. It is difficult to see how, in terms of current Government and Binding approaches, one could constrain such derivations so as to disallow this unwanted possibility. See Lapointe (1989) for one such account. Here, TNS = tense, PRED = predicate, R = relation, TP = term phrase, Τ = term, and LSP = logical specifier. Note also that the associations represented in (11c) are intended to show that the line between the TNS item in logical representation and the V in morphosyntactic representation lies on a different association plane from that of the lines connecting Τ and Ν and LSP and Det. See Lapointe (1987, 1991/1988) for details. In more recent work, Sadock (1991) has recast the earlier condition in (10) as the Constructional Integrity Condition. This condition

The government

constraint

181

amounts to the one given in (10) above for case (11a), i.e., where Xn contains Y°. Sadock states the Constructional Integrity Condition in terms of a strong and a weak form, depending on whether Y° is the head of Xn, as in (10), or is just some element of Xn, as part of his attempt to bring cliticization and incorporation under the same set of principles. He also notes "an observational law to be explained by further research" concerning the direction in which incorporation applies, which he states as in (i) (1989: 4-28). (i) The syntactic position of a complex morphological expression is the same as the syntactic position of its morphological head.

10.

11.

12.

13. 14.

Unfortunately, this does not help, even as an observational generalization, since nothing in (i) specifies which of the two elements L or Y° in the morphology in (11a) is to be taken as the head in various cases. Hence, Y° could be the head, in which case we would still expect the nonoccurring downward attachment of major items; likewise, if L is the head, we would expect the nonexistent upward attachment of f-elements. One possibility might be to say that the syntactic constituent derived by V-movement is more complex than that given in (15a), exactly because a major item is attaching to an f-element. The idea would be that when this occurs both the major item and the f-element contribute their features to the higher node, so that instead of (15a) we would have something like [V+XNS V + TNS], and by feature percolation, the T' and TP nodes above the original TNS element would become (V+TNS)' and (V+TNS)P. We would then still have a TNSP in the appropriate place in the structure, containing an appropriate TNS element for purposes of LF, while at the same time we would have a structure that would pose fewer potential problems for the morphology. Logically, there should be a further set of eight possibilities derived by varying whether L or Y is the head of the morphosyntactic representation constituent. However, since this property determines direction of association in the Autolexical analysis, this dimension of variation is already covertly included in the cases discussed in the text. Notice that I am concerned here only with elements which show clear effects in the logical structure of a sentence. Agreement markers may well register on the V nonlogical features of the syntactic phrase corresponding to LP, e.g., gender class features in Bantu languages, However, such features would not be represented by FNs in logical representation. Thanks to Mark Baker for pointing out the Vata examples to me. An exception is the discussion in Carlson (1983a). The condition stated in (21) below can be seen as an extension of the principle

182

15.

16.

17.

18.

Steven G. Lapointe Carlson presents to relate semantic and syntactic operations, in terms of a somewhat different theory of syntax from the one he adopts. Recent work within Government and Binding (Fukui 1986; Speas 1986) has proposed that this is in fact not the way to look at f-elements. Rather, on Fukui and Speas' account both f-elements and major lexical items form X' projections, the difference between the two being that only f-elements form maximal phrase projections while major lexical items do not. The notion of f-element that I am adopting here is more like the position in GPSG (Gazdar et al. 1985). It may be that there are further markedness relations among the possibilities mentioned in (21b) beyond the ones explicitly given there. Thus, it seems plausible that the option in which f-elements are expressed as function words in syntactic structure is less marked than the one in which they are expressed as inflections in morphosyntactic representation. This assumption might go some way toward explaining the lack of inflections typically observed in Creole languages, as well as the delayed acquisition by children of inflections as opposed to function words in the same semantic domain. However, this assumption might actually be part of a larger generalization to the effect that it is less marked for logical representation elements in general to be expressed as full words (whether as major items or function words) and more marked for them to be expressed as morphological affixes or processual operations. The validity of these various speculative suggestions will have to await further research. Hence, for purposes of this and the following demonstrations, (21b) is acting as if it were an absolute condition rather than a set of markedness conditions. Generally, this will not be the case, the behavior of the condition here resulting from the further constraints imposed by the restriction in the discussion on inflections. The conditions Γ and IV are given below (Sadock 1985: 404). Γ.

The Head Feature Convention applies in both morphology and syntax: the features of a constituent must be the same as those of the head of that constituent.

IV. Any lexeme which is independently represented in both the syntax and the morphology is immediately dominated by the same complex symbol in both syntactic and morphological trees. As Sadock notes, IV guarantees that features distributed on nodes in the syntax will occur on the corresponding nodes in the morphology. Hence, features that determine inflectional properties of the lexical head of a phrase in the syntax will appear on the

The government

19.

20.

21.

22. 23.

constraint

183

appropriate lexical category node there by Γ, by IV those features will cross over the syntax/morphology boundary and appear on the node immediately dominating the element corresponding to the syntactic lexical head, and by the definition of morphological head given in the introduction, the presence of those features on that node will mean that the element is the head of its unit in the morphology. Hence, we derive (22) from these two conditions under the definition of morphological head that we have been employing here. The present account also predicts that Baker's (1985) analysis of gerunds cannot be correct. If -ing really is a N, the movement should be upward, rather than downward as he has it in his analysis; however, upward movement in the case of gerunds will run into problems with the argument condition in (23) of a slightly different sort than those encountered by the Baker (1988) analysis of -PASS does. In particular, the -ing affix is not intended to be an argument of any sort on this account; hence, it is not an internal argument of the V to which it is attached, a situation ruled out by (23). This is a favorable result in light of the problems with this account noted above in the text and in note 3. I have regularized some of Sadock's notation here and made it consistent with some of the assumptions made above. Thus, Sadock uses N[2] rather than N" in the syntactic structure information for mellett, and I have employed the labels "syntactic structure" and "morphosyntactic representation". It should be noted that Sadock assumes that suffixes like ben in (30b) are listed with the full vowels that they occur with in the free-standing pronominal forms and that in the cases where they appear as suffixes the general phonological rules for vowel harmony will override the lexical vowel specification. The vowels in these forms are therefore to be taken as default values which occur only when other rules do not specify distinct vowels here. I will return to this issue below. One would need more facts about Hungarian than I presently have at my command to decide how best to carry this proposal out in detail. In particular, it is not entirely clear to me whether such "outer" inflections are to be generated in terms of the sorts of inflectional templates that Zwicky (1986) has argued for, or whether such templates only govern the distributions of the more usual "inner" inflections, with the "outer" ones being generated by more typical affixational rules. A question that immediately arises concerns the relations between entries like (33a) and those like (33b). I will return to this matter below. Although an analysis in terms of a coordinate structure like that in (i) certainly suggests itself for these cases.

184

Steven G. Lapointe _ NP "ccCASE ßLOCAT DEM

(i)

NP aCASE Ί ßLOCAT _ DEM

J

erröl

NP [aCASE Ί ßLOCAT I a häzröl

Positing such a structure would of course be compatible with either approach to the status of the locational suffixes. 24. A further fact about the suffix versus postpositional forms is a problem which both accounts will have to deal with. As Sadock notes, the suffix must be attached to the head Ν of the locational NP, as in the example (i) which he cites. (i) a szep kep-rol amit tegnap lattunk the beautiful painting-rol which, acc yesterday see, past, lpl 'about the beautiful painting that we saw yesterday' However, the regular locational postpositions show exactly the same behavior; they must occur immediately after the head Ν and cannot be separated from it by a modifier like a relative clause. (ii) a. a szep kep mellett amit tegnap the beautiful painting beside which, acc yesterday lättunk. see, past, lpl 'beside the beautiful painting that we saw yesterday' b. *a szep kep amit tegnap lattunk the beautiful painting which, acc yesterday see, past, lpl mellett beside

25.

Neither Sadock's account nor the purely lexical one sketched here handles this fact, since both assume that the postpositions are attached to the right ends of locational NPs as in (30a) and hence that they should appear to the right of relative clauses like amit tegnap lattunk, which they cannot. Thanks to Donka Farkas for helpful discussion about these sentences. Thanks to Jeff Harlig for pointing these forms out to me.

The government

constraint

185

References Anderson, Stephen 1977 "On the formal description of inflection", in: W. Beach — S. Fox - S. Philosoph (eds.), Papers from the Thirteenth Regional Meeting of the Chicago Linguistic Society. Chicago: Chicago Linguistic Society, 15-44. 1988 "Morphological theory", in: Frederick Newmeyer (ed.), Linguistics: The Cambridge survey I: Linguistic theory: Foundations. London: Cambridge University Press, 146-191. Baker, Mark 1985 "Syntactic affixation and English gerunds," in J. Goldberg - S. MacKaye - M. Wescoat (eds.), Proceedings of the West Coast Conference on Formal Linguistics 4. Stanford, CA: Stanford University, 1-11. 1988 Incorporation: A theory of grammatical function change. Chicago: University of Chicago Press. Baker, Mark - Kyle Johnson - Ian Roberts 1989 "Passive arguments raised", Linguistic Inquiry 20: 219-251. Burzio, Luigi 1981 Intransitive verbs and Italian auxiliaries. [Unpublished Ph.D. dissertation, MIT.] Carlson, Greg 1983a "Marking constituents" in: Frank Heny - Barry Richards (eds.), Linguistic categories: Auxiliaries and related puzzles. Vol. 1. Dordrecht: D. Reidel. 1983b "Logical form: Types of evidence", Linguistics and Philosophy 6: 295-317. Chomsky, Noam 1981 Lectures on government and binding. Dordrecht: Foris. 1986 Barriers. Cambridge, MA: MIT Press. 1991 "Some notes on economy of derivation and representation", in: Robert Freidin (ed.), Principles and parameters in comparative grammar. Cambridge, MA: MIT Press. Chomsky, Noam - Morris Halle 1968 The sound pattern of English. New York: Harper and Row. Fukui, Naoki 1986 A theory of category projection and its applications. [Unpublished Ph.D. dissertation, MIT.] Gazdar, Gerald - Ewan Klein - Geoffrey Pullum - Ivan Sag 1985 Generalized phrase structure grammar. Cambridge, MA: Harvard University Press. Gazdar, Gerald - Geoffrey Pullum - Ivan Sag 1982 "Auxiliaries and related phenomena in a restrictive theory of grammar", Language 58: 591-638.

186

Steven G. Lapointe

Huck, Geoffrey J . - Almerindo E. Ojeda (eds.) 1987 Syntax and Semantics, vol. 20: Discontinuous constituency. San Diego: Academic Press. Jackendoff, Ray 1973 "The base rules for prepositional phrases", in: Stephen Anderson - Paul Kiparsky (eds.), A Festschrift for Morris Halle. New York: Holt, Rinehart and Winston, 345-356. Karlsson, Fred 1983 Finnish grammar. Porvoo: Werner Söderström Osakeyhtiö. Koopman, Hilda 1984 The syntax of verbs. Dordrecht: Foris. Lapointe, Steven 1985 A theory of grammatical agreement New York: Garland. 1987 "Some extensions of the Autolexical approach to structural mismatches", in: Geoffrey J. Huck — Almerindo E. Ojeda (eds.), 152-183. 1989 The place of grammatical function change in a theory of Autolexical analysis. [Unpublished manuscript., Wayne State University.] 1991/1988 "Constraints on Autolexical analyses", Linguistic Analysis 18:123-155. Lieber, Rochelle 1983 "Argument linking and compounds in English", Linguistic Inquiry 14: 251-285. May, Robert 1985 Logical form: Its structure and derivation. Cambridge, MA: MIT Press. Perlmutter, David 1978 "Impersonal passives and the unaccusative hypothesis", in: J. Jaeger et al. (eds.), Proceedings of the Berkeley Linguistic Society BLS 4, 157-189. Pollock, Jean-Yves 1989 "Verb movement, universal grammar, and the structure of IP", Linguistic Inquiry 20: 365-424. Pullum, Geoffrey 1988 Heterocategorial heads in gerund constructions. [Paper presented at the annual meeting of the Linguistic Society of America. Revised version: "English nominal gerund phrases as noun phrases with verb-phrase heads", Linguistics 29(1991): 763-800.] Radford, Andrew 1988 Transformational grammar. London: Cambridge University Press. Rosen, Sara 1989 "Two types of noun incorporation", Language 65: 294-317.

The government

constraint

187

Sadock, Jerrold 1985 "Autolexical syntax: A proposal for the treatment of noun incorporation and similar phenomena", Natural Language and Linguistic

Theory 3: 379-439.

1987

"Discontinuity in Autolexical and autosemantic syntax", in: Geoffrey J. Huck - Almerindo E. Ojeda (eds.), 283-301.

1991

Autolexical syntax: A theory of parallel grammatical representations. Chicago: University of Chicago Press.

Selkirk, Elisabeth O. 1982 The syntax of words. Cambridge, MA: MIT Press. Smessaert, Hans 1988 An Auto-lexical syntax approach to pronominal cliticization in West Flemish. [Unpublished M.A. paper, University of Chicago.] Speas, Margaret 1986 Adjunctions and projections in syntax. [Unpublished Ph.D. dissertation, MIT.] Travis, Lisa 1984 P a r a m e t e r s and effects of word order variation. [Unpublished Ph.D. dissertation, MIT.] Woodbury, Anthony 1995 "On restricting the role of morphology in Autolexical Syntax", in this volume Woodbury, Hanni 1975 Noun incorporation in Onondaga. [Unpublished Ph.D. dissertation, Yale University.] Zwicky, Arnold 1985a "Heads", Journal of Linguistics 21: 1-30. 1985b "How to describe inflection", Proceedings of the Berkeley Linguistic Society BLS 11: 372-386.

1986

"The general case: Proceedings

314.

Basic form versus default form",

of the Berkeley Linguistic

Society BLS 12: 305-

Some pleasures and pitfalls of Autolexical Syntax Jerrold M. Sadock

While it is true that some linguistic theories are formalized to the extent that they might actually differ empirically, for the most part they are either too vague, or cover such distinct realms of data, that they are strictly speaking incomparable. The choice among linguistic theories is, then, largely a matter of taste, though emotional attachment, training, and propaganda often incline us to think otherwise. One of the things that I find attractive-that suits my tastes in grammar-about the autonomous view that brings us together today is its realism. At least in its present form, the constructs of Autolexical Syntax tend to be rather more real, rather more available to intuition, than the abstractions in certain other traditions of grammar, and much more so than in some. Others would disagree that this is an advantage, and since this is a Geschmackssache, it is their privilege to do so. It is the very abstractness, the very remoteness from anything we can lay our hands or ears on, that attracts other grammarians to other models. Levels of syntactic organization, semantic organization, and morphological organization are, to be sure, somewhat abstract. The phonic signal does not contain their hierarchies and categories any more than it does traces and co-indices, but I submit that it is considerably less taxing on the imagination to think about what referring expression is the semantic argument of the meaning of a verb, than it is to wonder about the grammatical function of an NP in the antepenultimate stratum in Relational Grammar, and it is much easier to have a pretheoretical idea of whether a certain NP is the surface-syntactic subject of a certain verb, than to ruminate upon the question of whether something one does not hear is PRO, NP-trace, WH-trace, pro, or deleted in PF. Chacun ä son gout. The greater reality of the bits and pieces of Autolexical descriptions shows up in the principles that are supposed to (but probably never actually do) keep the theory from being a theory of everything. There is something like a Theta Criterion and something like a Case Filter in the Autolexical model, but the

190

Jerrold Μ. Sadock

principle that, in different modules, covers much of the same conceptual ground as do the Theta Criterion and Case Filter in Government and Binding is so basic, and so mundane, that it scarcely deserves a name, let alone a scientificizing one sporting a Greek letter. The idea is simply that all the pieces of a representation must fit together according to the well-formedness rules of that level of representation. In other words, each level has its own tactics, to borrow a felicitous term from ancestral grammatical thinking. Applied to the semantics, this principle demands that every argument expression be an argument of something, and that every functor that takes an argument have one. This much is merely part of the tactics of the level of semantic representation. Here we have a blunt, homespun version of the Theta Criterion, but without abstract theta roles. Applied to surface syntax, our nameless principle requires that noun phrases (and everything else, for that matter) have some role to play in the syntax of a sentence, and since these roles are, inter alia, subject, object, indirect object, object of preposition, and possessor of N, we have a kind of ingenuous version of the Case Filter, but without abstract Case. Of course not all of the principles of the various "theories" of Government and Binding are automatic consequences of the structure of the Autolexical model (or should be), and there are principles here that have no basis in the architecture of the system, and must be stipulated. But even in such cases, I see a kind of plausibility in the Autolexical account that is lacking in competing ways of thinking about language. Autolexical descriptions are the sorts of things that one could imagine trying to explain to one's mother-in-law. If anything, these stipulations gain appeal as they become less abstract, at least as far as my preferences go. I see it as progress when a relatively arcane idea can be reduced to the conjunction of things we could imagine knowing without a sophisticated grammatical formalism in which to couch them. I'd like to illustrate this by describing a principle, to be found in my book (Sadock 1991), that I consider a real improvement over the way things were done in Sadock (1985). The problem has to do with specifying whether an incorporating form will show up in the position of the incorporator (i.e., the lexeme that needs to combine morphologically) or the incorporee (the morpheme that could stand as an independent element of syntax without attaching morphologically to something else). In other words, when should

Some pleasures and pitfalls of Autolexical Syntax

191

we expect to encounter the situation diagramed in (la), and when should we expect the one in (lb)? (1)

Syntax Morphology

Syntax Morphology In Sadock (1985) I attempted to approach this sort of problem by the relatively abstruse method of counting the number of crossing association lines between the two representations. The idea was that a diagram with fewer crossing lines was to be preferred to one with more. I now think that this effort, which met with only limited success anyway, was misguided. Lapointe (1987) has argued forcefully for a system in which association lines never cross, making the morpho-syntactic interface obey a fundamental constraint of the allied discipline of Autosegmental Phonology (Goldsmith 1976). If Lapointe is right, then we must abandon the attempt to account for ordering in such terms. As it turns out, the facts concerning the direction of incorporation seem to yield easily to a straightforward and intuitively sensible principle that makes no appeal to line crossing. Both kinds of cases clearly occur. All examples of noun incorporation that I know of are either compatible with (1), or demand to be analyzed in that way. In West Greenlandic, for example, the normal order of a stranded modifier with respect to an incorporating verb matches the normal order of an instrumentalcase object with respect to an intransitive verb: 1 (2)

Angisuu-nik

qimme-qar-poq.

big-INST/P dog-have-INDIC/3S 'He has big dogs.'

192

Jerrold Μ. Sadock

(3) Meeqqa-nik asannip-poq. child-INST/P love-INDIC/3S 'He loves children.' Since the order of elements in a West Greenlandic NP is N+MOD, this clearly shows that the incorporated N-stem is attracted to the verb (as in the mirror image of [ l a ] ) , rather than the verbal affix being attracted to the N-stem. As a case of the opposite kind, consider the incorporating determiners of Icelandic, as analyzed in Sadock (1991). If, as I suggest, the morphologically bound determiner in (4) occupies the same position in syntax that a free demonstrative occupies in (5), then we have the situation diagrammed in ( l b ) . (4) raudi hestur-inn sem tyndast red horse-DET which got.lost 'the red horse which got lost' (5) sä raudi hestur sem tyndast DEM red horse which got.lost 'the red horse which got lost' The survey in Chapter 5 of Sadock (1991) contains a near minimal pair of phenomena with respect to the direction of incorporation. Both Hungarian (see, for example, Tompa 1972) and Crow (Graczyk 1989) contain incorporating postpositions, but those of Hungarian descend to the head of their NP complements (if the derivational metaphor may be excused), while those of Crow rise to combine with the head of the phrase that governs them, i.e., the verb phrase. Let us consider Hungarian first. The language has both independent postpositions and those that are obligatorily joined to the head noun of their object phrase. (6)

egy/a feher häzrol 'from a/the white house' (7) egy/a feher häz mellett 'beside a/the white house' If the head noun of the object is preceded by an adjective and followed by a relative clause, the incorporating adposition still seeks the head noun, and winds up attached to it.

Some pleasures and pitfalls of Autolexical Syntax

(8)

193

a szep kep-röl amit tegnap lättunk the beautiful painting-rol which yesterday saw-we 'from the beautiful painting that we saw yesterday'

The Crow case is somewhat more complicated. This language has several clitic postpositions that attach as suffixes to the last word of their object NPs: (9)

Mary-sh ashtäahile is-kawüua-n awäachi-k. Mary-DEF teepee its-inside-LOC sit-DECL 'Mary is sitting inside the teepee.' (10) Bill bin-naask-etaa diili-k. Bill water-bank-along walk-DECL 'Bill is walking along the shore.' One of these, (ku)-ss-(ee), expressing the goal relation, is different from the rest in that it is usually incorporated into the verb. When it is not cliticized to a relational noun like awijua 'inside,' or piishi 'behind,' it still must cliticize to something, namely the empty noun-stem ku- 'it' (example [11]). When it cannot be incorporated (e.g., when there is no host, cf. example [12]), it has the desinence -ee/-aa. But when it is suffixed to a relational noun, both the postposition and the relational noun are incorporated into the verb (example [13]). (11)/} duleesh Jerry-sh Chichüche ku-ss-dee-k. yesterday Jerry-DEF Hardin it-GOAL-go-DECL 'Jerry went to Hardin yesterday.' (12) Shooss-da-lee? Chichüche kussee. where-GOAL-go-INTERR Hardin GOAL 'Where are you going? To Hardin.' (13) Charlie-sh aasüua piishi-ss-xalusshi-k. Charlie-DEF his.house behind-GOAL-run-DECL 'Charlie ran to the back of his house.' There is both phonological and morphological evidence that the postposition and the relational noun are incorporated into the verb. First, the phonological facts: Crow words contain only one stress, and there is only one stress in piishi-ss-xalusshi-k in (13). Note particularly that in the semantically and syntactically parallel sequence is-kawiiua-n awaachi-k'm (9), there are two word stresses.

194

J errold M. Sadock

Thus the division of the string into words that is given in (13) is quite secure. Next, the morphological evidence: In Crow, first and second person arguments are expressed obligatorily as verbal prefixes, e.g., diiwaalichi Ί hit you', biilaalichi 'you hit me', biilichi'(he/she) hit me', etc. According to Graczyk's research, the incorporating relational noun plus postposition must occur inside these verbal prefixes, next to the stem: (14) Mary-sh awaasüua bii-piishi-ss-dee-hche-k Mary-DEF house lS-behind-GOAL-go-cause-DECL 'Mary sent me to the back of the house.' (15) * Mary-sh awaasüua piishi-ssee bii-lee-hche-k Mary-DEF house behind-GOAL lS-go-cause-DECL Here one sees clearly that the complex morphological expression occurs where the verb does, rather than where the postposition does. One might think that the difference has something to do with what item is responsible for the incorporation, but this notion is falsified by the pair of cases at hand, since in both languages the incorporating power is a lexical feature of the adpositions, some of which do not participate in incorporation at all. We may also compare the Eskimo situation, where the lexeme that must be specifically marked as triggering the incorporation is clearly the suffixal verb, with Gta?,2 Sora, and other South Munda languages, where most nouns have special incorporating forms, but those that don't can't incorporate (Ramamurti 1931). In either case, the morphological construct consisting of noun and verb is found in the position of the verb. The actual difference resides in the relationship between the morphological and syntactic rules that the Autolexical element participates in. In both Crow and Hungarian we are dealing with an element that combines with NPs in the syntax to form adpositional phrases. But the morphological combinatorics of the postpositions differ greatly between the two languages. In Hungarian they participate in a morphological rule that combines a noun and a postposition to form a noun, whereas in Crow the adpositions combine with verb stems to form verbs stems. What we can now observe is that the complex noun of Hungarian occurs where the noun would, and the complex verb of Crow occurs where the verb would. In general, then, the complex morphological item occurs where its morphological category occurs in syntax.

Some pleasures and pitfalls of Autolexical Syntax

195

The reason that all noun incorporation involves raising, as Baker (1988) would have it, is simply that noun incorporation, among other things, is the morphological combination of a nominal and verbal form to form a verb, which, by the principle just adumbrated, will occur in the position of the syntactic verb. (16) Syntax Morphology X (17)

Y XP

Syntax

w

YP Y I Y

ZP

Morphology The situation can be understood by examining the diagrams in (16) and (17), which resemble those argued for by Lapointe (1987) in that non-terminal nodes of the morphological tree are associated with the syntactic tree. The diagram in (16) corresponds to the Crow situation, where X = V, and Y = P; The diagram in (17) analyzes the Hungarian facts, where X = P, and Y = N. Though it is quite possible, and certainly desirable, for the principle governing the direction of incorporation to follow from the architecture of the system, I do not see that it does, and so will state it as an observational law: (18)

Direction of Incorporation The syntactic position of a complex morphological expression is the same as the syntactic position of its morphological head.

As far as I know, the rule of Direction of Incorporation always works. In fact, it seems to be generalizable to the semantics/syntax interface as well, though there are some problems. A typical raising

196

Jerrold Μ. Sadock

verb, like English seem, presumably occurs as a propositional operator in semantic structure, but in the syntax it is a verb that takes an infinitive complement. The lexical entry in (19) will allow it to occur in semantic structures like (20) and syntactic structures like (21).3 (19) seem: syntax = [ VP _ VP[to] ] (SF5) semantics = O"1 morphology = V"° (20) F (21) O1 seem

T o see that something like what we found at the syntax/morphology boundary is at work here too, something must be said about the standard identification of categories between these two modules. I think it is reasonable to assume that there is a universal tendency for semantic predicates to be realized as VPs, and for semantic operators to be realized as adverbials, and vice versa. Perhaps the notation could be improved so as to include this cross-identification of categories in these two modules automatically (as it does between syntax and morphology), but I will not attempt that here. Rather, I will once again state these correlations as extrinsic interface constraints: (22) ADV ^ CT1 (23) VP ^ F"1 Now, if we suppose that (18) generalizes to the semantics/syntax interface, we can see why seem is attracted to VP, rather than VP being attracted to seem, giving rise to something like *(it) seems to love John Mary.4 The reason is that the syntactic combinatorics of seem give rise to VPs, which are cross-identified with predicates (F"1), and therefore occur in their slot. This kind of account rests crucially on the separation of grammatical information into encapsulated representational schemes that may or may not produce coincident structures.

Some pleasures

and pitfalls of Autolexical Syntax

197

Therein lies the power of this mode of grammatical description, and also a Pandora's box. The crucial assumption of automodular grammar, and the one that sets it apart from other current theories, is that the components are essentially independent mini-grammars, related to one another by an overarching interface system. How many components we need to recognize, what the reponsibilities of each of them are, and what formal power each of them should be endowed with are all questions concerned with the proper implementation of this basic idea. T h e s e are difficult, but absolutely central, questions in the development of an automodular approach to grammar, since i n c o r r e c t c h o i c e s will inevitably lead to c u m b e r s o m e , and counterintuitive accounts of real-language data, even if the basic idea of autonomous representations is right. Let me consider the question of the formal nature of each of the components first. As I have said before, Autolexical Syntax (despite its name) is not a theory of syntax, nor is it a theory of morphology, nor of semantics. One could, in principle, marry the basic idea concerning the relationship between modules with whatever theory of the individual components one wanted. Of course certain kinds of theories of the individual modules seem more congenial to the spirit of the Autolexical enterprise than others. For example, it would be distinctly incongruous to a s s u m e something like a Government and Binding syntactic component in the present theory. For one thing, the fundamentally derivational style of Government and Binding seems out of keeping with the static, non-derivational philosophy of the present orientation. For another, just the syntactic part of Government and Binding is at least partially a theory of some things that seem semantic and s o m e things that seem morphological. Finally, the Projection Principle, which requires all representations to be essentially similar, goes straight against the grain of Autolexical Syntax, where it is the very possibility of significantly discrepant representations in various dimensions that is the fount of explanation. For t h e s e reasons I have adopted context-free phrase structure grammar as the most suitable available theory of the syntactic component. In fact, it seems perfectly adequate (though not necessarily appropriate) as a theory of each of the big-three c o m p o n e n t s , syntax, semantics, and morphology. If it should turn out to be true that each subcomponent is a phrase structure grammar, this is an intriguing fact indeed. Lieberman (1984) argues that phrase

198

Jerrold Μ. Sadock

structure grammars are merely the non-motor equivalents of automatized motor control mechanisms. Thus the neuronal networks that control complex motor routines are "pre-adapted" to phrase structure grammar. Returning now to the central questions that concern us here, let us consider the problem of distributing the burden among components. Which phenomena are to be handled entirely within a single component, and which are to be dealt with at a particluar interface? Is there a general method of deciding which kind of solution is appropriate, or does it have to be decided on a case-bycase basis? Let us see how such questions apply to a particularly central grammatical construction, the passive. A great many theories of grammar have been founded on their treatment of passives, but not this one. I have never mentioned passives (except in passing) in my work of the last few years. Here something has to be done to make the autonomous model competitive in the empirical marketplace, but I don't know what. Several theories of passives suggest themselves. We could adopt a lexical analysis, which in the present theory would treat the passive morpheme as being a formative in the morphology only. The lexeme would have a syntactic and semantic effect on the verb-form it creates (presumably functions of the syntactic and semantic values of the stems to which it attaches), but would not be represented as an independent element in either of these components. Alternatively, the passive could be handled by a metarule in the syntax, creating new syntactic rules from old ones, though this is possibly too "derivational" an analysis to fit comfortably in the present framework. As a third alternative, the passive auxiliary could be treated as a formative in the syntax, with a semantic and morphological effect but no independent representation in the semantics, and no interesting properties in the morphology. For a language with a purely morphological passive, we might even adopt a hermaphrodite solution that places the passive morphology in the position of an auxiliary verb in the syntax, but gives it the status of an affix in the morphology, making it a classical incorporating element. Yet another possibility that springs to mind is that the core notion of passive cannot be found within any of the components that have been employed in Autolexical work so far, but must be located within a component where syntacto/semantic roles, like agent and theme, or subject and object, are primitives. This

Some pleasures and pitfalls of Autolexical Syntax

199

possiblity leads me to the last topic that I wish to discuss, namely the number and nature of components. There are two directions in which we can err-in the direction of having too many components, and in the direction of having too few. It is hard for me to imagine that any of the components now in vogue is superfluous, and I will proceed on the assumption that we need at least the components we presently have. Furthermore, it is obvious that the complete grammar must somehow include phonological description, but it is not at all clear how that is to be done. The only serious attempt I know of to fill in this glaring gap is the work of Shobhana Chelliah (1995). Besides the rather traditional triumvirate of generative components, there is actually a fourth source of structured representations, namely the lexicon. Even conceived of as a static list, the lexicon must contain complex items-semantic ally irregular morphological items, phrasal idioms, complex semantic categories like pronouns, and the like. Thus the lexicon can provide constituent structure that may not appear anywhere else in the grammar. Donka Farkas and I have employed this level of organization in our paper on the complex distribution of preverbs in Hungarian (Farkas — Sadock 1989). But are there other dimensions of representation, distinct from the traditional three and the lexicon? Is it possible that the components we now have are too coarse, including in some cases information that should actually be separated into autonomous modules? Consider the sort of semantic structures I have been assuming, courtesy of Russell, Carnap, and Quine. These contain two sorts of information, mushed together; function-argument information, and quantifier-variable information. While the putting of these two kinds of information into a single structure is absolutely traditional in modern logic, is there anything of a grammatical nature that compels us to do it that way? Perhaps there should really be two components here, one responsible for functionargument structure and the other for variable binding. In my own recent work, I have often found it convenient to suppress quantifiers so as to isolate function-argument structure. The down side is that if we separate the two kinds of information, something must be done to reconstruct the notion of scope that is so handily represented in the traditional semantic representations. We will need to say that somehow a quantifier can have scope over either predicate in Nancy wants to marry a Norwegian, just as either

200

Jerrold Μ. Sadock

quantifier may have scope over the other in Everybody

loves

somebody. Besides the possibility that there is too much in traditional logical structures, it is also imaginable (some might say obvious) that there is too little. One of the many things Autolexical Syntax in its current state does not do is make any discriminations on the basis of semantic content. As far as the present theory is concerned, intransitive agentive predicates like sing, "unaccusatives" like roll, and "undatives" (if I may call them that) like be hungry, are indistinguishable in semantic structure, all falling under the rather coarse heading of one-place predicates. Now it might be the case, as has frequently been suggested, that classes based not upon semantic configuration, but upon semantic content need to find a place in linguistic description. The question is, then, where to insert such information in an Autolexical model. Schiller (1989) has suggested that semantic structures could be enriched by features relating to content. Alternatively, abstract predicates of doing, undergoing, experiencing, and so on, might be added in the fashion of Generative Semantics. As a third possibility, an autonomous level of thematic organization, as in the work of Lapointe (1988) and Faarlund (1995), might be the appropriate method of capturing generalizations that depend upon semantic content. What about discourse functional information, viz. notions like topic, comment, focus, theme, and rheme, old information, new information, communicative dynamism, and so on? Do these belong in the semantics? Do they belong in the syntax? Both? Neither? Here I'd like to go out on a limb and suggest that it would be profitable to view such notions as defining a parallel, autonomous level of representation, an idea that has been taken up in recent work by Li (1989) and Smessaert (1995). After all, as students of these matters have pointed out (e.g., Firbas 1966), such notions find their expression in natural language in word order (as in Slavic), in terms of morphology, as in Japanese and Korean, in terms of stress (as in English), and even (if Atlas - Levinson 1981 are right) in the semantic representations of cleft and pseudo-cleft sentences. One of the strongest reasons for setting up an autonomous level of organization is, after all, the fact that the information it encodes can surface in the forms that are the responsibility of various components. Thus Eilfort (1989) has argued that illocutionary force be handled by an autonomous component, since here again the formal means by which languages encode speech-act

Some pleasures

and pitfalls of Autolexical

Syntax

201

distinctions make use of the resources of several different components. Another argument for an autonomous module is that discontinuities of structure at a single level can be exchanged for discrepancies between continuous constituents at two levels, as is done in the locus classicus of automodular phenomena, noun incorporation. Now topicalization produces discontinuities, so it might well be a perfect case for reanalysis in automodular terms. What would an automodular grammar of the topic-comment component of a language like Czech, said to have topic before comments, look like? In and of itself, it would not be particularly exciting, perhaps only the rules in (24-26), or only a subset of them. (24) Utterance (U) = Topic (TOP) Comment (COM) (25) TOP = TOP COM (26) COM = TOP COM Now since the simple structures that are the province of a component like this must be associated with both syntactic and semantic representation, it looks possible to explain some of the well-known interactions between discourse-functional notions and syntax or semantics. For example, the fact that subjects tend to be default topics is just what we would expect in a language where subjects are the default initial constituents of sentences. The normal case would correspond to a bimodular situation in which there is no conflict whatsoever between the autonomous representations: (27)

S NP I

VP I

TOP

COM

U Perhaps more interestingly, some of the effects of topic/comment structure on preferred or required semantic scope begin to make sense in a theory with an autonomous discoursefunctional component. As has often been observed (see Partee 1970, for example), there is a tendency to interpret quantifiers as having scope from higher to lower mapped onto linear order from

202

Jerrold Μ. Sadock

left to right. If a non-standard syntactic order occurs, this association is virtually categorical. In a familiar pair like (28) and (29), there is a clear tendency to interpret the first NP as having scope over the other, so to interpret some girls as non-specific in (28), and specific in (29). But in (30) it is almost impossible to interpret some girls as non-specific. (28) Most boys danced with some girls. (29) Some girls were danced with by most boys. (30) Some girls, most boys danced with. If we examine a rough-and-ready semantic/discourse-functional mapping for these examples, these facts begin to make sense. (31)

F Some girls

F

y

y

υ (32)

F

U If (29) or (30) has the specific reading of some girls then the dual diagram is as in (31). On the other scopal interpretation it is

Some pleasures and pitfalls of Autolexical Syntax

203

as in (32), a structure that involves a violation of one of the default mapping principles, namely Linearity. Structure (30) would also violate Linearity at the discourse/syntax interface on the disfavored scopal interpretation. There is much more to be done in the way of giving this kind of an analysis real teeth, but it does seem to hold a great deal of promise. But where will it all stop? How many components can we postulate before we end up with a separate component for each phenomenon, much as Relational Grammar went through a baroque phase in which there was a separate grammatical relation for each phenomenon? I can't tell you, but I can suggest that the spirit of realism that I mentioned at the outset should serve as a guide. My c o n s c i e n c e is fairly clear when it comes to suggesting a topiccomment plane of organization, b e c a u s e that strikes me as something I could convince my mother-in-law exists. I might have trouble sleeping, however, if it seemed necessary to set up a level of syntactic deep structure, or a structure in which superscripted co-indices were represented, or anything of that kind. Before closing, I'd like to point out one practical feature of the model that we should be aware of: the formalism is so easy to implement that it becomes a burden. The Autolexical model, with its relatively simple individual components, makes it possible to get quite explicit about the syntax, semantics, and morphology of a particular expression. But the syntax of that expression is only part of the syntax of the language, and similarly for the other components. A respectable description of a single phenomenon in Autolexical terms should really be embedded in at least a fragmentary, integrated description of the entire language. In other theories that contain much more complex components, it might seem reasonable to assume some version of X-Bar syntax, ccommand, or whatever, without actually saying what that version is. In the present theory it simply will not do to assume some version of English syntax, or some version of compositional semantics without saying what it is. While this might seem a chore, I have found that when one actually writes out at least the relevant rules in all components, a great deal, both about the language under scrutiny, and about the descriptive apparatus, is quickly learned.

204

Jerrold Μ. Sadock

Notes 1. 2. 3.

4.

The modifier or the argument may also appear after the verb, but with the same stylistic effect in both cases. I am grateful to Norman Zide for information on Gta?. The notation here is that of Sadock (1991). It owes much to suggestions of Eric Schiller's: "F" indicates formula, " 0 " indicates operator, and a negative superscript indicates degree of unsaturation. Thus "F~l" indicates a singly unsaturated proposition, i.e., a function from entity expressions to propositions. This would be ungrammatical anyway, since the lexical entry for seem in (19) requires it to take an infinitive complement. However, the present framework, in contradistinction to others, has no prejudice against overlapping accounts of the same facts.

References Atlas, Jay David - Stephen Levinson 1981 "/t-clefts, informativeness, and logical form: Radical pragmatics (Revised Standard Version)", in: Peter Cole (ed.), Radical pragmatics. New York: Academic Press, 1-62. Baker, Mark C. 1988 Incorporation: A theory of grammatical function changing. Chicago: University of Chicago Press. Chelliah, Shobhana 1995 "An Autolexical account of voicing assimilation in Manipuri", in this volume. Eilfort, William H. 1989 The illocutionary module. Paper presented at the Workshop on Autolexical Syntax, University of Chicago, April 16, 1989. Faarlund, Jan Terje 1995 "Autostructural analysis of semantic roles", in this volume. Farkas, Donka F. - Jerrold M. Sadock 1989 "Preverb climbing in Hungarian", Language 65: 318-338. Firbas, Jan 1966 "On defining the theme in functional sentence analysis", in: Josef Vachek (ed.), Travaux linguistiques de Prague 1: L'Ecole de Prague d'aujourd'hui. University of Alabama: University of Alabama Press, 267-280. Goldsmith, John 1976 "An overview of Autosegmental Phonology", Linguistic Analysis 2: 23-68.

Some pleasures and pitfalls of Autolexical Syntax

205

Graczyk, Randolph 1989 Postpositional phrase incorporation in Crow. Paper presented at the Siouan Languages Conference, June 2-3, 1989, Morley, Alberta, Canada. Lapointe, Steven G. 1987 "Some extensions of the Autolexical approach to structural mismatches", in Geoffrey J. Huck - Almerindo E. Ojeda (eds.), Syntax and Semantics, Volume 20: Discontinuous constituency. Orlando, FL: Academic Press, 152-184. 1988 "Distinguishing types of morphosyntactic cooccurrences: Mismatch resolution, agreement, government", in Diane Brentari - Gary Larson - Lynn McLeod (eds.), Proceedings of the 24th Annual Regional Meeting of the Chicago Linguistics Society, part 2: Parasession on Agreement in Grammatical Theory. Chicago: Chicago Linguistic Society, 181-201. Li, Ligang 1989 ßa-construction in Mandarin Chinese: An Autolexical approach. Paper presented at the Workshop on Autolexical Syntax, University of Chicago, April 16, 1989. Lieberman, Philip 1984 The biology and evolution of language. Cambridge, MA: Harvard University Press. Partee, Barbara Hall 1970 "Negation, conjunction, and quantifiers: Syntax vs. semantics", Foundations of Language 6: 153-165. Ramamurti, G.V. 1931 A manual of the So.ra: (or Savara) language. Madras: Government Press. Sadock, Jerrold M. 1985 "Autolexical syntax: A theory of noun incorporation and similar phenomena", Natural Language and Lingusitic Theory 3: 379-440. 1991 Autolexical syntax: A theory of parallel grammatical representations. Chicago: University of Chicago Press. Schiller, Eric 1989 The case for Autolexical case. [Unpublished manuscript, University of Chicago] Smessaert, Hans 1995 "Pronominal cliticization in West Flemish", in this volume. Tompa, Jozsef 1972 Kleine ungarische Grammatik. Budapest: Akademiai kiado.

Toward a tri-modular analysis of -ly adverbs* Robinson H. Schneider

Introduction By utilizing autonomous modules and no transformations, it will be shown that Autolexical Syntax can provide an analysis of -ly adverbs which is both more elegant and more explanatory than currently competing analyses. Drawing heavily from the work of Sadock (1985, 1987, 1991) and Lapointe (1987a, 1987b, 1988), that will be the goal of this paper. An overview of the problems inherent to the study of adverbs and a current analysis will be presented. A tri-modular Autolexical analysis of the same phenomena will follow, with emphasis given to the interface. In particular, Sadock's (1991) constraints on the morpho-syntactic interface will be applied to the syntactic-semantic interface and assessed for viability as more general constraints on intermodular association.

1. Definition of terms Adverb, for the purposes of this work, will refer to the so-called -ly adverbs. Since these adverbs, like probably, frankly, cleverly, etc., are universally considered to be prima facie adverbs, this seems to be the logical place to start constructing an analysis of adverbs. In addition, this work further restricts itself to those -ly adverbs which have been thought to semantically modify propositions (the ad-S), predicates (the ad-VP), and relations (the ad-V). The use of syntactic terminology-S, VP, and V-for semantic distinctions may initially be confusing; however, since these terms are already widely in use throughout the literature, it is felt to be unnecessary to introduce new terminology. "Module" is a term specifically associated with the theory-ofchoice for this work, Sadock's Autolexical Syntax. Autolexical Syntax

208

Robinson Η. Schneider

is an autonomous modular grammar in which each module handles only those aspects of the grammar which pertain to it, i.e., each module functions autonomously. The idea of modularity is not new to linguistic theory. However, these modules have standardly been arranged hierarchically, with the output of one module acting as an input to the next. In Autolexical Syntax, each module generates its own output independent of other modules. The interface then coordinates and constrains these possible outputs to the grammatical utterances of the target language. A familiarity with the basic tenets of Autolexical Syntax is assumed. A tri-modular analysis will refer to the three modules to be developed here: morphological, syntactic, and semantic. The interface between modules-in particular, the syntactic and semantic components-and the nature of the lexicon will also be discussed to provide a unified Autolexical analysis of adverbs.

2. The empirical challenge of adverbs Adverbs have been traditionally classified as modifiers of verbs, adjectives, and other adverbs. In addition to the so-called -ly adverbs, other candidates for adverb-hood have included not, very, almost, and the like. Despite the fact that grammarians have long recognized this heterogeneous class called "adverb", current linguistic theory has yet to legitimize an independent category of Adv within its four-way classification of ±N, ±V. The biggest challenge to any analysis of adverbs is accounting for the mismatch between surface syntactic structure and semantic representation. Consider the following: (1) a. Probably John will have arrived in Boston by now. b. John will have probably arrived in Boston by now. (2) a. Mary reluctantly poured syrup on her waffles. b. Reluctantly Mary poured syrup on her waffles. Probably in CI) is an example of an ad-S, an adverb semantically modifying the proposition, i.e., what is probable is that John will have arrived in Boston by now. Reluctantly in (2) is an ad-VP, i.e., Mary is reluctant to pour syrup on her waffles. (Quickly in Harry

Toward a tri-modular analysis of-ly adverbs

209

swam quickly would have the ad-V reading of 'in a quick manner'.) In the (a) examples, the nonitalic adverb is adjoined to the constituent that it modifies in both syntactic and semantic structure, as shown in (3) and (4), respectively. (3) a. Syntactic Structure for ( l a )

b. Logical form for ( l a ) Formula

Adv probably

ADV PROBABLY John will have arrived in Boston by now.

(4) a. Syntactic Structure for (2a)

JOHN WILL PAST ARRIVE IN BOSTON BY NOW. b. Logical form for (2a)

S NP Mary

Formula VP

Adv reluctantly

Formula

Arg MARY VP

Pred

ADV RELUCTANTLY λχ

Pred Formula

poured syrup on her waffles

χ POUR SYRUP ON HER WAFFLES In (la), probably is adjoined to an S in syntactic structure ([3a]) and to the S-counterpart, Formula, in (3b). The structures of (4a-b) show that reluctantly is adjoined to a VP in syntactic structure and to the VP-counterpart in semantic representation, Predicate (Pred). This is the default case, that syntactic and semantic structure are topologically alike. The (b) examples, however, show an apparent mismatch between the syntactic and semantic structures. In (lb), an adverb which is semantically modifying a Formula is syntactically adjoined to a VP. Example (2b) presents the reverse phenomenon; that is, an adverb which is semantically modifying a Predicate is syntactically adjoined to an S. Assuming that semantic

210

Robinson Η. Schneider

representation will remain unchanged from that specified in (3b) and (4b), examples (5) and (6) demonstrate this structural change in the syntax for (lb) and (2b), respectively. (5) Syntactic Structure (lb) S NP John

VP V

VP V

will

have

VP Adv

VP

probably

(6) Syntactic Structure (2b)

S

Adv reluctantly

arrived in Boston by now.

S Mary poured syrup on her waffles.

Note that adverbs are not the only phenomena to exhibit structural mismatches. Other examples include Dutch cross-serial dependencies and cliticization (Sadock 1991) and noun incorporation in West Greenlandic (Sadock 1985). The challenge for the linguist is to account for this "mismatch" within his chosen theoretical framework whether he utilizes transformations, projection rules, or modular components. The challenge for this work is to show that a modular analysis can provide the most accurate and elegant analysis possible.

3. A transformational analysis McCawley's (1983) analysis of some English -ly adverbs makes a strong case for the existence of adverbs which modify propositions (ad-Ss), adverbs which modify predicates (ad-VPs), and adverbs

Toward a tri-modular

analysis of-ly

adverbs

211

which modify verbs (ad-Vs). Building on Thomason - Stalnaker's (1973) observations of these adverbs' interactions with quantifier scope on the subject NP, McCawley (1983: 264) demonstrates that probably is an ad-S, while intentionally and completely are not. (7) a. Most of the voters probably prefer Reagan to Kennedy. b. Most of the voters intentionally left their ballots blank. c. Most of the voters completely reject Reagan. While (7a) asserts that what is probable is the proposition that most of the voters prefer Reagan to Kennedy, intentionally and completely in (7b) and (7c) cannot assert that it is intentional (?) or complete (?) that most of the voters left their ballots blank or reject Reagan, respectively. McCawley concludes that "probably is outside the S that serves as its host, while intentionally and completely are within the host S" (1983: 264). The adverbs in examples (7b) and (7c) are not combined with the proposition, but rather are within the proposition, i.e., that each of the voters by intention left his ballot blank or what was complete was his rejection of Reagan. The distinction between intentionally and completely is made apparent by a similar contrast of quantifier scope within the VP. Consider (8) (McCawley 1983: 264): (8) a. The invaders completely destroyed all three villages. b. Marvin intentionally sliced all three bagels. The adverb in (8a), as an ad-V, is interpreted as saying that for each of the three villages the invaders completely destroyed it; the adverb in (8b), however, is ambiguous depending on whether it is interpreted as being inside or outside the scope of the quantifier. It can either mean that for each of the three bagels, Marvin intentionally sliced it, or that Marvin had a single intention to slice all three bagels. Within the transformational framework in which McCawley is working, the canonical configuration for modifiers is one in which the modifier is Chomsky-adjoined to the constituent which it is semantically modifying. Therefore, (9a) would be the base form for ad-Ss, (9b) for ad-VPs, and (9c) for ad-Vs. Presumably each base form would also count as a specification of the semantic structure for each adverbial type. However, McCawley (1988) makes explicit that the semantic structure for ad-VPs in (9b) is derived from (10) by Equi-NP Deletion.

212

Robinson Η. Schneider

(9)

a.

b.

c.

S Adv

(10)

s NR

S NP V'

Adv

V' V

NP

Adv

?

Adv

S

NPj V' Once the semantic structure has been established, the various surface constituent structures in which these adverbial types may appear can be ascertained. McCawley makes this determinination by evaluating the grammaticality of sentences in which the position of the adverb changes, but which are identical in all other respects. Consider (11) (from McCawley 1983: 266): (11)

The enemy will have destroyed the village./,

probably intentionally completely

t

OK * *

ttt

OK OK * OK * ??

?? OK OK

t

.1

tt

2

* OK OK * OK*

Assuming the base configurations of (9), many of the findings of (11) are not unexpected. These are summarized in (12): (12) Position/type S-initial3 S-postfinal Before/after V S-final Within/adjacent to innermost VP

ad-S Y Y Y Ν Ν

ad-VP Ν Ν Υ Υ Ν

ad-V Ν Ν Ν Υ Υ

These include: (i) Ad-Ss can occur S-initially, but ad-VPs and ad-Vs cannot. (ii) Ad-Ss can appear in S-postfinal position.

213

Toward a tri-modular analysis of- ly adverbs

(iii) Ad-VPs can occur in S-final position, pronounced as a single intonational unit with the constituent they semantic ally modify; they cannot be pronounced parenthetically. (iv) Ad-VPs can precede or follow any VP other than the innermost one; that is, an ad-VP can precede any VP and can follow any VP except that it cannot follow the innermost VP unless in S-final position. (v) Ad-Vs cannot occur in S-final position with a parenthetical pronunciation. (vi) Ad-Vs must be within or adjacent to the innermost VP, i.e., an ad-V must precede or follow the VP which contains the verb that it semantically modifies. However, there are also at least two inconsistencies which cannot be accounted for by the canonical base structures alone. The first is that ad-Vs should not be able to occur in S-final position, i.e., (13) should be ungrammatical, given the base configuration. (13) The enemy will have destroyed the village completely. (14) a. They reduced the village to ruins.

t t

OK

completely a'.

generously b.

John

John

b'.

inadvertently

spread peanut

t t

They

butter

on the cracker

t

OK OK kept the car in the garage OK

stupidly

t

OK

reduced

t

OK

t

*

the village

.

t ?

t

.

?

to ruins

t

??

t

.

OK

However, examples like (14) argue against a re-analysis of an ad-V to an ad-VP (McCawley 1983: 267-268), because ad-Vs ([14a] and [14a']) can occur between the object NP and other material in the

214

Robinson Η. Schneider

VP; ad-VPs ([14b] and [14b']) cannot. On the basis of the data presented in (14), McCawley argues that these kinds of sentences in which the ad-V appears after the object of the verb it modifies have the discontinuous surface structure of (15a) or are generated by some sort of transformation which would flatten the structure, as in (15b). (15) a.

V'

b.

V' V

destroy

NP

Adv

the village

completely

The second finding which must be accounted for is that ad-Ss can occur on either side of a tensed auxiliary verb (but not within a non-finite VP). Examples like (16) (McCawley 1983: 268) argue for an optional transformation of Raising which yields the surface structure of (17). (16) a. The invaders both [probably have destroyed and [certainly have killed the hostages]. b. The invaders both [have probably destroyed and [have certainly killed the hostages].

the

village]

the village]

(17)

the invaders the invaders have destroyed the village probably have destroyed the village

Toward a tri-modular analysis of- ly adverbs

215

An additional transformation of Attraction-to-Tense (McCawley 1988) would generate (16b). This is detailed in (18).

destroy

the

village

This is just one of several possible analyses for the Autolexical analysis to follow. For adverbial semantics, of particular interest are Thomason — Stalnaker (1973) and McConnell-Ginet (1982), although Huck (1984) has shown that the structural incongruities of adverbs pose a serious problem to Montague grammarians. Jackendoff (1972) and, more recently, Ernst (1984), use a combination of transformational syntax and semantic interpretation rules, which has been shown by Schneider (in prep.) and others to be unnecessarily complex and empirically inadequate. McCawley's analysis presented above is the most viable of the current analyses of the empirical facts, but it sheds no light on why adverbs behave in this manner.

4. An Autolexical analysis Having established the parameters for an Autolexical analysis of adverbs, the contents of the individual components and their interface can be specified. Given space constraints, the following analysis will concentrate on the paradigm cases presented in the last section. The system of morphology to be used will be that of Sadock (1991) which contains the following rules:

216 (19)

Robinson Η. Schneider (Ml)

(M2) (M3) (M4)

X[-0] X[-l] X[-2] X[-0]

= AF, Y[-0] = X[-0] AF = Y [ - l ] AF = Y[-0, CF] X[-0]

The bar levels in these rules specify distinct morphological categories: stems [-0], words [-1], and "superwords" [-2]. Rule Ml adds derivational affixes to bound morphemes; M2, inflectional affixes; M3 adds clitics to words; and M4 generates compound words. The -ly suffix in English is a derivational affix which converts many adjectives to adverbs. Its lexical entry is in (20):4 (20)

-ly morphology = -ly, AF [-infl]; Ml where X=Adv and Y=Adj syntax = nil semantics = nil

All -ly adverbs fall into one morphological category: Adv[-0]. This category is either generated by rule Ml introduced in (19) such that X[-0] = AF, Y[-0] where X = Adv, Y = Adj, and AF = -ly, or individual adverbs can be directly assigned to this morphological category. For example, those adverbs which have an obvious semantic relationship to their corresponding adjective, like

probably, intentionally, completely, and rudely would have in their lexical entry for the morphological component (21). Adverbs like hardly would have the partial lexical entry of (22). (21) (22)

probably hardly

probably; Adv[-0 ]; Μ1 hardly; Adv[-0]

The proposed semantic analysis, like Montague grammar, will consist of two parts: a syntax and an interpreted intensional logic. The syntax, similar to McCawley's (1981) predicate logic, will only specify the combinatorics for logical categories, that which constitutes a well-formed semantic representation. These rules will function as node-admissability conditions, whose constituents are unordered, like the context-free phrase structure rules of the syntax proper. The constant of intensional logic which constitutes the logical form of the lexical item of English will be referred to by the capitalized variant of the appropriate lexeme, e.g., walk V WALK (= Montague's walk').

Toward a tri-modular analysis of-ly adverbs

217

The basic definition of a well-formed semantic phrase is (23): If γ is a well-formed category, then for γ : α β, where γ dominates both α and β, α is a function which associates β to γ. LF fragment (24) LF1 Formula: ~ Formula LF2 Formula: ν Formulaj, (n > 2) Formula: Λ F o r m u ^ (n > 2) LF3 LF4 Formula: -> Formula Formula LF5 Formula: IOP Formula LF6 Formula: Binder Formula Binder: Q Formula LF7 Formula: Pred Arg LF8 Pred: R Arg LF9 TOP Formula LF10 Pred: Some Semantic Syntactic Category Definition Equivalents Formula formula S or proposition walk, swim Pred predicate V love(s) Mary VP Arg argument (Formula, Pred, etc.) basic entities (a, b, c, . ..) 5 relation V R love, give 6 quantifier phrase NP Binder all boys quantifier Det every, the, some Q (or determiner) intransitive operator V seem IOP7 7 transitive operator V want TOP

(23)

The condition on well-formedness in (23) should not be interpreted as having directionality. It simply states that semantic representations consist of branching nodes such that each nonterminal node is made up of two categories in a function-argument relationship. Given this definition, a variety of formulae (cf. McCawley's Ss [1981: 91]) can be generated as in (24).7 Most of the category names are self-explanatory; however, an explication with syntactic equivalents and examples is given in (25). This lists

218

Robinson Η. Schneider

the kinds of syntactic/semantic pairs which might be associated when relating syntactic structure to semantic representation. As will be seen below, adverbs will belong to several different semantic categories. The logical operators of LF1-LF4 have their usual interpretation and, like other lexical categories, can be found in the lexicon. Nonlexical categories like Formula and Binder are semantic primes. What distinguishes this analysis from others of the same type is that much of the semantic work is done by the lexicon. The lexicon will specify how many arguments a particular predicate will require and will also delineate to which category (or categories) a lexeme belongs. This allows for a simplification of the semantic description. Given the semantic program outlined here, it is clear that the ad-S is an intransitive operator, IOP, like seem. No evidence has been found to reject the standard analysis that ad-Ss are functors which take propositions-Formula in this analysis-and return propositions. By saying that ad-Ss are IOPs, one is saying that they participate in LF5-Formula: IOP Formula. This captures the standard paraphrase of it is AD J that (Formula) or it is x-ly true that (Formula) because the adverb has scope over the entire proposition. The basic semantic structure for the ad-S can be found in (26). (26) Probably John loves Mary. Formula IOP

Formula

PROBABLY JOHN LOVE MARY The fact that one can also say John probably loves Mary and John loves Mary, probably is a syntactic matter, not a semantic one and will be dealt with below. Another semantic property of ad-Ss is that one can get scope ambiguities with the logical subject. For example, (27) can mean that it is certain that all of the men left or that for all men, it is certain that each man left. These readings can be captured by the structures of (27a) and (27b), respectively. This is possible because two formulas-FormulaQ and Formula^are available for combination with the IOP of certainly in (27).

Toward a tri-modular analysis of Ay adverbs

219

(27) a. Certainly all men left. b. All men certainly left. Formula

b) CERTAINLY The ad-VP functions as a TOP, a transitive operator, like want. The basic semantic structure for the ad-VP can be found in (28).

(28) Joan reluctantly instructed Mary. Formula Arg JOANx

Pred TOP

Formula

RELUCTANTLY χ INSTRUCT MARY This captures the standard paraphrase of Joan is ADJ to instruct Mary. Ad-VPs should also be able to exhibit scope ambiguities with the object-place argument, but not with the logical subject. For example, (29) can mean that Joan sliced all of the bagels slowly or that for each bagel, Joan sliced it slowly. These readings can be captured by the structures of (29a) and (29b), respectively. For ad-Vs, a new semantic category of ADV is created for adverbs which modify verbs. 8 This is standardly paraphrased as V'd in an ADJ manner. The manner-reading for carefully is detailed in (30). Given the semantic structure of (30), it is clear that ADV will always be within the scope of the Binders on the subject-place or object-place arguments.

220

Robinson Η. Schneider

(29) Joan slowly

sliced

all the

a.

Formula

Arg I JOAN

Pred TOP

bagels.

Formula

SLOWLY

Binder, / V ALL BAGEL

Formula χ SLICE y

Formula Binder.,

Formula

ALL BAGEL Arg

Pred

JOAN.

TOP I SLOWLY

(30) Joan sliced the bagel carefully. = Joan sliced the bagel in a careful

Formula χ SLICE y manner.

Formula

THE BAGEL Arg I JOAN

SLICE CAREFULLY In sum, semantically the -ly adverbs fall into one of three classes: IOP, TOP, or ADV. This analysis only necessitates one new category of ADV which is designed to handle ad-Vs. The other adverbs can be accommodated by existing semantics. Example (31) is offered as evidence that this analysis can handle the different interpretations of an adverb in the same sentence. Interpretations

Toward a tri-modular

analysis of-ly

adverbs

221

(a) and (b) are expected: the fact that there is a scope ambiguity between the adverb and subject-place argument is evidence of an ad-S. The ad-VP interpretation of rudely in (31c) can be distinguished from the ad-S reading of (31b) in that the ad-S reading is commenting on the fact of all of the students leaving, whereas the ad-VP reading implies that the students were agents in the rude leaving. The final interpretation is the ad-V manner reading. (31)

All of the students

t a

rudely a. b. c. d.

It was It was All of All of

departed

L

b,c,d

./,

.

U

c,d a,D

rude for all of the students to depart. (ad-S) rude for each student to depart. (ad-S) the students were rude to depart (ad-VP) the students departed in a rude manner. (ad-V)

Formula Formula (a) ΙΟΡ , ^ I Binder x Formula RUDELY ^ ^ ALL STUDENT (b) IOP Formula RUDELY

χ DEPART

Formula Binder, ALL STUDENT

Formula Arg I χ

Pred (c) TOP

Formula

RUDELY χ DEPART

Robinson Η. Schneider

222

Formula Binder,

Formula

ALL STUDENT Arg Pred I χ (d) ADV I RUDELY

Pred I DEPART

Given the summary of syntactic positions available to each of the individual classes of adverbs summarized in (12) and the semantic considerations discussed above, the specification of the syntactic phrase structure rules for adverbs is straightforward. The syntactic component will incorporate many of the rules and principles of Generalized Phrase Structure Grammar detailed in Gazdar et al. (1985). A familiarity with this context-free phrase structure grammar is assumed. In order to fully understand the syntactic phenomenon of adverb position, some discussion of the data in (32) and (33) is necessary. (32) a.

inadvertently b. completely

They reduced the village

to ruins

.

t

t

t

OK

??

OK

They reduced the village

t

OK

to ruins

t

OK

t *

The basic assumption is that when "other material" appears in the VP, e.g., to ruins in (32), then ad-Vs like completely can appear VP-initially and before the "other material", but ad-VPs like inadvertently can only appear VP-initially and VP-finally. Ad-VPs cannot appear within the verb phrase or only marginally so.9 On the other hand, if the "other material" appears outside the VP, like time-adverbials, then, not surprisingly, a different distribution occurs, as in (33).10

Toward a tri-modular analysis of- ly adverbs (33) a.

They

inadvertently b. completely

plundered

t

the village

OK They

plundered

t

OK

the village

after 5 PM

223 .

t

t

OK

OK after 5 PM

t

OK

.

t *

Any analysis of these two types of sentences must predict these observations. Starting from the basic assumption that syntactic structure is topologically similar to semantic structure, (34a-b) is generated as a (partial) template for those sentences like (33a-b), respectively. In (34a), if the ad-VP is a sister to the available VPs, then the correct predictions are obtained for positions (a), (b), and (c). For (34b), if the ad-V acts as a sister of the V in syntactic structure, then positions (d), (e), and (f) are available with position (e) being ruled out unless the object is "heavy" (Keyser 1968). This, too, predicts the correct placement of ad-Vs. (34) a. inadvertently: ad-VP with VP-external PP

after 5 PM

224

Robinson Η. Schneider

When extra material appears within the VP, as in (32), there are two problems for the analysis of ad-Vs. Consider (35a-b). (35) a. inadvertently: ad-VP with VP-internal VP

reduced

the village

b. completely: ad-V with VP-internal VP

Again, in (35a), one gets exactly the right predictions for the ad-VP positions: (a) and (b). However, position (d) and position (f) in (35b) are not allowable positions for the ad-V and must be excluded. Position (d) can be ruled out because general constraints on the grammar do not allow anything to appear between a verb and an object which is not "heavy". By utilizing the notion of immediate precedence (symbolized by " 1 PROBABLY; IOP

b.

intentionally

intentionally, Adv[-0]; Ml ; Adv[0]; SF1 where a=l INTENTIONALLY; TOP

c.

completely

completely, Adv[-0]; Ml ; Adv[0]; SF2 COMPLETELY; ADV

Following Lapointe (1988), the entry for the morphological component is italicized; the entry for the syntactic component in angled brackets; and the entry for the semantic component in capital letters. Entry (42a), for example, states that probably is an adverb morphologically consisting of an adjective and an affix. Syntactically, it is a sister to any V[Bar α ] where α is greater than or equal to 1. Semantically, it is an intransitive operator. It is this kind of straight-forward description that makes Autolexical Syntax so appealing.

6. The interface Now that the contents of the individual components have been determined and specified via the lexicon, the interface between components can be examined. Since the interaction of the morphological component with the syntactic and semantic components does not exhibit the kind of structural incongruities needing resolution at the interface, it will not be examined here.

Toward a tri-modular analysis of-ly adverbs

231

The following discussion will concentrate on the interface between the syntactic and semantic modules. Sadock (1991) has proposed two constraints on the interface between the morphological and syntactic components. Each of these constraints has a strong and a weak version. Using the data on adverbs presented above, it can be shown that these constraints are, in fact, more general conditions on the association of modules. Examples (43) and (44) are Sadock's morpho-syntactic conditions on allowable discrepancies re-written as conditions on the semanticsyntactic interface. (43) Linearity Conditions a. Strong The associated elements of semantic and syntactic representations must occur in the same linear order. b. Weak The associated elements of semantic and syntactic representations must occur in the same linear order except where the semantic and syntactic requirements of lexemes make this impossible. (44) Constructional Integrity Conditions a. Strong If the maximal SS lexical category of an element combines with an SS category Ρ in the syntactic structure and the maximal LR lexical category for its associated element combines with another LR category to yield category Q in semantic structure, then LR category Q must be associated with the head of the SS category P. b. Weak If the maximal SS lexical category of an element combines with an SS category Ρ in the syntactic structure and the maximal LR lexical category for its associated element combines with another LR category to yield category Q in semantic structure, then LR category Q must be associated with some descendent of the head of SS category P. The claim is that Sadock's General Morphosyntactic Homomorphism Condition can be re-written as a general semantic-syntactic homomorphism condition. This condition is in (45).13 Given the great flexibility in adverbial position in surface structure, it is not surprising that adverbs obey a combination of weak and strong

232

Robinson Η. Schneider

homomorphic conditions. The ad-S, the ad-VP, and the ad-V will be examined in turn. (45) General Semantic-Syntactic Homomorphism Condition Let each of the strong homomorphism conditions count as two degrees of similarity, and let the weak homomorphism conditions count as one. The total degree of similarity between autonomous syntactic and semantic representations must be at least two. Consider (46), which shows a comparison between semantic ([46a]) and syntactic([46b]) structure for the ad-S. John must have left,

(46)

ta tb Μ c

certainly

ίe

d

a.

y χ

x

LEAVE

b.

a

b

c

d

left

e

Toward a tri-modular

analysis

of Ay adverbs

233

Each Formula in (46a) meets the requirements of LF5 for intransitive operators, so certainly can appear in any of these semantic positions (v-z). In (46b), positions (a-e) for ad-Ss are filled syntactically via SF1 if the bar level for V is greater than or equal to 1. Semantic position (v) is associated with syntactic position (a), (w) with (b), (x) with (c), (y) with (d), and (z) with (e). Linear order is only weakly maintained because the logical argument in semantic structure is associated with an NP which must occur S-initially in English, syntactically. However, the association between these two structures fulfills the strong Constructional Integrity Condition. The SS lexical category which combines with Adv is V[Bar a] where a > 1 in syntactic structure (= SS lexical category P). The associated element to Adv is IOP in semantic structure which combines with Formula to yield a Formula (= LR category Q). These Formulae are associated with the heads of SS lexical category Ρ in syntactic structure. This brings the degree of similarity to 3, which maintains the General Semantic-Syntactic Homomorphism Condition in (45). Example (47a-b) shows a comparison between the semantic and syntactic structures, respectively, of the ad-VP. John

(47)

must

have

left,

11 11

intentionally

a

b

e

d

Formula

a.

Pred

Arg JOHN,

PAST

w

x

LEAVE

ζ

Formula Formula

IOP MUST

χ

ζ

234

Robinson Η. Schneider Formula Formula

ΙΟΡ I MUST

ΙΟΡ I PAST

Formula Ar§

JOHN

Pred Formula^ LEAVE

John

Semantic position (w) is associated with syntactic position (a), (x) with (b), (y) with (c), and (z) with (d). For the adverb,14 both the Linearity Constraint and the Constructional Integrity Condition in their strong forms are met. Therefore, the degree of similarity for ad-VPs is 4, well within the demands of the General SemanticSyntactic Homomorphism Condition in (45). (48)

John bypassed the accident..

completely a. Formula Arg"

t

t

a

b

ζ

Arg I ACCIDENTy

Toward a tri-modular analysis of- ly adverbs b.

235

S VP

NP John a

bypassed the accident

b

Finally, one must consider the ad-V. Examples (48a-b) detail the syntactic and semantic structures for the ad-V. As a sister of bypass in (48b), one would expect the adverb to be found in the syntactic positions (a), (b), and immediately after the verb. However, this latter position is not eligible for consideration because of general constraints on the syntax via IP1. In order to get syntactic position (b), the Constructional Integrity Condition is only weakly met. The SS lexical category which combines with Adv is V[Bar 0] (and an NP) in syntactic structure (= SS lexical category P) whose head is a VP. The associated element to Adv is ADV in semantic structure which combines with the R (dominating BYPASS) to yield another R (= LR category Q). This R (=LR category Q) is associated with the descendent (=V[Bar 0]) of the head (V[Bar 1]) of SS lexical category Ρ in syntactic structure. The strong version of the Linearity Condition is then violated. In order to maintain the strong version of the Constructional Integrity Condition, LR category Q would have to be associated with VP. Again the General Semantic-Syntactic Homomorphism Condition is fulfilled with a degree of similarity of 2.

7. Conclusion What this work has tried to show, through the analysis of some paradigm cases of adverbs, is that Autolexical Syntax can provide an more elegant and explanatory model of adverbs' behavior than previous analyses. In addition, it is empirically more accurate in that the transformational analysis cannot not generate ad-Ss after two auxiliaries without the addition of another rule. The advantage of this analysis over a transformational analysis is that now the behavior of adverbs becomes transparent. Each adverb modifies a different semantic entity; the ad-S has scope over the proposition,

Robinson Η. Schneider

236

the ad-VP has scope over the semantic equivalent of the VP, and the ad-V has s c o p e over the syntactic equivalent of the V. Syntactically, these adverbs can occur within the domain defined by the semantics. The closer the syntactic form mirrors the semantic form, the more grammatical the utterance becomes. Jackendoff (1972: 10-11) wrote that a structurally simple grammar, like a contextfree phrase structure grammar, is preferable to a structurally complex theory, like a transformational grammar, as long as the two theories are empirically equivalent. A specification of each of the components of grammar in the lexicon and some general conditions on their interaction at the interface provides that "simpler" theory which Jackendoff was proposing. It is felt that Autolexical Syntax can provide the kind of flexibility needed to handle adverbial phenomena.

Notes *

1.

An earlier version of this paper was presented at the Workshop on Autolexical Syntax at the University of Chicago, April 1989. Many thanks to Bill Darden, Jim McCawley, Jerry Sadock, and Cyndi Welsh for comments on drafts of this paper; however, all errors and shortcomings are my own. McCawley (1983: 279) writes: T o save space, I have used carats [sic] to indicate different positions in which an adverb might occur and have given below the carat a judgement of the acceptability of the adverb in question in that position. By a comma followed by a carat, I indicate a pronunciation in which the terminal intonation contour precedes the comma and the following adverb is destressed, e.g., I have indicated that probably is acceptable at the end of a sentence if it is pronounced as a "parenthetical", but not as part of the same intonational unit as the preceding matter.

2.

3.

This I am calling the S-postfinal position. Conversely, probably is ungrammatical terminally when it and its host sentence are pronounced as a single intonational unit. This position between the main verb and its object is not eligible for consideration due to general constraints preventing anything intervening between a verb and its direct object unless the object is "heavy" (Keyser 1968). The question of S-initial ad-Vs and ad-VPs will not be discussed here; however, it has been addressed by Schneider (in prep.).

Toward a tri-modular

4.

5. 6. 7. 8. 9.

10. 11. 12. 13. 14.

analysis of-ly

adverbs

237

The semantic and syntactic information that this affix converts adjectives to adverbs would not be stated in the lexical entry for -ly. This could easily be stated in the form of a lexical redundancy rule or some other general principle in the lexicon, since this information follows naturally from the morphological rule. Some of the particulars of (24) and (25) have been borrowed from Sadock (1991) who bases his semantic program on Cresswell (1973). From Lapointe (1987a). From Sadock (1991). Note that the term "verb" here does not refer to a unified semantic category. In this work, syntactic Vs will correspond to simple predicates (intransitive verbs) and relations (transitive verbs). Willingly is even w o r s e in (32a). McCawley d o e s not discuss the distribution of ad-Ss in these sentences. They can appear in all three positions described in (32a-b); however, the central position yields a very different reading, that is, what is certain or probable is not the entire S, but only the content of the "other material". For example, in a sentence like John went running, probably in the park, what is probable is not the proposition that John went running, but where he did the running, i.e., in the park. Note that the intonation is different; the entire phrase probably in the park is occurring S-postfinally. It is proposed that in these sentences the adverb is modifying the phrase in the park semantically. The discussion of (32) and (33) is restricted to the usual interpretation of ad-Ss which pattern like the ad-VPs as expected. Whether these phrases are S- or VP-adjuncts is not important here, only that they appear VP-externally. The rule in (36) is sufficient for this paper; however, a more complex rule would be necessary to handle all cases of the interactions of adverbs with object NPs. The exclusion of position (f) has been detailed in Schneider (in prep.) and will not be discussed here. Sadock (1991) concurrently made this claim. Obviously t h e r e are discrepancies in linear order between the auxiliaries and the subject-place argument when associating the SS to LR. However, this is not a problem for the adverb specifically.

References

Cresswell, M.J. 1973

Logics and languages. London: Methuen.

Ernst, Thomas 1984 Towards an integrated theory of adverb position in English. [Ph.D. Dissertation, Indiana University, Bloomington, IN. Distributed by the Indiana University Linguistics Club.]

238

Robinson Η. Schneider

Gazdar, Gerald - Ewen Klein - Geoffrey Pullum - Ivan Sag 1985 Generalized phrase structure grammar. Oxford: Basil Blackwell. Huck, Geoffrey J. 1984 "Correspondence, compositionality, and discontinuity", in: Joseph Drogo - Veena Mishra - David Testen (eds.), Proceedings of the 20th Annual Regional Meeting of the Chicago Linguistic Society. Chicago: Chicago Linguistic Society, 203-227. Huck, Geoffrey J. - Almerindo E. Ojeda Ceds.) 1987 Syntax and Semantics, Vol. 20: Discontinuous constituency. San Diego: Academic Press. Jackendoff, Ray 1972 Semantic interpretation in generative grammar. Cambridge, MA: MIT Press. Keyser, Samuel 1968 "Review of Sven Jacobson. Adverbial positions in English", Language 44: 357-374. Lapointe, Steven 1987a "Some extensions of the Autolexical approach to structural mismatches", in: Geoffrey J. Huck - Almerindo E. Ojeda (eds.), 152-183. 1987b A three-level Autolexical analysis of deverbalized phrases. [Unpublished manuscript.] 1988 Constraints on Autolexical analysis. Lecture delivered at the University of Delaware, Newark, DE. McCawley, James D. 1981 Everything that linguists have always wanted to know about logic. Chicago: University of Chicago Press. 1983 "The syntax of some English adverbs", in: Amy Chuckerman - Mitchell Marks - John F. Richardson (eds.), Proceedings of the 19th Meeting of the Chicago Linguistic Society. Chicago: Chicago Linguistic Society, 261-282. 1988 The syntactic phenomena of English, Vol. 2. Chicago: University of Chicago Press. McConnell-Ginet, Sally 1982 "Adverbs and logical form", Language 58: 144-184. Sadock, Jerrold M. 1985 "Autolexical syntax: A proposal for the treatment of noun incorporation and similar phenomena", Natural Language and Linguistic Theory 3: 379-439. 1987 "Discontinuity in Autolexical and autosemantic syntax", in: Geoffrey J. Huck — Almerindo E. Ojeda (eds.), 283-301. 1991 Autolexical syntax: A theory of parallel grammatical representation. Chicago: University of Chicago Press.

Toward a tri-modular analysis of-ly adverbs

239

Sag, Ivan 1976

Deletion and logical form. [Dissertation, MIT, Cambridge, MA.] Schneider, Robinson in prep. An Autolexical analysis of some English adverbs. [Ph.D. dissertation, University of Chicago, Chicago.] Thomason, Richmond - Robert Stalnaker 1973 "A semantic theory of adverbs", Linguistic Inquiry 4: 195220.

Zwicky, Arnold 1986 "Immediate precedence in GPSG", in: Interfaces: Working Papers in Linguistics (Department of Linguistics, Ohio State University) 32: 133-138.

Pronominal cliticization in West Flemish* Hans Smessaert

Introduction In order to get a first idea of how syntax and morphology may interact in a particular language, let us take a look at the following example from West Flemish,1 the language under scrutiny in this paper: (1)

Kwil -de -n -t -ik -em KNOM)- want -PAST -1PSG -itCACC) -I(NOM) -him(DAT) [ook geven ]. too give Ί wanted to give it to him too.'

From the point of view of morphology, the whole string before the brackets is to be considered one word. There can be no pause or insertion between any of the morphemes. Apart from the verbal stem wil- this "word" contains six bound morphemes. These six can clearly be divided into two categories: the inflectional morphemes: -de for the past tense and -n for person and number, and the pronominal clitics: k- and -ik for the subject, -t for the direct object, and -em for the indirect object. It is this second category of pronominal clitics that will be the topic of this paper. The key property of these elements is that they do not have an independent status in the morphology-being bound morphemes they cannot function as independent words-but they do function as independent constituents in the syntax, i.e., as NPs or PPs. For the traditional sequential and hierarchical theories of grammar this poses serious problems: in this view of grammar all morphology should be handled before going to the next higher level of analysis, namely syntax. In other words, the output of the morphology component contains only independent items, i.e., words, and it is only these elements that can be the minimal input units in the syntax. Clearly, clitics do not conform to this requirement. The

242

Hans Smessaert

solution in such grammatical theories then is to create a new morphological component, exclusively for clitics, which is situated after the syntactic component. This obviously runs counter to the idea of linearity and hierarchy, and furthermore greatly complicates the theory by splitting up the morphology into a part that precedes syntax and a very small part that follows syntax. In Sadock (1985, 1991) and Farkas - Sadock (1989), however, a radically different view of grammar is presented: it is argued that-unlike what is assumed in most theories of g r a m m a r - t h e idea of autonomous modules (components) in grammar is quite independent from the concept of hierarchiality (linear sequence). The Autolexical Syntax theory presented is highly modular but non-sequential: each module of grammar (so far this is basically morphology, syntax, and semantics) operates completely independently. A lexical item is specified with respect to each of the modules and a string of the language is given an independent representation by each of the modules if it does not violate the module's rules. In a separate interface component, it is checked whether the three representations are congruous. In this framework, both inflectional and clitic morphology belong to the same module, which also includes principles of derivation and compounding. They differ from one another, though, in their interface properties: inflectional morphology, such as the tense or agreement in (1), is the default case in that the bound morphemes do not have an independent value specification in the syntax module. The clitics do have a value specification in the syntax, however, and their position in the morphology will be determined by general principles of the syntax/morphology interface. In this paper we will illustrate how this modular theory of Autolexical Syntax can be used to describe and account for the phenomenon of pronominal cliticization in West Flemish. In section 1 we discuss subject pronouns. The effects of inversion in questions are dealt with in section 2. In sections 3 and 4 we go into direct objects and the closely related reflexive pronouns. Indirect object pronouns are the topic of section 5. And finally, pronominal objects of prepositions and the related partitive constructions are considered in sections 6 and 7 respectively.

Pronominal cliticization in West Flemish

243

1. Subject pronouns In West Flemish the unstressed subject pronouns are clitic prefixes to the inflected verb. Furthermore, there is the option of adding the more emphatic form of the subject pronoun as a clitic suffix to the inflected verb. This optional suffix, which has to agree with the prefix in person, number, and gender, then may or may not receive strong stress. The full subject paradigm with both the obligatory prefix and the optional suffix then runs as in (2). Let us also consider how these facts could be accommodated within the theory of Autolexical Syntax. Beginning with the prefixes, the following lexical entry can be given for /r-in (3). (2)

K-zoun (-ik) ook kommen. Je-zou(-gij) ook kommen. Ne-zoud(-ij) ook kommen. Ze-zou(-sij) ook kommen. T-zout(-et) ookgoan.

Ί would come too.' 'You would come too.' 'He would come too.' 'She would come too.' 'It would go too.'

Me-zoun(-wijder) ook kommen.'We would come too.' Je-zou(-gijder) ook kommen. 'You would come too.' Ze-zoun(-zijder) ook kommen. 'They would come too.' (3) ksyntax = N[2,N0M/1PSG] semantics = BINDER [ Q [DEF] FORM [ PROP [ _ ] i ] ] morphology = V[-2]:AFF V[-l] The syntax specifies that k- has the function of an N[2] (the 2 refers to the bar level of the N, thus N[2] = NP). The semantics states that k- has the value of a definite quantifier plus a "domain expression" (see McCawley 1981: 118), which can be combined with an "open sentence" containing only one unbound variable, namely i, to form a full proposition. We will not go any further into the semantics of the clitics, because they all have basically the same semantic value, and because the focus of this paper is the morphology/syntax interface. The morphological entry for kindicates that it is a prefix taking an inflected verb (= V[-l]) as its host, and yielding a "super-verb" (=V[-2p when combined with that host. The interaction between syntax and morphology can be

244

Hans Smessaert

illustrated in the following tree corresponding to the first example in (2) above, but without the suffix: (4) N[2,NOM] V[1,+FIN] I kV[0,+FIN] zoun

V[1,-FIN]

Syntax

V[0,-FIN] kommen

zoun

kommen

Morphology

The subject clitic k- clearly observes the general interface constraints. First of all it conforms to the Linearity Constraint which requires that: (5) All lexemes that are projected on two dimensions must occur in the same linear order in both. (Sadock 1991: 61) In the syntax the subject N[2] precedes the finite verb, and in the morphology k- is a prefix to the inflected verb. In addition, also the Constructional Integrity Constraint is respected. The Constructional Integrity Constraint is formulated as follows: (6) If a lexeme is in construction with Ρ on dimension Dl, and is in c o n s t r u c t i o n with Q on dimension D2, then t h e correspondent to Q in Dl is a descendent of P, or the correspondent to Ρ in D2 is a descendent of Q. (Sadock 1991: 61)

If Dl is the syntax dimension and D2 is the morphology dimension then we can redraw (4) in terms of (6). This would give (7). The lexeme k- (an N[2]) is in construction with Ρ (namely the VP) on Dl (the syntax) and in construction with Q (the inflected verb) on D2 (the morphology), and therefore conforms to (6) since the correspondent to Q in Dl, namely the V[0], is dominated by P, namely the VP.

Pronominal

cliticization

in West Flemish

245

S

(7)

LEX k-

Ρ V[0]

V[1]

zoun

V[0]

D1

kommen k-

LEX

zoun

kommen

D2

Q

A lexeme which precedes the phrase it is in syntactic construction with and which attaches as a prefix to the first word of that phrase ( h e n c e o b s e r v i n g b o t h t h e Linearity Constraint and t h e Constructional Integrity Constraint) is called a "simple clitic" (Sadock 1991: 67). The subjectival suffix, such as -ik in (2), however, poses some serious problems from the point of view of syntax. What is the syntactic function of this lexeme? Do we want to posit two syntactic N[2] subjects, namely a prefix k- and a suffix -ik? Suppose for the moment that we do. Then a further question arises, namely: what is the syntactic position of the second N[2]? Two possibilities seem to suggest themselves: either we put two subject N[2]s under the subject N[2] node before the finite verb, the latter of which is then obligatorily suffixed in the morphology, or we create a new subject N[2] node after the verb for the suffix. Both of these possibilities, however, seem to complicate unnecessarily the syntax of the language. Furthermore, from the point of view of semantics only one of the N[2]s can be a "binder", because a predicate can only take one subject argument. Since the suffix is optional but the prefix not, we might argue that it is the suffix that does not have independent semantics. This might then lead us to posit that the suffix does not have any independent syntax either, which would solve the problem of where to put the extra N[2]-node, since we would no longer need one in the syntax. Then, what value specifications would the suffix get in its lexical entry? In the way we have given lexical entries so far, only the morphology would be specified:

Hans Smessaert

246

(8) -ik-

syntax = nil semantics = nil morphology = V[-2] : V[-2] AFF The morphology states that -ik- is a suffix that takes a V[-2], i.e., a super-verb that already "carries" a clitic (the subject prefix must be present), and creates another V[-2] super-verb, i.e., one carrying two clitics. Even though the suffix does not strictly speaking have a semantic value, it does have a pragmatic function, either of stress or of extra clarification. This information probably belongs to an as yet undefined pragmatics or discourse module. (9) N[2,NOM]

V[1,+FIN]

k-

V[0,+FIN] I

zoun

V[1,-FIN] Syntax I V[0,-FIN]

kommen zoun

-ik

-ik

zoun

I stressee

kommen ι

kommen

stressor

[Topic]

Morphology

Pragmatics [Comment]

Therefore, the question as to whether the suffix violates the Linearity Constraint and the Constructional Integrity Constraint cannot really be answered until such a fourth module is worked out in any detail. Let us assume for the moment, however, that in the pragmatic dimension the "stress-indicator", or "stressor" is in construction with the "stressee", i.e., the subject item to be stressed, and can either precede or follow it. Whichever order we choose the Linearity Constraint will always be violated. Remember from (5) above that the Linearity Constraint requires that all lexemes that are projected on two dimensions must occur in the same linear order in both.

Pronominal

cliticization

in West Flemish

247

This clearly does not hold here, as can be inferred from the crossed lines in the pragmatics/morphology interface of figure (9) which combines the information from the three relevant dimensions: syntax, morphology, and pragmatics. Notice that if the stressor were to precede the stressee in the pragmatics dimension there would be one more line crossing. How about the Constructional Integrity Constraint? Consider the more detailed structure of the morphology/pragmatics interface in (10). V[-2]

(10)

V[-2] Morphology AFF I k-

V[-l] I zoun

-ik

-ik

stressor

stressee

Pragmatics [Topic] In dimension Dl, i.e., the morphology, the stressor suffix -ik- is in construction with P, namely the V[-2] k-zoun. In dimension D2, i.e., the pragmatics, the stressor suffix is in construction with Q, namely the stressee prefix k-. This configuration conforms to the Constructional Integrity Constraint requirements of (6) above: the correspondent to the stressee Q in Dl, namely the subject prefix k- is a descendent of the superverb Ρ k-zoun. So we can conclude that, unlike the subject clitic which observes both the Linearity Constraint and the Constructional Integrity Constraint, the stress suffix is not a simple clitic since it violates the Linearity Constraint. This violation of an interface constraint, which causes the stress suffix to be in a sense a "delayed" clitic, is fully motivated, however, by the morphological requirements of the items involved: the stressor is obligatorily a suffix and the stressee, which it is in construction with, is itself a prefix. Hence the latter cannot function as the morphological host of the former. Therefore, the suffix must go on the first available host, which is the V[-2] superverb. In addition, we might want to

248

Hans Smessaert

say that it is an example of a "second-position clitic", in that it occurs as a suffix to the first word of a phrase (Sadock 1991: 68). Even though the above analysis of the stress suffix clitic is based on the "rules" of a pragmatics module that has not yet been fully developed, it does have some theoretical relevance for the general Autolexical framework in two important respects. First of all, the theory of clitics, which was mainly developed on the basis of the morphology/syntax interface, states that a clitic may violate at most one of the homomorphism conditions, i.e., either the Linearity Constraint or the Constructional Integrity Constraint but not both at the same time (Sadock 1991: 68). The analysis of (9) and (10) suggests that this general requirement on the behaviour of clitics can be extended to the morphology/pragmatics interface as well, since the stress suffix only violates one of the interface constraints, namely the Linearity Constraint. Secondly, it is interesting to see that the Constructional Integrity Constraint seems to assign a somewhat different status to the dimension of morphology depending on which interface it is operating on. If we take the morphology/syntax interface, as in (7) above, we see that the clitic combines with a phrase in the syntax but with a descendent of that phrase in the morphology. On the morphology/pragmatics interface, as illustrated in (10), however, the clitic combines with a phrase in the morphology but with a d e s c e n d e n t of that p h r a s e in t h e pragmatics. The latter configuration also seems to hold for the morphology/lexicosemantics interface in the case of Hungarian Preverb Climbing (Farkas - Sadock 1989: 333), where the preverb combines with a phrase in the morphology, namely a compound verb, but only with a descendent of that phrase in the lexicosemantics. In addition, the morphology/semantics interface of the verbal suffixes in Eskimo (De Reuse 1988) displays essentially the same characteristics as well. The analyses of these three different languages clearly show that, from the point of view of interface properties, the dimensions of t h e semantics (Eskimo), lexicosemantics (Hungarian), and pragmatics (West Flemish) behave very much in the same way with r e s p e c t to the morphology dimension, w h e r e a s t h e morphology/syntax interface is of a different nature. This suggests that the interface constraints in Autolexical Syntax lead to an intrinsic hierarchy of grammatical components, that is, without having to impose an external ordering on them. This hierarchy

249

Pronominal cliticization in West Flemish

(see also Farkas - Sadock 1989: 333, footnote 15) would then look as follows: (11) (lexico)semantics > morphology > syntax pragmatics

2. Inversion in questions Let us now consider what happens with the two subject clitics under inversion in questions. Take a look at the following paradigm: (12) Zoun-k(-ik) Zou-je(-gij) Zout-en(-ij) Zou-se(-zij)

ook kommen? " " ? " " ? " " ?

Zoun-me(-wijder) Zou-je(-gijder) Zoun-ze(-zijder)

ook kommen? " ? " ?

'Would 'Would 'Would 'Would

I come too you " he she

'Would we come too 'Would you " 'Would they "

?' ?'

?' ?' ?'

What we see here is that the original subject prefix has now become a suffix, whereas the pragmatic stress suffix is still a suffix. Not surprisingly, the unstressed subject affix is sensitive to the syntactic process of inversion which puts the auxiliary verb in sentence-initial position (preceding the N[2] subject) in the case of questions. Its morphological status changes according to sentence type: prefix in statements and suffix in questions. This means that we have to modify its lexical entry as given in (3) slightly, as in (13). The only change is the comma between AFF and V[-l] in the morphology, which indicates that linear order is not fixed in advance but will be determined independently by principles of the syntax. (13) -ksyntax = N[2,N0M/1PSG] semantics = BINDER [ Q [DEF] FORM [ PROP [ _ ] i ] ] morphology = V[-2] : AFF , V[-l] Now the question arises whether or not any of the interface constraints is violated when the subject N[2] is a suffix. The Linearity

Hans Smessaert

250

Constraint is clearly observed since both in the syntax and the morphology the finite verb immediately precedes the subject. In order to determine whether or not the Constructional Integrity Constraint has been violated we have to decide what exactly the syntactic tree of a yes/no interrogative looks like. Three possible structures suggest themselves, as can be seen in (14) below: b.

(14) a. AUX

VP NP

VP

AUX

NP

c. AUX

NP

VP

The difference between (14a) and (14b) for our purposes would be that in the former case the subject is not in syntactic construction with the AUX and therefore violates the Constructional Integrity Constraint, whereas in the latter case the subject and the AUX do make up a syntactic constituent and therefore conform to the Constructional Integrity Constraint. Both configurations contain an as yet unidentified syntactic node, though, and it is not clear at all that there is any syntactic evidence for the reality of these syntactic constituents. Evidence from the realm of coordination, for instance, would interact with the phenomenon of "gapping" (of the AUX) in (14a) and with the phenomenon of "right node raising" (of the VP) in (14b). I will therefore adopt a more neutral position and take an "inverted" sentence (in this case a yes/no interrogative) to have the "flat" structure of (14c), a position also adopted within the Head-driven Phrase Structure Grammar framework (Pollard Sag 1987: 156-157). This structure can be given in more detail as: _S[INV]

(15) V[0,AUX,+FIN] I zoun

N[2,NOM] I -k-

V[1,-FIN] I V[0,-FIN] kommen

Pronominal cliticization in West Flemish

251

We might argue that in the syntax N[2] is in construction both with the finite V[0] and the non-finite V[ 1 ]. Therefore, the fact that the affix goes as a suffix on the preceding finite verb does conform to the Constructional Integrity Constraint as formulated in (6). In addition it is a "second-position clitic" in that it occurs as a suffix to the first word of the phrase that it is syntactically associated with (Sadock 1991: 68). So we can conclude that both in statements and questions the subject clitic observes both the Linearity Constraint and the Constructional Integrity Constraint, and is thus a "simple clitic" (Sadock 1991: 67). Remember that we saw in connection with (9) and (10) in the previous section that the pragmatic stress-suffix in statements conforms to the Constructional Integrity Constraint but violates the Linearity Constraint. This violation was reflected in the crossed lines of the morphology/pragmatics interface. Now, consider what happens in questions:

(16) V[0,AUX,+FIN] zoun

N[2,NOM] I

zoun ι zoun

V[1,-FIN] I V[0,-FIN] Syntax I kommen kommen *

stressee [Topic]

o m

Morphology

L Pragmatics

[Comment]

Clearly, in the case of questions, the stress-suffix observes the Linearity Constraint: all the lexemes that occur both in the morphology and the pragmatics dimension occur in the same linear order in both, so w e get no crossed lines in the morphology/pragmatics interface a n y m o r e . As for the Constructional Integrity Constraint, exactly the same situation as the one described in connection with (10) holds here: in the morphology the stress suffix is in construction with a V[-2], which we might call a "morphological phrase", and in the pragmatics it is

252

Hans Smessaert

in construction with the correspondent of a "daughter" of that phrase. Hence we may conclude that in questions (but not in statements) the stress suffix is a "simple clitic" since both the Linearity Constraint and the Constructional Integrity Constraint are observed. Notice, as well, that we do not have to modify its original lexical entry at all (remember that we did have to make some changes in the lexical entry of the subject-affix). It is here repeated as: (17) -iksyntax = nil semantics = nil morphology = V[-2] : V[-2] AFF Both in questions and in statements, the lexeme is a suffix with a superverb as its morphological host. The only difference then lies in the internal structure of that host (the subject-affix can be a prefix or a suffix), but the stress-suffix is not sensitive to this difference.

3. Direct object pronouns In West Flemish the direct object pronouns, which may or may not be stressed, are clitic suffixes to a [V-2] superverb. Consider the full direct object paradigm in (18). The lexical entry for the direct object clitic suffix would be as in (19). The direct object is an NP in syntax and a binder in semantics. Morphologically, it is a suffix that takes a V[-2] superverb as its host. (18) N-e-mijn N-e-joun N-ed-em N-e-neur N-ed-et N-e-nuus N-e-junder N-e-nunder N-e-ze

ezien. ezien. ezien. ezien. ezien.

'He 'He 'He 'He 'He

has has has has has

seen seen seen seen seen

me.' you.' him.' her.' it.'

ezien. ezien. ezien. ezien.

'He 'He 'He 'He

has has has has

seen seen seen seen

us.' you.' them.' them [neuter].'

Pronominal cliticization in West Flemish

253

(19) -mijn

syntax = N[2,ACC/1PSG] semantics = BINDER [ Q [DEF] FORM [ PROP [ _ ] i ] ] morphology = V[-2] : V[-2] AFF In order to see how examples such as (18) behave with respect to the interface constraints we should first notice that the word order within the non-finite V[l] in West Flemish is with the non-finite V[0] in final position, as we can infer from: (20) Jan

et

[ die

vint gisteren

up de mart

ezien].

John has [ that man yesterday on the market seen ] 'John saw that man at the market yesterday.' Therefore, the syntactic structure of the first sentence in (18) looks as follows: (21)

S N[2,NOM] I n-

V[1,+FIN] V[0,+FIN]

V[ 1 ,-FIN]

s

e N[2,ACC] -mijn [

n-

ι

-mijn

V[0,-FIN] I ezien

ezien ]

From the above tree we can infer that the direct object suffix -mijn does not violate the Linearity Constraint: all the lexemes that occur both in the syntax and the morphology occur in the same linear order in both dimensions. The Constructional Integrity Constraint, however, is violated: syntactically the direct object is in construction with the non-finite main verb ezien, but morphologically it combines with the finite auxiliary, with which it is not in syntactic construction. So we see that the object clitic differs from the subject clitic in that the latter observes both the Linearity Constraint and the Constructional Integrity Constraint (and is therefore a "simple clitic"), whereas the former clearly violates the Constructional Integrity Constraint. As was the case

254

Hans

Smessaert

with the stress suffix in (10) of section 1, however, the violation of one of the interface constraints, in this c a s e the Constructional Integrity Constraint, is fully motivated by the morphological requirements of the direct object clitic: it has to be a suffix. A clitic of this particular type is called an "anticipatory clitic" (Sadock 1991: 67) since it agrees with what follows, i.e., it determines the characteristics of an item that follows, in this case the transitivity of the main verb. One may have noticed that the stress suffix in (17) and the d i r e c t o b j e c t clitic in (19) have identical m o r p h o l o g i c a l specifications, here repeated as: (22) morphology = V[-2] : V[-2] AFF The obvious question now arises: which of the two goes first? Let us therefore take a look at the following paradigm in which both are present as suffixes on the inflected verb: (23) N-ed-ij-mijn N-ed-ij-joun N-ed-ij-em N-ed-ij-neur N-ed-et-ij N-ed-ij-nuus N-ed-ij-junder N-ed-ij-nunder N-e -z -ij

ezien. ezien. ezien. ezien. ezien.

'He 'He 'He 'He 'He

has has has has has

seen seen seen seen seen

me.' you.' him.' her.' it.'

ezien. ezien. ezien. ezien.

'He 'He 'He 'He

has has has has

seen seen seen seen

us.' you.' them.' them[neuter].'

What we s e e is that the stress suffix precedes the direct object clitic except when the latter is third person neuter/inanimate (i.e., the "prototypical object"). Now, how could we formally express this difference in linear order? Let us assume that if two clitics with the same morphological requirements, in this case those of (22), occur simultaneously, they form a "clitic cluster" (CL) which is itself a morphological constituent and which as a whole still takes a V[-2] as its morphological host. The combination of the V[-2] host and the clictic cluster then gives us a V[-3]. To get this result we have to posit the following rules: (24) a. V[-3] : V[-2] CL b. CL : ( AFF ) n

Pronominal

cliticization

in West Flemish

255

It is important to keep in mind, though, that the subject clitic is not a part of the clitic cluster. First of all it does not have the morphological specification of (22), i.e., its host is not a V[-2] superverb but rather an inflected V[-l]. In addition, it changes its morphological status from prefix to suffix under inversion, whereas all the clitics that go into the clitic cluster are obligatorily suffixes. The next thing to do, then, is to specify the linear order within the clitic cluster. For the sentences in (23) the following two linear precedence rules would suffice: (25) Morphological linear precedence (MLP) rules: MLP 1: [ACC/NEUT] < [STRESS] MLP 2: [STRESS] < [ACC] In order for these morphological linear precedence rules to achieve the intended effect, the original lexical entry of the stress suffix in (17) above would have to be modified so as to include the feature [STRESS]. In addition, the result of combining a clitic with the V[-2] gives us a V[-3] according to (24a). Furthermore, we somehow want to guarantee that the stress suffix and the subject affix agree in person, number, and gender. Since agreement is at least to some extent a matter of syntax, and stress is somehow related to semantics in a general sense, the lexical entry of a stress suffix (in the examples of (23) this is the third person singular masculine) will henceforth look as follows: (26) -ij-

syntax = [3PSG/MASC] semantics = [STRESS] morphology = V[-3] : V[-2] AFF As far as the direct object clitics are concerned, we only have to make sure that reference to the gender feature [NEUT] is included in the lexical entry: 2 (27) -(e)t syntax = N[2,ACC/3PSG/NEUT] morphology = V[-3] : V[-2] AFF -ze

syntax = N[2,ACC/3PPL/NEUT] morphology = V[-3] : V[-2] AFF

256

Hans Smessaert

In the following two examples the linear order within the clitic cluster is determined by Morphological Linear Precedence Rule 1 and 2 respectively: (28) a. N-ed-et-ij b. N-ed-ij-mijn

ezien. ezien.

'He has seen it.' 'He has seen me.'

Let us now take a look at the trees for (28a) and (28b) in order to see which, if any, of the interface constraints are violated when both the stress suffix and the object clitic occur. As for the neuter object clitic -et in (28a) we have to look at the syntax/morphology interface. Figure (29) below shows that the Linearity Constraint is observed since all lexemes that are present both in the syntax and in the morphology occur in the same linear order in both dimensions. (29)

S N[2,NOM] I n-

V[1,+FIN] V[0,+FIN]

V[1,-FIN]

ed N[2,ACC] I et

Syntax

V[0,-FIN] I ezien IJ

ezien

-et

ezien

Morphology

Pragmatics [Topic] As we noted in connection with (21) above, however, the Constructional Integrity Constraint is violated: syntactically the direct object is in construction with the non-finite main verb but morphologically it combines with the finite auxiliary, with which it is not in syntactic construction. As for the stress suffix -ij, we again see that the Linearity Constraint in the morphology/pragmatics interface is violated: the

Pronominal cliticization in West Flemish

257

crossed lines indicate that not all lexemes occur in the same linear order in both dimensions. Remember from our discussion of (10) in section 1, however, that the Constructional Integrity Constraint is observed: in the morphology the stress suffix combines with a superverb, which has the value of a "phrase", and in the pragmatics it combines with the stressee, which corresponds to a "descendent" of the superverb, namely the subject clitic (29) and (30). (30)

S N[2,NOM]

V[1,+FIN] V[0,+FIN] I ed

n-

n-

ed

V[1,-FIN] N[2,ACC]

-ij

-ij

stressee

stressor

ed

V[0,-FIN1 I

mijn

ezien

-mijn

ezien

X I

n-

-mijn

Syntax

Morphology ezien

Pragmatics

[Topic] From figure (30), which gives the structure of (28b), we can infer that the non-neuter direct object clitic -mijn does not really differ from the neuter one with respect to the Linearity Constraint and the Constructional Integrity Constraint. As in (29), the direct object clitic in (30) observes the Linearity Constraint but violates the Constructional Integrity Constraint. Also for the stress suffix in (30) the same observations hold as for (29), i.e., the Constructional Integrity Constraint is observed but the Linearity Constraint is violated. Again, the fact that the stress suffix is in a sense a "delayed clitic" (in the case of [29] even further than in that of [30]) can clearly be accounted for in terms of its morphological requirements: in both cases it has to be a suffix to a V[-2] and in (29) it has to observe the first morphological linear precedence rule of (25), whereas in (30) it has to observe the second morphological linear

258

Hans Smessaert

precedence rule of (25). Notice as well that the general restriction on clitics still holds in (29) and (30): even though the direct object clitic violates the Constructional Integrity Constraint and the stress suffix violates the Linearity Constraint, neither of them violates both at the same time. As for the morphology/pragmatics interface the crucial difference between (29) and (30) lies in the number of "crossed lines": in (29) there are two crossed lines, whereas in (30) there is only one. What the exact theoretical relevance of this difference in the number of crossed lines could be is not really obvious for the moment. Nevertheless, the pattern with only one crossed line is much more common than the one which has a double crossing: the latter is restricted to the third person neuter direct objects. This might suggest a tendency in natural language to keep the number of crossed lines as low as possible.

4. Reflexive pronouns An interesting property of the West Flemish pronominal system is that reflexive pronoun clitics have exactly the same form as the direct object clitics, as we can infer from the following paradigm: K-en-ik-mijn J-e-gij-joun N-ed-ij-em Z-e-sij-neur

ewoschen. ewoschen. ewoschen. ewoschen.

Ί washed myself.' 'You washed yourself.' 'He washed himself.' 'She washed herself.'

M-en-wijd er-n u us ewoschen. ewoschen. J-e-gijder-j under Z-en-zijder-n under ewoschen.

'We washed ourselves.' 'You washed yourselves.' T h e y washed themselves.'

K-en-ik-mijn J-e-gij-joun N-ed-ij-em Z-e-sij-neur

Ί hurried up.' 'You hurried up.' 'He hurried up.' 'She hurried up.'

goast. goast. goast. goast.

M-en-wijd er-n u us goast. J-e-gijder-j un d er goast. Z-en-zijder-n under goast.

'We hurried up.' 'You hurried up.' 'They hurried up.'

Pronominal

cliticization

in West

Flemish

259

Notice as well that the three affixes, namely the subject prefix, the stress suffix, and the reflexive suffix, all have to agree in person, number, and gender. This matter of agreement, however, is not really of crucial importance for the issues under discussion. I will therefore simply assume that the theory provides an agreement rule that takes care of this. In order to determine the lexical entry of the reflexive pronoun, let us first recall that of the direct object clitic, here repeated as: (32) -mijn

syntax = N[2,ACC/1PSG] semantics = BINDER [ Q [DEF] FORM [ PROP [ _ ] i ] ] morphology = V[-3] : V[-2] AFF The reflexive pronoun clearly has the same morphological specification: it takes a V[-2] as its host and is therefore a part of the clitic cluster. And indeed, the second morphological linear precedence rule of (25) in section 3 correctly predicts the linear order within the clitic cluster: 3 the reflexive has to follow the stress suffix: (33) MLP 2 : [STRESS] < [ACC] There are some serious problems, however, with the semantic specification of (32) since the reflexive pronoun does not have the semantic value of a binder. In the case of the true reflexive verb em woschen 'to wash oneself in (31) the reflexive pronoun has the value of a second semantic variable bound by the definite quantifier of the subject, whereas in the case of the so-called "reflexiva tanta" such as em oasten 'to hurry up' the reflexive pronoun has no semantic value whatsoever. We will therefore have to modify the lexical entry (32) in the following way: (34) a.

-mijn

syntax = N[2,ACC/REFL/1PSG] semantics = i morphology = V[-3] : V[-2] AFF

b.

-mijn

syntax = N[2,ACC/REFL/1PSG] semantics = nil morphology = V[-3] : V[-2] AFF

260

Hans

Smessaert

Reflexive verbs will then subcategorize for either of these types of pronouns. Notice, as well, that the morphological linear precedence rule of (33) still applies to (34). I will not go any further into the semantics of reflexive verbs (such as the exact nature of the binding conditions that the variable i in (34a) is subject to), except to point out that "reflexiva tanta", i.e., verbs taking a reflexive pronoun like (34b), can still take a "real" direct object, in which case they give rise to very interesting combinations within the clitic cluster. Consider for instance sentences with the verb em etwot voernstellen 'to imagine (oneself) something': (35) a. K-an-t-ik-mijn anders voernesteld. I-had-it-I-myself differently imagined Ί had imagined it differently.' b. N-ad-et-ij-em anders voernesteld. He-had-it-he-himself differently imagined 'He had imagined it differently.' The real direct object is neuter and must therefore, according to the first morphological linear precedence rule, precede the stress suffix, whereas the reflexive is non-neuter and must therefore, according to the second morphological linear precedence rule, follow the stress suffix. Examples like these suggest that we can really collapse the two morphological linear precedence rules of (25) above into one: (36) Morphological linear precedence rule: [ACC/NEUT] < [STRESS] < [ACC] We may now ask what happens if the real object of a "reflexivum tantum" is non-neuter, i.e., when we get two non-neuter accusative pronouns in the clitic cluster. Not surprisingly, the situation is not very clear and neither order is perfect according to my intuitions: (37) K-an-ik-mijn-em anders voernesteld. I-had-I-myself-him differently imagined Ί had imagined him differently.' ? K-an-ik-em-mijn anders voernesteld. ? I-had-I-him-myself differently imagined Ί had imagined him differently.'

Pronominal cliticization in West Flemish (38) ? M-an-wijder-nuus-em

anders

261

voernesteld.

? We-had-we-ourselves-him differently imagined 'We had imagined him differently.' M-an-wijder-em-nuus

anders

voernesteld.

We-had-we-him-ourselves differently imagined 'We had imagined him differently.' Since cases like these are really exceptional I will not venture to posit any morphological linear precedence rule for the linear order between two non-neuter accusative pronouns, and simply stick to the unified version of (36) above. Let us instead take a look at the interfaces of an example such as (35a) in which the three suffixes in the clitic cluster occur in the linear order predicted by the morphological linear precedence rule of (36), in order to determine whether or not any of the interface constraints are violated. As far as the stress suffix and the neuter object in figure (39) are concerned, exactly the same observations hold as for (29) in section 3: the stress suffix observes the Constructional Integrity Constraint, but violates the Linearity Constraint, whereas the direct object clitic observes the Linearity Constraint but violates the Constructional Integrity Constraint. (39) N[2,N0M] I le-

an N[2,ACC] N[2,ACC] ADV[2] V[0,-FIN] I I -mijn

voernesteld

I

I

I

-mijn

anders

voernesteld

anders

voernesteld

I

-mijn

Topic

anders

262

Hans Smessaert

Again, the fact that the stress suffix is in a s e n s e a "delayed clitic" can clearly be accounted for in terms of its morphological requirements: it has to be a suffix to a V[-2] and it has to observe the first part of the morphological linear precedence rule of (36). The reflexive pronoun behaves in exactly the s a m e way as the neuter object: it observes the Linearity Constraint but violates the Constructional Integrity Constraint since in the syntax it is in construction with the non-finite main verb but in the morphology it combines with the finite auxiliary. Notice a s well that the general restriction on clitics still holds in (39): even though the two object clitics violate the Constructional Integrity Constraint and the stress suffix violates the Linearity Constraint, none of them violates both at the s a m e time.

5. Indirect object pronouns In c a s e s where there are two "real" object clitics, that is, where both pronominal objects have the semantic value of an argument, the unmarked configuration is with a neuter/inanimate direct object and a non-neuter/animate indirect object. In West Flemish this can be expressed in the following way: (40) N-ed-et-(ij)-mijn N-ed-et-(ij)-joun N-ed-et-(ij)-em N-ed-et-(ij)-neur N-ed-et-(ij)-nuus N-ed-et-(ij)-j under N-ed-et-(ij)-nunder

egeven. egeven. egeven. egeven.

'He 'He 'He 'He

has has has has

given given given given

it it it it

to to to to

me.' you.' him.' her.'

egeven. egeven. egeven.

'He has given it to us.' 'He has given it to you.' 'He has given it to them.'

Alternatively, it is also possible for the indirect object clitic not to be part of the clitic cluster on the superverb but instead to take the preposition a(n) 'to' as its morphological host (41). In this c a s e the whole dative PP usually follows the main verb, and tends to be more emphatic than the dative clitic on the verb in (40) especially when the stress suffix on the subject is absent. Let us also consider the lexical entries for the indirect object clitics. In order to account for both (40) and (41) we have to specify two possible morphological hosts as in (42).

Pronominal cliticization in West Flemish (41)

(42)

N-ed-et-(ij) N-ed-et-(ij) N-ed-et-(ij) N-ed-et-(ij)

egeven egeven egeven egeven

a-mijn. a-joun. an-em. a-neur.

N-ed-et-(ij) N-ed-et-(ij) N-ed-et-(ij)

egeven a-nuus. egeven a-junder. egeven a-n under.

'He 'He 'He 'He

has has has has

given given given given

it to it to it to it to

263

me.' you.' him.' her.'

'He has given it to us.' 'He has given it to you.' 'He has given it to them.'

-mijn syntax = N[2,0BL/1PSG] morphology = V[-3] : V[-2] AFF Ρ[-2] : P[-l, "a(n)"] AFF

First of all the indirect object clitic can take a V[-2] as its morphological host, and is therefore part of the clitic cluster. And indeed the unified morphological linear precedence rule formulated in (36) predicts the correct linear order within the clitic cluster, on the assumption that we drop the distinction between [ACC] and [DAT] in favour of the cover-feature [OBL]. The generalized morphological linear precedence rule would then look like: (43) Morphological linear precedence rule: [OBL/NEUT] < [STRESS] < [OBL] As it stands this rule states that any non-neuter, non-nominative pronoun in the clitic cluster has to follow the stress suffix. This holds for real direct objects, reflexives, and indirect objects, and therefore the feature [OBL] nicely captures these linear order generalizations. The corresponding change to the cover-feature [OBL/NEUT] in the first part of the rule captures a similar generalization, as will be shown in section 6 where we discuss pronominal objects of prepositions. Remember from our discussion of transitive reflexiva tanta in (37) and (38) of section 4 that it is possible to get two [OBL] forms next to one another. The same holds here, with somewhat odd semantics though, as can be seen in (44) below. Again the linear order among them is not so clear, with a preference, though, for interpreting the first [OBL] form as the indirect object:

264

Hans Smessaert

(44) N-ed-ij-mijn-nunder etoogd. He-has-he-me-them shown 'He has shown them to me.' ? 'He has shown me to them.' By virtue of the second morphological specification in (42), the [OBL] suffix can also take the preposition a(n) as its host and turn it into a super-preposition. It is this extra morphological property that will then distinguish the indirect object clitic from the other, direct [OBL] objects. Notice that this mechanism not only holds for the "pure" indirect objects with the preposition aCn) but can also be extended to expressions of "goal" with the preposition no(r) 'to', or of "beneficiary" with the preposition voe(n) 'for': (45) K-en-t-ik-em ezonden. I-have-it-I-him sent K-en-t-ik ezonden nor-em. I-have-it-I sent to-him Ί have sent it to him.' ? K-en-t-ik-em ekocht. ? I-have-it-I-him bought K-en-t-ik ekocht voen-em. I-have-it-I bought for-him Ί have bought it for him.' Now, how about the interface constraints in the case of indirect object clitics? In sentences like (40), the Constructional Integrity Constraint is violated, as can be seen in figure (46). 4 (46)

S V[1,+FIN]

N[2,NOM] n-

V[0,+FIN] ed

V[1,-FIN]

N[2,OBL] -et

ri-

ed

-et

-ij

Syntax

N[2,OBL]

V[0,-FIN]

-mijn

egeven

-mijn

egöven Morphology

Pronominal cliticization in West Flemish

265

The indirect object takes the finite auxiliary as its morphological host whereas syntactically it is in construction with the non-finite main verb. The Linearity Constraint is not violated since all lexemes that occur both in the syntax and the morphology occur in the same linear order in both dimensions. When we turn to the examples in (41), where the indirect object clitic takes the preposition as its morphological host, we see that both the Constructional Integrity Constraint and the Linearity Constraint are met: both in the syntax and in the morphology the N[2] clitic is in construction with the preposition, and all lexemes occur in the same linear order in both dimensions. Therefore, the indirect object clitic is a "simple clitic" (Sadock 1991: 67) if it takes the preposition as its host: (47)

S N[2,NOM] I n-

V[1,+FIN] V[0,+FIN] I ed N[2,ACC]

V[ 1 ,-FIN] V[0,-FIN] egeven

P[2,DAT] P [ l ] N[2,DAT] -mijn

ed

\

-ij

egeven

I -mijn

6. Pronominal objects of prepositions Consider the paradigm in (48) where a non-neuter pronoun is the object of a preposition. In these cases the pronoun is a suffix clitic with the preposition as its morphological host. The morphological specification of such a non-neuter object of a preposition differs from that of the indirect object in two ways. First t h e r e is no reference to a particular preposition functioning as the host (for the indirect object this had to be a[n]). Secondly, this clitic does not have the alternative of taking a V[-2] as its host, and is therefore not a part of the clitic cluster (49).

266

Hans Smessaert

(48) N-ed-ij

N-ed-ij

an-mijn an-joun an-em an-neur

epeisd.

an-nuus epeisd. an-junder an-nunder

'He has thought of me.' 'of you.' 'of him.' 'of her.' 'He has thought of us.' 'of you.' 'of them.'

(49) -mijn syntax = N[2,0BL/1PSG] morphology = P[-2] : P[-l] AFF It may have become clear that the neuter pronoun occupies a special position in the paradigm. Remember that West Flemish does not have a neuter reflexive pronoun, nor a neuter indirect object pronoun. When the neuter form does occur as a direct object we have seen that it has to preceed the stress suffix, whereas all the non-neuter forms follow the stress suffix. In the case of neuter objects of prepositions the form (d)er shows up instead of the regular -f. The former is obviously related to Standard Dutch er, a detailed analysis of which is presented in Leys (1979) and Swiggers — Van den Eynde (1985). As in Standard Dutch, West Flemish (d)er can occur in two different positions: (50) a. ? N-ed-ij ol der-an epeisd. ? He-has-he already there-of thought 'He has already thought of it.' b. N-et-er-ij ol an epeisd. He-has-there-he already of thought 'He has already thought of it.' First of all, we see in (50a) that the pronoun can take the preposition as its host. Notice that in this case the clitic is a prefix instead of a suffix.5 Much more common, however, is the form in (50b), where the neuter pronoun (d)er again has the familiar morphological property of taking a V[-2] superverb as its host. (51) (d)er syntax = N[2,OBL/3PSG/NEUT] morphology = V[-3] : V[-2] AFF Ρ[-2] : AFF P[-l]

Pronominal cliticization in West Flemish

267

Therefore it is again part of the clitic cluster. The lexical entry of this neuter pronoun then looks as in (51). In those cases where the pronoun takes a P[-l] preposition as its morphological host the Constructional Integrity Constraint is clearly observed since both in the syntax and in the morphology the pronoun is in construction with the preposition. As for the Linearity Constraint we have to distinguish between cases such as (49) where the pronoun is a suffix, and those such as (51) where it is a prefix. In the former case the Linearity Constraint is observed whereas in the latter case it is violated. This can be seen in (52) which gives the relevant part of the respective structures. Again, however, the violation of the Linearity Constraint in (52b) is fully motivated by the morphological requirement of the neuter pronoun that it be a prefix to a preposition. As we have seen so many times before, morphology once more overrides syntax. (52) a.

P[2]

b.

P[2]

P[T]

N[2]

P[l]

I an

I.

I an

I an

-mi^n

der-

N[2] I deran

-mijn Let us now turn to those cases in which the neuter pronoun takes the V[-2] as its morphological host. The first half of our morphological linear precedence rule, which is simplified in (53) by reducing [OBL/NEUT] to [NEUT], indeed predicts the right linear order for the example in (50b), i.e., the neuter pronoun has to precede the stress suffix: (53) Morphological linear precedence rule: [NEUT] < [STRESS] < [OBL] One may now wonder whether it is still possible to get a non-neuter oblique pronoun following the stress suffix. Remember that the feature [OBL] covers three cases: the direct object, the indirect object and the reflexive pronoun. And indeed, all three are possible in combination with a neuter prepositional object and a stress suffix, and the linear order within the clitic cluster is exactly as predicted by (53). Respective examples are:

268

Hans Smessaert

(54) a.

K-en-der-ik-em ol van overtuugd. I-have-there-I-him already of convinced Ί have already convinced him of it.' b. K-en-der-ik-em ol van eklapt. I-have-there-I-him already about talked Ί have already talked to him about it.' c. K-en-der-ik-mijn ol mee emoeid. I-have-there-I-myself already with interfered Ί have already interfered with it.'

From the point of view of the interface constraints, however, the behaviour of the neuter pronoun when it is part of the clitic cluster is really problematic. Consider the following figure corresponding to (54a): (55)

S N[2,N0M] k-

yiJj+FINJ V[0,+FIN] I en N[2,OBL]

V[1 ,-FIN] ADV

P[2]

V[0,-FIN]

P [ l ] N[2,0BL] ol k-

en

I van

I dervan

overtuugd ι overtuugd

Restricting our attention to what happens to the neuter prepositional object in (55), we see first of all that it manifestly violates the Linearity Constraint. Remember that this was also the case in (52b). Unlike in (52b), however, the Constructional Integrity Constraint is also violated: the pronoun is in syntactic construction with the P [ l ] to form a P[2], whereas in the morphology it is a suffix to a V[-2], i.e., the finite auxiliary. The fact that this clitic violates both the Linearity Constraint and the Constructional Integrity Constraint at the same time is a very serious problem for the theory of clitics presented in the Autolexical framework, according to which a clitic may not violate both interface constraints at the same time. Nevertheless, these violations can again be accounted for in terms of the morphological requirements of the

Pronominal

cliticization

in West Flemish

269

neuter pronoun form: it has to go on a V[-2], and within the clitic cluster it has to obey the morphological linear precedence rule which puts it in first position. In the following paragraphs I will further illustrate some of the particular phenomena that relate to this neuter form (d)er-. Remember, first of all, from our discussion of (37) and (38) in section 4 and (44) of section 5 that it is possible to get two non-neuter [OBL] forms next to one another. The same holds for two [OBL/NEUT] forms, where one is the direct object, and the other an object of a preposition both of whose referents are clear from the discourse, as in: (56) K-en-t-er-ik

ol

up

I-have-it-there-I already on Ί have already put it on it.'

ezet.

put

In cases like these the direct object clitic (e)t has to precede the prepositional object (d)er-, and consequently has to precede any other member of the clitic cluster, so that we can formulate the following additional morphological linear precedence rule: (57) Morphological linear precedence rule 2: [ (e)t ] < X

S

(58)

V[1,+FIN]

N[2,NOM] k-

en

k-

V[1,-FIN]

V[0,+FIN]

en

N[2,OBL]

ADV P[2]

V[0,-FIN]

ezet

A possible explanation for the relative order between (e)t and (d)er- might be the following: (d)er- violates both the Linearity Constraint and the Constructional Integrity Constraint anyway, as

270

Hans Smessaert

can be seen in (58), but the violation would be even worse if it were to precede (e)t, i.e., if we did not have the second morphological linear precedence rule of (57). Another remarkable fact about this neuter pronoun (d)er- is that with verbs whose prepositional object is necessarily (or at least preferably) non-neuter, the neuter form has the non-neuter reference, as in: (59) K-en-der-ik ol mee eklapt. I-have-there-I already with talked Ί have already talked to him/her/?? it.' With this type of verb it is also possible to get examples similar to (56) where there is a neuter direct object as well, and which must therefore conform to the second morphological linear precedence rule given in (57), i.e., the neuter direct object must precede the neuter prepositional object. So in addition to (60) with the regular non-neuter clitics following the stress suffix, we also get examples like (61) where both objects precede the stress suffix: (60) a. K-en-t-ik-em ol I-have-it-I-him already Ί have already given it to b. K-en-t-ik-neur ol I-have-it-I-her already Ί have already given it to (61) K-en-t-er-ik ol I-have-it-there-I already Ί have already given it to

egeven. given him.' egeven. given her.' an egeven. to given him/her/?? it.'

Notice that in (61) the preposition an which marks the indirect object, has to be added. Hence, the distinction between (60) and (61) corresponds to the difference in English between to give someone something and to give something TO someone where the "goal" is a simple NP in the former case, but a prepositional object in the latter. Thirdly, it should be noted that the neuter form (d)er- also has a deictic or emphatic counterpart, namely (d)or- (corresponding to the distinction between the pronoun it and the demonstrative that). As was the case for the non-deictic form in (50a) above, the deictic form can also be a prefix to a preposition:

Pronominal cliticization in West Flemish

271

(62) ΊΚ-en-ik ol dor-an epeisd. ? I-have-I already there-of thought Ί have already thought of that.' Much more common, though, are sentences where the (d)or- goes as a suffix on the V[-2], and is therefore a part of the clitic cluster. If it does, however, it differs radically from its non-deictic counterpart in that it has to follow the stress suffix. Compare, for instance: (63) a. K-en-der-ik ol I-have-there-I already Ί have already thought b. K-en-ik-dor ol I-have-I-there already Ί have already thought

an epeisd. of thought of it.' an epeisd. of thought of that.'

Now, what happens if we get a non-neuter object clitic, i.e., one bearing the feature [OBL], in addition to a deictic prepositional object? Remember from (54) above that [OBL] covers three types, namely direct object, indirect object, and reflexive pronoun. In all three cases the deictic has to follow the [OBL]. The respective emphatic counterparts of (54) are: (64) a. K-en-ik-em-dor ol van overtuugd. I-have-I-him-there already of convinced Ί have already convinced him of that.' b. K-en-ik-em-dor ol van eklapt. 1-have-I-him-there already about talked Ί have already talked to him about that.' c. K-en-ik-mijn-dor ol mee emoeid. I-have-I-myself-there already with interfered Ί have already interfered with that.' Let us now consider the lexical entry for the deictic neuter prepositional object. Remember that we introduced the "pragmatic" feature [STRESS] into the semantic specification of the stress suffix. I would like to do something similar here. The difference between (d)er- and Cd)or- does not lie in the syntax nor in the "pure" semantics. In addition, the morphological requirements are basically the same as well: it is either a prefix with a P[-l ] host, or a suffix with a V[-2] host. The only difference lies in the "pragmatic"

Hans Smessaert

272

realm of deixis. Since so far we have not really included an independent pragmatics specification in the lexical entries, I will introduce the feature [DEIXIS] in the semantic specification of (d)or-. Its complete lexical entry will look like: ( 6 5 ) (d)or-

syntax = N[2,OBL/3PSG/NEUT] semantics = BINDER [ Q [DEF] FORM [ PROP [ _ ] i ] ] [DEIXIS] morphology = V[-3] : V[-2] AFF Ρ[-2] : AFF P[-l] Remember that the pragmatic feature of [STRESS] was of crucial importance in determining the linear order of the morphemes within the clitic cluster. Exactly the same holds for the pragmatic feature of [DEIXIS] since it has to come last in the clitic cluster. We can therefore make our first morphological linear precedence rule even more general by adding this feature at the end: (66) Morphological linear precedence rule 1: [NEUT] < [STRESS] < [OBL] < [DEIXIS] Notice that the crucial difference between [STRESS] and [DEIXIS] is that items with the former feature do not have independent "real" semantics, whereas items with the latter feature do have the value of a binder. As it stands, (66) does not exclude the possibility of getting all four morphemes within the clitic cluster, and this is clearly right. Although a normal main verb can hardly take three non-subject arguments, which would correspond to the [NEUT], the [OBL], and the [DEIXIS], perception verbs and causative verbs do allow all three at the same time. The [OBL] is then the subject of the infinitival complement clause and the [NEUT], and the [DEIXIS] are the objects of the infinitive. In these cases the [STRESS] refers to the main clause subject: (67) a.

K-en-t-ik-em-dor

ol

zien

up

zetten.

I-have-it-I-him-that already see up put. Ί have already seen him put it on "that".' b.

K-en-t-ik-em-dor

ol

doen/loaten

up

zetten.

I-have-it-I-him-that already make/let up put. Ί have already made/let him put it on "that".'

Pronominal cliticization in West Flemish

273

Any possible combination of the four "features" that conforms to the required linear order among them is well formed. 6 In many cases a combination will even have more than one realisation, since the feature [NEUT] can be realised both as the direct object (e)t and the prepositional object (d)er. In addition [OBL] can be realised as the direct object, the indirect object, or the reflexive. Which of these is realised in any particular case is exclusively determined by the syntactic and semantic properties of the main verb, i.e., its subcategorization frame. Notice, however, that some restrictions hold on the possible combinations. 7 One is that there cannot be two prepositional objects in the clitic cluster, i.e., if the [DEIXIS] is present, the [NEUT], if present, has to be the direct object (e)t. Other restrictions are the agreement requirements between the subject affix and the [STRESS], or between the [STRESS] and the [OBL] if the latter is a reflexive, for which we have simply assumed the theory to provide the necessary agreement mechanism. Let us now take a look at the syntax/morphology interface of examples such as (67a) where we get all four clitics: (68)

N[2,NOM]

V[1,+FIN] V[0,+FIN]

V[1,-FIN] S[-FIN]"

N[2,0BL] k-

en -em

V[1,-FIN]

N[2,OBL] ADV I -t

ol en -t

V[0,-FIN]

-ik

-em

P[2] P[l] I up

-dor

V[0,-FIN] N[2,0BL] I

-dor zetten zien

ol

zien up

zetten

One thing to be noticed is the violation of the Linearity Constraint between the two verbs at the end of the sentence. This phenomenon has become known as cross-serial dependency and holds in a very similar way for Standard Dutch and West Flemish. This phenomenon

274

Hans Smessaert

also interacts with that of preposition reanalysis which accounts for the possible alternative ordering up zien zetten at the end. These problems, however, are not really relevant to our discussion and I will therefore not go into them any further. More important is the behaviour of the emphatic prepositional object do(r). As was the case for its non-emphatic counterpart in (55) and (58), both the Linearity Constraint and the Constructional Integrity Constraint are violated. Nevertheless, these violations can again be accounted for in terms of its morphological requirements: it has to go on a V[-2], and within the clitic cluster it has to obey the morphological linear precedence rule which puts it in final position. Notice, incidentally, that this final position again corresponds in a sense to its "final" position in the syntax. In other words, if do(r) did not come at the end of the clitic cluster the Linearity Constraint would be violated even more. In these particular infinitival complement constructions an additional problem arises with the relationship between the [OBL] subject of the infinitive zetten, namely -em, and the [OBL] direct object of that same infinitive, namely -t. As we can infer from (68), both the subject and the object violate the Constructional Integrity Constraint since they are in morphological construction with the finite auxiliary, whereas in the syntax the subject combines with the non-finite V[l] and the object with the non-finite V[0]. In addition the crossed lines indicate that there is a violation of the Linearity Constraint as well. It is not so easy, however, to decide which of the two items is in the wrong place. In view of the fact that the [NEUT] prepositional objects -der and -dor are the only lexical items so far that violate both the Linearity Constraint and the Constructional Integrity Constraint at the same time, we will assume for the sake of uniformity that also in the case of (68) it is the [NEUT] direct object which is out of place. 8 As with the two prepositional objects, the violation of both interface constraints by the [NEUT] direct object -t is due to its morphological requirements: it takes a V[-2] as its host, and within the clitic cluster it has to obey the second morphological linear precedence rule which puts it right at the beginning of the clitic cluster. Constructions like the one in (68) once more illustrate the fact that morphological requirements cannot be overruled even though this may lead to violation of both of the interface constraints at the same time.

Pronominal cliticization in West Flemish

275

7. Quantified pronominal direct objects As in Standard Dutch9 the neuter pronoun (d)er also shows up in a number of other configurations which involve direct objects rather than prepositional objects. Consider for instance the difference between (69a) and (69b), both of which have a direct and an indirect object: (69) a.

K-en-t-ik-em

ol

egeven.

I-have-it-I-him already given Ί have already given it to him.' b.

K-en-der-ik-em

ol

egeven.

I-have-there-I-him already given Ί have already given him some (of them).' In both cases the [NEUT] is clearly the direct object, but its meaning in (69b) corresponds roughly to what traditional grammar calls a "partitive" genitive, i.e., it indicates the set of items (whose reference is clear in the discourse), a part of which something is predicated of in the sentence. In (69b) very little information is given as to which set is involved or which part of that set is relevant. It is possible, however, to be somewhat more specific by adding a quantificational premodifier, i.e., a numeral or quantifier, as in: (70) K-en-der-ik-em

ol

twee/vele

egeven.

I-have-there-I-him already two/many given Ί have already given him two/many (of them).' The idea is that der pronominalizes definite or indefinite NPs that occur in a partitive relationship (Leys 1979: 242). When it is not further specified quantificationally, as in (69a), the indefinite reading such as (some) books is strongly preferred, whereas in (70) both the definite and the indefinite reading are possible. Hence, the combination der .. twee in (70) could mean 'two books' as well as 'two of the books'. If we want to focus on the definite reading, however, we can add the preposition van (='of) to the quantificational element: (71) K-en-der-ik-em

ol

twee/vele

van

I-have-there-I-him already two/many of Ί have already given him two of them.'

egeven.

given

Hans Smessaert

276

What the exact syntactic structure is of the string twee (van) in (70-71) is not very clear. One could argue that the first part is a determiner whose head noun is absent, and that the second part is the preposition introducing the PP postmodifier whose NP is absent too. Such a structure is shown in (72). At first sight this seems to be plausible, since there are cases corresponding to (72) where the head noun of the N[2] is not absent, and which still allow the (d)er to occur in the clitic cluster (73). There are some serious problems with this analysis, though. First of all, the Cd)er in (73) no longer has the partitive meaning, i.e., its referent is no longer a set but simply neuter singular. (72)

N[2] DET I twee

N[l] N[0] I

0

P[2] P[l]

N[2] I

I (van) (73) K-en-der-ik-em ol twi fotos van I-have-there-I-him already two pictures of Ί have already given him two pictures of it.'

0

egeven. given

Secondly, (73) only allows the "weak" (i.e., unstressed) form of the numeral twi. In other words, the "strong" (i.e., stressed) form twee is excluded. Exactly the opposite holds for (70) and (71) where the weak form is impossible. Thirdly, the example in (73) has a counterpart in which the object of the preposition is emphatic, i.e., has the feature [DEIXIS], whereas (70) and (71) do not have this counterpart (74).10 This evidence clearly suggests that (72) is not the correct structure for the direct object NPs in (70) and (71). K-en-ik-em-dor ol twi fotos van I-have-I-him-there already two pictures of Ί have already given him two pictures of that.' b. * K-en-ik-em-dor ol twee egeven. * I-have-I-him-there already two given Ί have already given him two (of those?).'

(74) a.

egeven. given

Pronominal cliticization in West Flemish

277

c. * K-en-ik-em-dor ol twee van egeven. * I-have-I-him-there already two of given Ί have already given him two of (those?).' The alternative I would like to suggest considers the strong numeral to be the head "noun" which is postmodified by a partitive PP whose preposition van may or may not be overtly realised. The structure of such an NP would look like: (75)

N[2] I N[l] N[0] twSe

P[2] P[l] I (van)

N[2] I (d)er

This view is also supported by the behaviour of an item such as e poar 'a few' which may refer to a subset of a particular set, in which case it has basically the same properties as the strong numeral: as in (70-71) it can take the partitive (d)er, but, as in (74b), it does not take the emphatic form: (76) a. K-en-der-ik-em ol e poar (van) egeven. I-have-there-I-him already a few (of) given Ί have already given him a few of (them).' b. * K-en-ik-em-dor ol e poar egeven. * I-have-I-him-there already a few given Ί have already given him a few (of those?).' It is interesting to see that the example corresponding to (74c) is possible, but then e poar looses its broader subset-meaning of 'a few', and has to get its original, stricter meaning of 'a pair': (77) K-en-ik-em-dor ol e poar van I-have-I-him-there already a pair of Ί have already given him a pair of (those).'

egeven. given

An analysis along the lines of (75) would then capture the difference between the weak and the strong numerals,11 as well as between

278

Hans

Smessaert

the two meanings of e poar. In addition it would enable us to distinguish two types of (d)er. one with a partitive meaning as in (70) and (71), which does not have the emphatic counterpart and postmodifies a numeral (or related) head noun, and another one with the neuter singular meaning as in (73), which does have an emphatic counterpart and is part of a direct object with a "regular" head noun. Remember, however, that both types of (d)er have identical morphological properties: they take a V[-2] as their host, and are [NEUT], i.e., they go in the first slot of the clitic cluster. Furthermore, these morphological requirements cause the violation of both the Linearity Constraint and the Constructional Integrity Constraint for both types of (d)er. The only difference between the figure in (55) of section 6 and the partitive construction in (78) below, corresponding to (71), is that the P[2] is not a daughter of the V [ l ] but is embedded in an N[2] which is the daughter of the V[ 1 ]. Perhaps one might even argue that in the c a s e of partitive der there is a stronger violation of the Linearity Constraint than in that of the prepositional der. S

(78) N[2,NOM]

V[1,+FIN] V[0,+FIN]

V[ 1,-FIN]

N[2,OBL]

N[2,OBL]

V[0,-FIN]

N[(

k-

en

-em

k-

en -ik -der

twee

van

der- egeven

-em

twee van

egeven

Pronominal cliticization in West Flemish

279

8. Conclusion In this paper I have tried to show that the Autolexical framework provides the adequate tools to account for the various phenomena involving pronominal cliticization in West Flemish. The most important feature of our analysis is the radical separation between immediate dominance relations and linear precedence relations. The immediate dominance relations are specified for each particular clitic in the form of a "morphological phrase structure rule" in the morphological part of its lexical entry. The linear precedence relations, on the other hand, are determined by a general morphological linear precedence rule which is to a large extent independent of the particular clitics involved. Within Autolexical Syntax two general constraints have been formulated that limit the degree of mismatching in the interface between two modules: one is the Linearity Constraint and the other is the Constructional Integrity Constraint. Mainly on the basis of work on the syntax/morphology interface, an additional principle had been formulated according to which a lexeme may violate at most one of the interface constraints, i.e., either the Linearity Constraint or the Constructional Integrity Constraint but never both at the same time (Sadock 1991: 68). From the above discussion of the West Flemish pronominal system, where most of the action takes place in the syntax/morphology interface as well, it has become clear that this is too strong a claim. In West Flemish at least three instances can be found of clitics that violate both the Linearity Constraint and the Constructional Integrity Constraint at the same time. First there is the [NEUTER] prepositional object (d)er (= 'it'), which may also be used in a partitive construction. Secondly, we get the [DEIXIS] counterpart of the prepositional (d)er, namely do(r) (= 'that'). And finally, there is the [NEUT] form -t when it functions as the direct object of the infinitival complement of perception verbs and causatives. These three clitics are brought together in the following (somewhat far-fetched) example in which they all violate both the Linearity Constraint and the Constructional Integrity Constraint at the same time: (79) K-en-t-er-ik-em-dor twee zien loaten up I-have-I-there-I-him-there two see let up Ί have seen him let two of them put it on that.'

zetten. put

280

Hans Smessaert

S[0,+FIN]

(80)

Ν 2,NOM]

V[1,+FIN]

V[0,+FIN]

V[1,-FIN] S[0,-FIN]

N[2,C)BL]

V[07FIN]

"vfl^FIN] S[0,-FIN]

N[2,OBL] N[0]

k-

en en

-em twee

Ρ

V[0,-FIN]

V[1,-FIN]

2] N[2,OBL] ΡΓ2]

-er

-t

P[0] I up

-er -ik -em -dor twee

V[0,-FIN]

N[2] I

-dor zetten loaten zien zien

loaten up

zetten

In the light of this evidence, the general principle that lexemes cannot violate both interface constraints at the same time has to be relaxed. It undoubtedly remains a very strong tendency, but some items, such as the ones in (79) and (80), may override it. It is important to notice that the "double" violation of the items in (80) is clearly due to their morphological requirements: they have to go as a suffix on the finite V[-2] and the linear order among them within the clitic cluster is fixed by the Morphological Linear Precedence Rule. The final version of this rule, which has been elaborated and refined step by step in the course of our analysis, is here repeated as: (81) Morphological Linear Precedence Rule: [-1] < [NEUT] < [STRESS] < [OBL] < [DEIXIS] The "power" of this morphological template illustrates once again that in the case of tensions or mismatches between morphological and syntactic requirements it is the morphology that wins out. In addition, the discussion of the [STRESS] and [DEIXIS] clitics has

Pronominal cliticization in West Flemish

281

urged us to consider the development of a separate pragmatics module for Autolexical Syntax. An important difference between the two pragmatic features is that [DEIXIS] is a property of a lexical entry which has an independent syntactic and semantic specification, whereas the [STRESS] clitics are not specified for syntax or semantics. The latter are nevertheless truly productive clitics. As far as the pragmatics/morphology interface is concerned, we have shown that the Constructional Integrity Constraint principle is also applicable. It is interesting to notice that with these stress clitics in West Flemish, and in similar cases in Hungarian and Eskimo, the dimension of morphology seems to acquire a somewhat different status. In the morphology/syntax interface the "phrasal units" for the Constructional Integrity Constraint are in the syntax, whereas the "simple items" are in the morphology. If the second dimension in the interface is not the syntax, but rather the pragmatics or the (lexico)-semantics, however, then the "phrasal units" for the Constructional Integrity Constraint are to be found in the morphology whereas the "simple items" are in the pragmatics (or the lexico-semantics). This difference in "status" depending on which interface is involved may very well be a more generell tendency across languages and may suggest some sort of an intrinsic hierarchical ordering of the different modules. As far as the applicability of the Linearity Constraint to the pragmatics/morphology interface is concerned, though, it should be noticed that it is much more difficult to talk about matters of linear order in the pragmatics. Hence any claims made in this connection should be dealt with very carefully. In the second part of this conclusion I would like to abstract away from the particular theoretical framework of Autolexical Syntax and consider the various West Flemish clitics in terms of the general constellation of properties that recent grammatical literature has claimed "typical" clitics to display. In doing this, I will follow the classification of properties discussed by Sadock (1991: 52) who distinguishes five dimensions, namely morphology, syntax, semantics, phonology, and lexicon. The West Flemish clitics are: (82) a.

"syntactico-semantic": 1. [NOM] : -k 2. [NEUT]: -t/-(d)er 3. [OBL] : -em

282

Hans Smessaert b.

"semantico-pragmatic": 1. [STRESS]: -ik 2. [DEIXIS] : -do(r)

Consider the morphological properties of clitics first: (83) Morphology: a. clitics are bound morphemes; b. they are completely productive; c. they attach outside inflection; d. they block further morphology; e. they attach without regard to the morphological class of the host. As to (83a-c) there can be no doubt at all: the West Flemish clitics are completely productive bound morphemes which attach outside of inflection. With (83d), however, w e get into trouble since the existence of the clitic cluster in West Flemish clearly shows that most West Flemish clitics do not block further morphology. In fact, only one does, namely the [DEIXIS] clitic since it is the last one in the clitic cluster. As for (83e), all the examples in this paper have taken a finite verb or an auxiliary as the morphological host of the clitics. In chapters 4 and 5 of Smessaert (1988) and in Smessaert (1995), however, it is shown how the West Flemish clitic cluster can also take a complementizer as its morphological host, and how the [NOM] subject clitic even allows a third type of host, namely a polarity particle. Turning to syntax next, we find the following two properties: (84) Syntax: a. clitics are independent elements of syntax; b. they are syntactically adjacent to their morphological host. The only West Flemish clitic listed in (82) for which the property in (84a) does not hold is the [STRESS] suffix. All the others do indeed have an independent syntactic function, such as subject, direct or indirect object, or object of a preposition. As far as (84b) is concerned, only one clitic is always syntactically adjacent to its morphological host, namely the [NOM] subject. If the other clitics occur in the clitic cluster, then they may, but need not, be syntactically adjacent to their morphological host, depending on the latter's class. If they are outside of the clitic cluster, which

Pronominal cliticization in West Flemish

283

may be the case for the [NEUT] prepositional object (d)er and its [DEIXIS] counterpart do(r), then they are syntactically adjacent to their morphological host, namely the preposition, although they come in the wrong linear order. The following semantic properties have been attributed to typical clitics: (85) Semantics: a. clitics are semantic functions; b. they take the meaning of a phrase as their argument. In the case of the [STRESS] suffix neither of these two properties holds since the item d o e s not have any semantic value in t h e strict sense. One might argue, however, that it is a pragmatic function, namely of "stressing", that takes the meaning of a phrase, namely the subject, as its argument. If we look at the clitics that do have a real semantic value we see that they are not semantic functions, but rather semantic arguments. As a consequence they cannot have the property of (85b) either. Let us now turn to t h e phonological properties that clitics typically display: (86) Phonology: a. clitics are affixes; b. they are agglutinative; c. they are subject to automatic rules only; d. they are stressless. As far as the first three properties are concerned, they clearly hold for all the West Flemish clitics. Not all West Flemish clitics are stressless as suggested by (86d), however. First of all t h e [STRESS] suffix may or may not receive primary stress. In addition, the [DEIXIS] prepositional object do(r)- may receive heavy stress, in which case it is phonetically realised with a long vowel. The final class of clitic properties belongs to the lexicon: (87) Lexicon: a. "host + clitic" forms are not lexicalized; b. clitics alternate with free words. Clearly, none of the West Flemish "host + clitic" combinations would be listed in the lexicon. Even though we did not mention t h e m explicitly, t h e r e are two West Flemish clitics that have sentence-initial free form counterparts, namely the [NOM] subject

Hans Smessaert

284

clitic and the [OBL] object clitic. With the latter, however, the free form is only possible in the direct object function, and not in its reflexive or indirect object function. In addition, one might argue that the free form of the [NOM] subject is actually the free form of the [STRESS] clitic, since they are identical. (88) [PROP] [INFL] [NOM] [NEUT] [NEUT] [STRESS] [OBL] [DEIXIS]

-n

-k

-t

-er

-ik

-em

(83a)

+

+

+

+

+

+

+

(83b)

+

+

+

+

+

+

+

(83c)

-

+

+

+

+

+

+

(83d)

-

-

-

-

-

-

+

(83e)

3

3

2

2

2

2

2

(84a)

-

+

+

+

-

+

+

-

+

?

-/+

?

?

-

-

-

-

?

-

-

-

-

-

-

?

-

-

(86a)

+

+

+

+

+

+

+

(86b)

+

+

+

+

+

+

+

(86c)

+

+

+

+

+

+

+

(86d)

+

+

+

+

-/+

+

-/+

(87a)

+

+

+

+

+

+

+

(87b)

_

+

_

_

?

(84b) (85a) (85b)

-dor

-

Pronominal cliticization in West Flemish

285

All the clitic properties discussed above are brought together in (88), where we have also included the verbal inflectional ending for the sake of comparison between clitic properties and inflectional properties. From the general survey in (88) we can infer that the West Flemish clitics differ among one another with respect to how many of the typical clitic properties they display. Clearly, there seems to be some sort of a hierarchy: [NOM] and [OBL] come closest to the ideal, next come the two [NEUT] clitics and the [DEIXIS], and then comes the pragmatic [STRESS] suffix. We should not overlook the fact, however, that this list of properties is biased towards clitics that operate on the syntax/morphology interface, which is maybe the most common one, but certainly not the only one, as our discussion of the West Flemish pragmatic clitics has illustrated. As a final remark, I would like to point out that this detailed analysis of pronominal cliticization in West Flemish is also very likely to shed some light on what is going on in Standard Dutch. Even though Standard Dutch does not have the pragmatic clitics, it clearly shows a strong tendency towards cliticization with those unstressed personal pronouns that do have syntactic and semantic value. It would therefore surely be worthwile to investigate the possibility of describing those phenomena in Standard Dutch in terms of a morphological template or a morphological linear precedence rule similar (if not identical) to the one that accounts for the West Flemish data. Appendix: Sixteen logical combinations of Morphological Linear Precedence Rule 1: [NEUT] < -/+ (1)

[

]

[STRESS] -b

< [OBL] < [DEIXIS] -b

K-en olgeten. Ί have already eaten.'

(2) [ — + ]

K-en-dor ol an epeisd. Ί have already thought of that.'

(3)

K-en-em ol ezien. Ί have already seen him.'

[--+-]

-b

286

Hans Smessaert

(4) [ - - + + ]

K-en-em-dor ol van eklapt. Ί have already talked about that to him.'

(5) [- + - - ]

K-en-ik ol geten. Ί have already eaten.'

(6) [ - + - + ]

K-en-ik-dor ol an epeisd.

Ί have already thought of that.'

(7) [-+ + -] (8) [ - + + + ]

K-en-ik-em ol ezien.

Ί have already seen him.'

K-en-ik-em-dor ol van eklapt. Ί have already talked about it to him.'

(9) [ + — ] (10) [ + - - + ] (11) [ + - + - ] (12)[+- + + ]

(13) [ + + - - ] (14)[ + + - + ] (15)[ + + + - ] (16) [ + + + + ]

K-en-t ol ezien.

Ί have already seen it.'

K-en-t-dor ol up ezet.

Ί have already put it upon that.'

K-en-t-em ol ezeid.

Ί have already told it to him.'

K-en-t-em-dor ol zien up zetten. Ί have already seen him put it upon that.'

K-en-t-ik ol ezien.

Ί have already seen it.'

K-en-t-ik-dor ol up ezet.

Ί have already put it upon that.'

K-en-t-ik-em ol ezeid.

Ί have already told it to him.'

K-en-t-ik-em-dor ol zien up zetten.

Ί have already seen him put it upon that.'

Pronominal cliticization in West Flemish

Notes *

1.

2.

3.

4.

5.

6.

7. 8.

This paper is a revised version of the first three chapters of Smessaert CI988). I would like to thank Jerry Sadock and Willem De Reuse for their comments on the latter document. West Flemish is a Dutch dialect spoken in the province of West Flanders in Belgium. I have basically used Standard Dutch spelling throughout. West Flemish pronunciation deviates from Standard Dutch pronunciation in the following ways : =[u], =[i], = [y], =[D:]. In addition : =[a], =[e], and =[e:]. Word-final is reduced to a syllabic nasal [n]. (See also Sercu 1972.) I will henceforth omit the semantic specification in the lexical entries unless they are relevant to our discussion, as is the case for the stress suffix. T h e first morphological linear p r e c e d e n c e rule is irrelevant here since West Flemish, unlike French or English, does not have a neuter reflexive pronoun. I will henceforth omit the morphology/pragmatics interface since the behaviour of the stress suffix has been dealt with in sufficient detail and is less important in further discussion. The same holds for Standard Dutch (Swiggers - Van den Eynde 1985: 71) where ungrammatical combinations of a preposition plus the neuter pronoun het have to be replaced with er-Ρ, as in: *Ik denk aan het ==> Ik denk er-aan Ί think of it'. In Leys (1979: 243-246) it is furthermore stressed that this er- fits into a full paradigm with a manifestly locative, adverbial origin : hier-, daar-, waar-, ergens-, nergens-, overal-, which literally mean 'here, there, where, somewhere, nowhere, everywhere', but which due to a functional shift get the nonlocative meaning of 'this, that, what, something, nothing, everything' when they combine with a preposition or adverb. In the Appendix I give one example of each of the sixteen logical possibilities that the morphological linear precedence rule in (66) allows. For a detailed description of the co-occurence restrictions of different types of er in Standard Dutch see Swiggers — Van den Eynde (1985: 75-81). An additional argument for treating the interface behaviour of the direct object -f on a par with that of the prepositional object der in these cases comes from a comparison with Standard Dutch. In Standard Dutch the form er also occurs in sentence initial position of two particular constructions (Leys 1979: 244; Swiggers - Van den Eynde 1985: 73-75), namely :

287

288

Hans Smessaert [a]

[b]

Er There 'There Er There There

Staat een man op straat. stands a man on street is a man standing in the street.' werd veel gedanst. became much danced was a lot of dancing [being danced].'

In [a] the indefinite subject comes after the verb, and in [b] passivization has removed the original subject. In both cases the sentence-initial subject position is filled by a so-called dummy or expletive er. The interesting thing now is that in West Flemish the -t form may occur in this position : [a'] [b'] 9. 10.

11.

T-stoat e vint up stroate. It-stands a man on street T-wier vele gedanst. It-became much danced

This type of er in Standard Dutch is discussed in Leys C1979: 242-243) and Swiggers - Van den Eynde (1985: 71-72). The same thing holds for Standard Dutch (Leys 1979: 243) where in contrast to the prepositional er, the partitive er is not a member of the so-called "locative paradigm" mentioned in footnote 6. This distinction is phonologically marked in many cases where the strong form has an additional schwa at the end: eg. drie versus drie-e (3), vier versus vier-e (4), vuuf versus vuuv-e (5), zes versus zes-e (6), acht versus acht-e (8), tien versus tien-e (10), ef versus ev-e (11), twoaf versus twoav-e (12). It is not, however, in een (1), zeven (7) and negen (9), and in all forms ending in jgsuch as twintig (20), dertig (30), and so on.

References De Reuse, Willem 1988 "The morphology/semantics interface: An Autolexical treatment of Eskimo verbal affix order", in Lynn MacLeod Gary Larson - Diane Brentari (eds.) Papers from the 24th Annual Regional Meeting of the Chicago Linguistic Society, Part One: The General Session. Chicago: Chicago Linguistic Society, 112-125. Farkas, Donka F. - Jerrold M. Sadock. 1989 "Preverb climbing in Hungarian", Language 65: 318-338.

Pronominal cliticization in West Flemish Leys, Odo 1979

289

"De bepaling van het voornaamwoordelijk bijwoord en de systematisering van ndl. er [The determination of the pronominal adverb and the systematization of Dutch er]", De Nieuwe Taalgids 72: 240-246. McCawley, James D. 1981 Everything that linguists have always wanted to know about logic but were ashamed to ask. Chicago: University of Chicago Press. Pollard, Carl - Ivan A. Sag. 1987 Information-based syntax and semantics. Vol. 1: Fundamentals. CSLI Lecture Notes, 13. Stanford: The Center for the Study of Language and Information. Sadock, Jerrold M. 1985 "Autolexical syntax: a proposal for the treatment of noun incorporation and similar phenomena", Natural Language and Linguistic Theory 3: 379-439. 1991 Autolexical syntax: A Theory of parallel grammatiical representation. Chicago: University of Chicago Press. Sercu, Aurel 1972 Het dialekt van Oostduinkerke en omgeving [ The dialect of Oostduinkerke and the adjacent area], Gent: Koninklijke Vlaamse Academie voor Taal- en Letterkunde. Smessaert, Hans 1988 An Auto-lexical syntax approach to pronominal cliticization in West Flemish. [Unpublished M.A. paper, Department of Linguistics, University of Chicago.] 1995 Morfo-syntaxis van het Westvlaamse "be-jaa-k-gie" [Morphosyntax of the West Flemish polarity particle expression "bejaa-k-gie"]. [To appear in : Tabu 25/1.] SWiggers, Pierre — Karel Van Den Eynde. 1985 "Distributie- en kombinatiemogelijkheden van Nederlands 'er': een Studie in syntaktische klassifikatie [Distribution and combinatorial properties of Dutch er. a study of syntactic classification]", in: Marc Dominicy - Hugo Baetens Beardsmore (eds.), Linguistics in Belgium 7. Brussels: Vrije Universiteit Brüssel, 67-86.

Two types of mismatch between morphology and semantics Gregory T. Stump

One of the basic tasks which must be fulfilled by an Autolexical theory of grammar is that of specifying the permissible associations between morphological structures and semantic representations. In the simplest cases, these associations are just given; that is, the lexical listing of a word simply stipulates that its morphological representation is linked to its semantic representation. But the associations at the interface between morphology and semantics won't all necessarily be this simple; consider, for instance, the possibility of pursuing the hypothesis in (1). (1) Partial listing hypothesis: Not all of the words in a given paradigm are necessarily listed in the lexicon; in particular, a word belonging to a given paradigm may be omitted if both its morphological structure and its semantic representation are uniquely determined by some other member(s) of the same paradigm. If this hypothesis is taken to be valid, then there are at least some w o r d s w h o s e morphological s t r u c t u r e s and semantic representations are not simply given in the lexicon. Once this assumption is made, two important questions arise. First, what kinds of associations exist between lexically unlisted morphological structures and the corresponding semantic representations? And second, what sorts of principles regulate these associations? At least two distinct ways of answering these questions are imaginable. One possibility is the structure-based approach in (2). This principle is predicated on the assumption that a lexically unlisted word can be exhaustively analyzed into terminal elements which are themselves lexically listed, and that it is the individual semantic associations into which these parts enter that determine the kind of semantic representation with which the unlisted whole may be paired.

292

Gregory Τ. Stump

(2) Structure-based approach to morphology/semantics associations: The association of a lexically unlisted morphological structure Μ with its semantic representation S is at the level of terminal elements, and is subject to the requirement that every meaningful terminal element of Μ must be associated with a terminal element of S, and vice versa. In principle, such an approach seems perfectly feasible; but a very different alternative to this approach is also conceivable. Following Lapointe (1987), one might adopt the restriction in (3): (3) Maximal word restriction: Associations between morphological structures and semantic representations are at the maximal word level. Taken together with the partial listing hypothesis, the maximal word restriction entails that if a morphological structure is lexically unlisted, then its association with a particular semantic representation can neither be directly stipulated in the lexicon, nor be reduced to the semantic associations of its terminal elements; thus, this restriction excludes the structure-based approach in (2) as a possible means of regulating semantic associations involving lexically unlisted morphological structures, and makes it necessary to assume that in the case of lexically unlisted words, associations between morphology and semantics must be predicted by rules or principles of some sort. Among the possible ways of doing this is the paradigm-based approach in (4). (4) Paradigm-based approach to morphology/semantics associations: The association of a lexically unlisted morphological structure Μ with its semantic representation is stated as a function of M's membership and position within a particular paradigm. On this approach, there is no expectation that the morphological structure and the semantic representation of a lexically unlisted word can necessarily be factored into matching terminal elements; rather, the link between a word's structure and its semantic representation is here assumed to be determined by its relatedness to one or more words in the same paradigm.

Two types of mismatch

293

In this paper, I wish to examine the relative merits of the structure-based and paradigm-based views of the associations between morphology and semantics. My discussion will focus on two very different types of mismatches between morphological structure and semantic representation: the first of these is a structural mismatch exhibited by certain plural nouns in the Breton language; the second is a mismatch not of structure but of content exemplified by the system of nominal derivatives in Tigre. My claim is that the most satisfactory Autolexical analysis of these mismatches is one which exploits the possibilities afforded by the paradigm-based approach in (4) rather than the structure-based approach in (2). More specifically, I shall argue for a novel conception of the morphological component of a grammar which permits the paradigm-based approach to be cleanly implemented. What is distinctive about this new approach to morphology is the assumption that there are actually two kinds of morphological rules-morpholexical rules and paradigm rules—which have very different properties. The details of this two-way rule distinction will be discussed shortly; first, however, it will be useful to consider one of the kinds of mismatch which this distinction can help resolve.

1. Plurals of diminutives and loose compounds in Breton1 This first type of mismatch between morphology and semantics is clearly illustrated by the plural forms of diminutive nouns in Breton. Before turning to the problems presented by this formation, consider first the usual morphology and semantics of plural nouns and singular diminutives. Breton nouns are ordinarily pluralized through simple suffixation, as in (5a); (5) a. bag + -ού boat PLURAL b. bag + -ig boat DIMINUTIVE

= =

bagou boats bagig little boat

in addition, the meaning of a plural form can ordinarily be regarded as a function of the meaning of the corresponding singular; for instance, if bag 'boat' has the semantic representation in (6a), then the corresponding plural form bagou can be assumed to have

294

Gregory Τ. Stump

the representation in (6b). (For concreteness, the semantic function Fpl in (6b) may be identified with the {*)' operator proposed by Link (1983), the definition of which is given in (7a); nothing in the following discussion hinges on the assumption of this particular definition, however.) (6) a. Semantics of bag 'boat': b. Semantics of bagoü 'boats': c. Semantics of bagig 'little boat':

bag' Fpl(bag') Fdim(bag')

(7) Semantics of Fpland Fdim: a. Fpl may be identified with Link's ' © ' operator: where Ρ denotes a set S of atomic individuals, © P denotes the set of non-atomic individual sums of members of S (Link 1983). b. Semantics of Fdim: identity function. Pragmatics of Fdim: where Ρ denotes a set S of individuals and χ denotes an individual i, Fdim(P)(x) conventionally implicates that the speaker's attitude towards i is as if i were a diminutive member of S. Singular diminutives are, by and large, similarly uncomplicated in their morphology and semantics. The diminutive form of a noun is made through the addition of the suffix -ig, as in (5b). This is an extremely productive process, and one whose semantic consequences are quite regular, though not entirely so. In the following discussion, I shall assume that the usual semantic counterpart of diminutivization is a function Fdim, which might be defined as in (7b).2 On this assumption, the singular diminutive bagig 'little boat' has the semantic analysis in (6c). So far, no problem; as long as linear order is assumed to be irrelevant in semantic representations (cf. Sadock 1987), then the semantic representations in (6) are in all relevant respects isomorphic to the corresponding morphological structures in (5). The problem arises once plural diminutives come into view; some examples of these are given in (8). Brief scrutiny of these examples reveals the key fact, namely that, given a singular diminutive consisting of a nominal base suffixed with -ig, the corresponding plural diminutive consists of the plural form of the nominal base followed by the diminutive suffix -ig, which is itself followed by the default plural suffix -oil

Two types of mismatch

295

(8) Breton diminutives: a. b. c. d. e. f. g· h. i. j·

singular bag pedenn tra labous merc'h potr bugel maen ti askorn

a. b. c. d. e. f. g· h. i. j·

singular diminutive bagig pedennig traig labousig merc'hig potrig bugel ig maenig tiig askornig

'boat' 'prayer' 'thing' 'bird' 'girl' 'boy' 'child' 'stone' 'house' 'bone'

plural bagou pedennou traoü laboused merc'hed potred bugale mein tiez eskern

'little 'little 'little 'little 'little 'little 'little 'little 'little 'little

plural diminutive bagoüigoü pedennouigou traouigou labousedigou merc'hedigoü potredigoü bugaligou meinigoü tiezigoü eskernigoü

boat' prayer' thing' bird' girl' boy' child' stone' house' bone'

Thus, the plural form of the diminutive noun meaning 'little boat' in (8a) has (9) as its morphological structure. (9)

Ν

bag

-ού

-ig

-ού

('little boats')

Clearly plural diminutives are much more complicated in their structure than the ordinary plurals considered earlier; yet, this structural difference is not accompanied by any clear semantic difference. That is, the semantic relation between a singular

296

Gregory Τ. Stump

diminutive and its plural is exactly the same as the one between an ordinary singular and its plural; in both cases, the meaning of the plural form is apparently a function of the meaning of the singular form. This can be seen most clearly in examples such as those in (10). (10)

singular base

a. tok b. moc'h

plural

'hat'

tokou

'swine' (collective)

-

singular diminutive 'robin' a. tokig b. moc'hig 'sand-flea'

plural diminutive

tokedigou moc'hedigoü

In each of these examples, the singular diminutive is not simply interpreted as the diminutive of its nominal base, but has instead taken on some special, idiosyncratic piece of meaning.3 The point is that in each case, the plural form of the diminutive shares whatever semantic idiosyncrasy happens to be associated with the singular form of the diminutive; the plural form tokedigou in (10a), for example, means 'robins', and not 'little hats'. Thus, here and generally, the meaning of a singular diminutive is a proper part of that of the corresponding plural diminutive, even though the singular form is not itself a proper part of the plural form. For this reason, the semantic analysis of a plural diminutive like bagouigoiι 'little boats' is parallel to that of an ordinary plural noun such as bagoü 'boats', as in (11). (11) a. Semantics of bagoii 'boats': F p l (bag') b. Semantics of bagouigou 'little boats': F pl (bagig') [= F pI (F dim (bag'))] In view of these facts, it is clear that a mismatch exists between the morphological structure of a plural diminutive and its semantic structure, as the example in (12) shows. On the assumption that association lines cannot cross, this mismatch must somehow be resolved or avoided. Before proceeding, I should point out-without going into details-that there is a class of compounds in Breton whose plurals exhibit a systematic mismatch similar to the one in (12); these are the so-called loose compounds, exemplified in (13).4

Two types of mismatch

(12)

297



('little boats')

Ν (13) Loose compounds: singular

a. loa-vansoner

'trowel'

plural

loaio ύ-manson er

spoon-mason b. milin-zour 'water mill' mill-water

milinoü-dour

c. pod-soubenn 'cooking pot'

podoü-soubenn

pot-soup

d. tok-kolo

'straw hat'

tokou-kolo

'lighthouse'

tourioü-tan

hat-straw

e. tour-tan tower-fire

T h e plural form of a l o o s e compound, like a plural diminutive, does not contain the corresponding singular f o r m as a p r o p e r part. Yet, the meaning of the plural form of l o o s e c o m p o u n d s must contain that of the corresponding singular form, since any semantic idiosyncrasy inherent in the singular f o r m invariably shows up in the plural form as well. Given these facts, it is clear that the plural form of a l o o s e compound exhibits a structural mismatch parallel to the one observed in plural diminutives; thus, alongside the mismatch in (12) is the one in (14). N o w , the mismatches in (12) and (14) could be instantly resolved b y direct lexical stipulation; that is, one could simply assume that both the morphological structures and the semantic representations in (12) and (14) are listed in the lexical entries of bagigand tok-kolo.

298

Gregory Τ. Stump

kolo

('straw hats') But as was noted above, diminutive morphology is extremely productive in Breton; moreover, given the form and the meaning of a singular diminutive, both the form and the meaning of the corresponding plural diminutive are virtually always uniquely determined. Similar remarks apply to the loose compounds, which remain the most highly productive formation for nominal compounds in Modern Breton. Since this is so, it is plausible to assume that in accordance with the partial listing hypothesis, the plural form of a diminutive or loose compound will not, in general, be listed lexically. But if it isn't, then how can the association b e t w e e n its morphological s t r u c t u r e and its semantic representation be accounted for?

2. Structure-based accounts of the Breton mismatches The structure-based approach to predicting associations between morphology and semantics is not obviously applicable here, since the terminal nodes of the morphological structures and the semantic structures in (12) and (14) cannot be associated without violating the general ban on line crossing. There are, however, two ways in which a structure-based account of the associations in these examples might be devised. For instance, one could assume that at the word level, inflectional affixes have no semantic content (cf. Sadock 1988). If such were the case, then the plural suffixes in (12) and (14) would not have to match up with anything at the

Two types of mismatch

299

semantic level, and would thus fail to give rise to the indicated mismatches. Note that this approach would not have to commit one to the view that inflection as a whole is semantically empty, but only to the view that inflectional affixes fail to make any semantic contribution at the word level. That is, this approach would be perfectly consistent with the possibility that an inflectionally realized feature might be semantically relevant when it appears at a higher, syntactic level. Consider, for instance, the example in (15), 'hundred-franc bills'. (15)

bilhed bill

-ού

(i.e. worth 100 francs) ('hundred-franc bills')

In this expression, the prepositional phrase is interpreted distributively-that is, Ί 0 0 francs' names the value of each individual bill and not of all the bills together. In order to account for this fact, one might assume that the plural operator has scope over the entire expression, as in (16): (16) F p i a x [ b i l l ' ( x ) & worth-lOO-francs'(x)]) In other words, one might assume that the plural feature on the top N-bar node in (15) has semantic relevance, but that this same feature is semantically empty in the smaller constituents of (15)-in particular, in the plural suffix -oil In this way, one might conclude that inflectional affixes are simply irrelevant at the interface between morphology and semantics. While such an approach to resolving the mismatches in (12) and (14) might appear to be workable, it would in fact be incompatible with certain independent facts about Breton morphology; in particular, there are certain clear cases in which plural inflectional affixes are semantically relevant at the word level. T o see this, consider the denominal verb-deriving process exemplified in (17). This process applies to a noun χ to form a

300

Gregory Τ. Stump

verb whose meaning is roughly 'to seek after x' and whose infinitive form is marked by the special suffix -a. (17) Denominal verbs of seeking: a. singular mass noun bijin 'seaweed' keuneud 'firewood' koc'h-kezeg'horse manure' raden

'ferns'

b. collective noun brinnig 'limpets' buzhug 'earthworms' logod 'mice' melved 'slugs' sivi 'strawberries' c. singular count noun aval 'apple' brezhel 'mackerel' labous 'bird' merc'h 'girl' pesk 'fish'

verbal derivative bijina 'to gather seaweed' keuneuta 'to gather firewood' koc'hkezeka 'to gather horse manure' radena 'to gather ferns'

verbal derivative brinnika 'to gather limpets' buzhuka 'to gather earthworms' logota 'to hunt for mice' melveta 'to look for slugs' sivia 'to pick strawberries'

plural

verbal derivative

avaloü brizhili laboused merc'hed pesked

avalaoua 'to look for apples' brizhilia 'to fish for mackerel' labouseta 'to hunt for birds' merc'heta 'to chase girls' pesketa 'to fish'

[N.B.: -a devoices a preceding obstruent, and οι) becomes aou in stressed (penultimate) position.] What's interesting about this process is that it only applies to nouns which have plural reference-that is, nouns which don't simply denote sets of atomic individuals. Thus, this process can apply to mass nouns, as in (17a); to basic collective nouns, as in (17b); and to count nouns, provided that these are in their plural form, as in (17c). Note that this is a semantic rather than a purely morphological restriction; one couldn't, for example, simply say that the process applies to nouns marked with the feature 'plural', since the mass nouns in (17a) behave as singulars for purposes of agreement. But if this is a semantic fact about the verb-deriving process in (17), then clearly the plural morphology in (17c) is semantically significant.

Two types of mismatch

301

This conclusion is confirmed by a second type of evidence, the denominal adjectives in (18). (18) Denominal privative adjectives: a. singular adjectival derivative penn 'head' dibenn 'scatter-brained' komz 'word' digomz 'mute, silent' kabell 'cap' diskabell 'bareheaded' b. singular

botez 'shoe' loer 'sock' draen 'thorn'

plural

boutoü lerou drein

adjectival derivative divoutou 'without shoes on' dilerou 'without socks on' dizrein 'without thorns'

As these examples show, an adjective with the meaning 'without x' can be formed by joining the prefix di(s)- to a noun x. What is significant here is that the number of the noun to which di(s)- is prefixed is clearly reflected in the meaning of the resulting adjective; so here again, plural inflections must be recognized as having semantic relevance at the interface between morphology and semantics. It should be noted, too, that this fact is perfectly consistent with the distributive meaning assigned to the prepositional phrase in the complex expression in (15). This distributive meaning doesn't have to be attributed to a plural operator with wide scope; instead, it could be attributed to the presence of a kind of distributive operator DISTR having the semantic effects in (19). (19)DISTR may be identified with Link's '*' operator: where Ρ denotes a set S of atomic individuals, *P denotes the set of individual sums of members of S (see Link 1983). With this operator, the expression in (15) might be assigned the semantic representation in (20), in which the plural operator itself has word-level rather than phrase-level scope. (20) Xx[DISTR(worth-100-francs')(x) &F p l (bill')(x)] (equivalent to [16]) I conclude from these facts that plural morphology cannot just be regarded as semantically empty, at least not in Breton; so

302

Gregory Τ. Stump

some other way of resolving the structural mismatches in (12) and (14) must be found. There does, in fact, seem to be one additional way to provide a structure-based resolution of these mismatches; this would be to assume that for associations between semantics and morphology, inflectional formatives define a separate association plane. On this assumption, the association between the structure of a plural diminutive and its meaning would be as in (21). Fdim

(21)

[[[[ ('little boats') This seems like a promising approach, but one clear difficulty with it stands out. The association in (21) presumes that every terminal element in the morphology of a plural diminutive is matched with some element in its semantics; but this can't possibly be right for those expressions mentioned earlier which have the form of diminutives but whose meaning is idiosyncratic. Consider again the noun tokig 'robin' in (10a). Though it can be formally analyzed as a noun with a diminutive suffix on it, it is semantically unanalyzable; in particular, its meaning cannot be viewed as the simple result of applying the diminutive operator to the meaning of foic'hat'. (22) F pl (tokig')[where tokig' is not analyzable as Fdim(tok')] (23)

tokig' [[[[

tok]

-ed]

-ig)

-ου]

Crobins'/*'little hats') On this assumption, the plural diminutive form tokedigoii has the semantic representation in (22); but if this is so, then how is the morphological structure of the plural form to be associated with

Two types of mismatch

303

its semantic representation? The association in (23) is very dubious, since the meaning of tokig has no connection either to tok or to the diminutive suffix. Similar difficulties arise with l o o s e compounds, nearly all of which exhibit some degree of semantic idiosyncrasy. The problem here, I claim, is the assumption that terminal elements must be associated at the interface between morphology and semantics. As an alternative, consider now the possibility of adopting the maximal word restriction given above in (3). Under this restriction, the only possible association between the morphological and semantic trees in (12) is between the two nodes which exhaustively dominate the two trees; and similarly for the compound example in (14). On this view, the match between the morphology and the semantics of a plural diminutive or compound never amounts to more than a single, unproblematic association line, as in (24). (24) a. M

N

M

N

b.

F pl (F dim (bag')) I bag]-oü]-ig]-oü] F pl (tok-kolo') I

U N

[N

I O / F

] "

Ο Ι }

ΠΝ

kolo]]

Assuming that this is so, the question now becomes: what are the principles which regulate this sort of association?

3. A paradigm-based account of the Breton mismatches The analysis which I shall develop here embodies the paradigmbased approach given in (4) above, and presupposes a novel conception of the morphological component of a grammar. T w o basic premises underlie this alternative approach. The first of these is that some morphological expressions exist as members of a paradigm, where a paradigm is simply viewed as a sequence of ordered triples, comparable to the list in (25). The first member of each ordered triple identifies the role which the second member plays within the paradigm; for instance, 'PLURAL' identifies the role of bagoü in the paradigm in (25).

304

Gregory Τ. Stump

(25) Paradigm of bag'boat': < SINGULAR: [ N b a £ ] , b a g ' > < PLURAL: [N [ N bag ] -ού ], F pl (bag') > In addition, the third member of each triple is the semantic representation associated with the second member; thus, bag' is the semantic representation of the singular form bag in (25). Now, I assume that the lexical listing of a Breton noun has space for a paradigm comparable to the one in (25). For nouns with irregular or suppletive plurals, both the singular and the plural form may be specified, as in (26a); but I assume that in accordance with the partial listing hypothesis, the ordinary situation is one in which only the singular form is given, with its plural unspecified, as in (26b) and (26c). (26) a. Partial lexical entry for 'man': < SINGULAR: [N den ], den' > < PLURAL: [N tud ], > b. Partial lexical entry for 'boat': < SINGULAR: [N bag], bag > < PLURAL: , > c. Partial lexical entry for 'hat': < SINGULAR: [N tok ], tok' > < PLURAL: > Moreover, I assume that the semantic representation of a plural form is in general left unspecified, being in virtually all cases predictable as a function of the semantic representation of the corresponding singular form. The s e c o n d premise underlying the paradigm-based conception of morphology is that the morphological component of a language comprises two distinct types of rules, which I shall call morpholexical rules and paradigm rules. The morpholexical rules are responsible for generating complex morphological expressions from more basic expressions; as such, they carry out the full range of possible morphological operations, whether concatenative or not, whether inflectional, derivational, or compositional. The paradigm rules, on the other hand, serve to relate members of the same paradigm; in effect, they allow the form of a complex, lexically unlisted expression to be deduced from that of a related expression, even in those cases in which the former expression doesn't actually come from the latter.

Two types of mismatch

305

T o account for the Breton data discussed above, three morpholexical rules are necessary: one to generate suffixed plurals, another to generate diminutive nouns, and a third to generate loose compounds. In stating the first of these rules, I assume that Breton nouns carry a 'PL-SUFFIX' feature, whose value for a given noun is simply the suffix by which that noun is pluralized. On this assumption, the noun laer 'robber' in (27) has the partial lexical entry in (28a). The value of this feature does not have to be explicitly specified for most nouns, however: since the vast majority of Breton nouns are pluralized by means of the suffix -οιi, it can simply be assumed that this suffix is the unmarked, default value of the 'PL-SUFFIX' feature in those instances in which no value is explicitly specified, as in (28b). (27) (28) a.

laer 'robber', plural laeron Partial lexical entry for 'robber': < SINGULAR: [ N laer ], laer' > < PLURAL: >; [PL-SUFFIX:on] b. default value: [PL-SUFFIX:ou] (e.g., bagoCi 'boats', tokou 'hats')

With this formalism, the morpholexical rule for suffixed plurals can be stated as in (29), along with the morpholexical rules for diminutives and loose compounds. (29) Morpholexical rules for Breton nouns: a. b. c.

Plural suffixation: M L R - P L ( [ n X ] ) = [ N [ N X ] - a].where χ is [ P L - S U F F I X : a ] Diminutivization: MLR-DIM([N χ ] ) = [N [N x]-ig] Loose compounding: MLR-CMPD([n X ],[N y ] ) = [N [N X ][n y ]]

Note that the morpholexical rules are stated not as context-free rewrite rules, but as operations on morphological expressions; 5 as will be seen momentarily, this is necessitated by the manner in which they interact with the paradigm rules. The rules in (29) are responsible for generating plurals, diminutives, and compounds; the paradigm rules, on the other hand, relate the morphological structure of a singular noun to that of the corresponding plural. More precisely, the paradigm rules

306

Gregory Τ. Stump

serve as clauses in the recursive definition of a function "plural-of", which, when applied to any singular noun in the language, gives the corresponding plural form as its value. For singular nouns whose plurals are suppletive or irregular, the value of the plural-of function is simply stipulated in the lexicon, as in (26a). For all other nouns, however, the value of this function is determined by the paradigm rules in (30). PI is the default paradigm rule, which predicts that the plural form of a given noun is simply the result of appending the appropriate plural suffix to that noun; PI may, however, be overridden by P2 or P3, which specify the more complex s t r u c t u r a l relations which singular diminutives and loose compounds bear to their plurals. (30) Paradigm rules for Breton plurals: Where α is a singular noun, PI. Default: plural-of (a) = MLR-PL(a). P2. Diminutives: If a = MLR-DIM(x), then plural-of(a) = MLR-PL(MLR-DIM(plural-of(x))). P3. Loose compounds: If α = MLR-CMPD(x,y), then plural-of(oc) = MLR-CMPD(plural-of(x),y). Note that the paradigm rules are not necessary for generating any morphological expression; on the contrary, the value of the plural-of function for any argument will always be an expression that is generated independently by the morpholexical rules. In other words, the paradigm rules just relate expressions whose existence is already guaranteed by the morpholexical rules. Nevertheless, the paradigm rules serve three important functions. First, they make it possible to maintain the claim that regularly inflected plural forms are not lexically listed, since they can predict these forms from the corresponding singular forms, even in those cases in which the plural form does not come from the singular form; thus, given the lexically listed singular forms in (31a-b), rules P2 and P3 predict the corresponding plurals in (31c). The second function of the paradigm rules is to serve as a filter on the morpholexical rules. As they are stated, the morpholexical rules overgenerate profusely; for instance, besides producing such acceptable words as (32a), the rules for pluralization and diminutivization jointly produce such nonwords as those in (32b-d).

Two types of mismatch

307

(31) a.

Partial lexical entry for 'little boat': < SINGULAR: [ N [ N bag)-ig], bagig' > < PLURAL: , > b. Partial lexical entry for 'straw hat': < SINGULAR: [ N [ N tok][N kolo ]], tok-kolo' > < PLURAL: , > c. Lexically unlisted plural forms:

(32) a. b. c. d.

InUMn bag]-oü]-ig]-oü] [N [s[Ntok] -ou][N kolo ]] bagouigou [= MLR-PL(MLR-DIM(MLR-PL(bag)))] *bagigou [= MLR-PL(MLR-DIM( bag)) ] *bagoüig [= MLR-DIM(MLR-PL( bag)) ] *bagowgowg [=MLR-DIM(MLR-PL(MLR-DIM(MLR-PL(Z?a£))))]

Though all of the forms in (32) are generated by the morpholexical rules (and though all would seemingly count as plural forms, under standard assumptions about feature percolation), only the first form, bagouigoü, is linked to a singular form by the paradigm rules (specifically, by rule P2). T o account for this fact, all lexically unlisted plural forms generated by the morpholexical rules could be assumed to be subject to the well-formedness condition in (33): (33) An unlisted plural form is well-formed only if for some well-formed expression β, α = plural-of(ß). This constraint guarantees that the three forms in (32b-d) will be filtered out by the paradigm rules. The third and final function of the paradigm rules is to facilitate the association of lexically unlisted morphological structures with the appropriate semantic representations. Once the plural-of function is defined by the paradigm rules, the rule of association for plural nouns in Breton can be given the trivially simple formulation in (34). (34) Morphology/semantics association rule for Breton plurals

If

a' I α

is a match, then

F p l (a') I plural-of( a )

is a match,

Gregory Τ. Stump

308

This rule guarantees that the two unlisted plurals in (31c) will be associated with exactly the right semantic representations, as in (35): (35) a.

b.

F pl (bagig')[=F pl (F dim (bag'))] I bagoüigou [= plural-of(bagig)] F.(tok-kolo') p. ι tokou-kolo [= plural-of(tok-kolo)]

This sort of analysis can be easily extended to accommodate the problematic diminutive forms given earlier in (10). Consider again the expression tokig 'robin' in (10a). This expression is problematic for two reasons: first, its idiosyncratic meaning cannot be computed from the meanings of the nominal base tok 'hat' and the diminutive suffix -ig, and second, the plural suffix -ed which is attached to the nominal base tok in the formation of the plural diminutive tokedigou 'robins' is not the plural suffix which tok 'hat' ordinarily requires. To account for these facts within the proposed paradigm-based framework, one need only assume the partial lexical listing in (36), in which tok2 is a bound root which bears the feature specification [PL-SUFFIX:ed] (and is hence distinct from tok 'hat' in [26c]). (36) Partial lexical entry for 'robin': < SINGULAR: [ N [ N tok2 ] -ig ], tokig' > < PLURAL: On this assumption, the paradigm rule P2 in (30) correctly predicts that tokig has tokedigou as its plural form, and the association rule in (34) correctly links this plural form with the desired semantic representation, as in (37). Note that neither the semantic representation of tok 'hat' nor the semantic counterpart of the diminutive suffix (i.e., the operator Fdim) is associated with tokedigoiι in this analysis. (37) F pl (tokig') tokedigoü [= plural-of (tokig)]

Two types of mismatch

309

Because they allow a plural diminutive or loose compound to be associated with its semantic representation at the maximal word level, the paradigm rules make it possible to avoid crossing association lines and semantically unmotivated associations at the morphology/semantics interface; in this way, the paradigm-based approach to lexically unlisted associations provides an account of Breton plurals superior to that of the structure-based approach considered earlier. And, as I shall now show, the paradigm-based approach also affords the cleanest account of a second, v e r y different type of mismatch between morphology and semantics.

4. Derivative nouns in Tigre6 This second type of mismatch is exemplified by the system of nominal morphology in Tigre, a Semitic language of Ethiopia. In Tigre, nouns are distinguished by gender and number; in addition, singular nouns may be classed as either singulative or collective. Both singulative and collective nouns induce singular concord in a verb or adjective; the collectives, however, lack individual reference, and therefore fail to appear with numerals, unlike the singulatives. Whether singulative or collective, Tigre nouns give rise to a number of derivative formations: for instance, collective nouns may be converted into forms with an augmentative meaning or into singulatives; singulatives, in turn, may be converted into diminutives or pejoratives. These various possibilities are schematized in (38).The conversions indicated by the arrows in (38) are carried out by means of the four suffixes listed in (39). This seems like a simple enough system until some actual forms are encountered. (38) Derivative formations among singular nouns in Tigre coll^tive

—•

augmentative

singulative

—•

pejorative

diminutive

310

Gregory Τ. Stum ρ

(39) Derivative suffixes -ät -it

I

-ay -etay

( p r o d u c e feminine nouns)

} ( p r o d u c e masculine nouns)

Consider, for example, the forms in (40). (40) S o m e singular nouns and their derivatives: collective singulative diminutive pejorative augmentative 'cow': (bound s t e m c l a s s ] ) w a w a ' - ä t wa '-ät-it 'charcoal': (basic c o l l e c t i v e a a s s l ) fahäm fähm-ät fähm-ät-it 'woman': (basic singulative f e m ) 'assit 'assit-ät 'man': (basic singulative m a s c ) 'anas 'anes-ay 'ox': (bound stem c l a s s I I ) ba'rba'r-ay ba'r-etay 'nits': (basic collective c l a s s I I ) £ib cib-ay

wa '-etay

wa '-ay

fähm-etay

fähm-ay

'assit-ay

-

'anes-ät

-

As t h e s e e x a m p l e s show, there is no e a s y way to gloss the Tigre derivative suffixes. Notice, for instance, that the suffix -ay marks the singulative form of 'ox', the diminutive f o r m of 'man', the pejorative form of 'woman', and the augmentative form of 'charcoal'. By the s a m e token, notice that any one of the four suffixes listed in (39) may be the mark of a diminutive form: -ät marks the diminutive of 'woman', -ay marks the diminutive of 'man', -etay m a r k s the diminutive of 'ox', and -it marks the diminutive of 'charcoal'. Clearly there is a major mismatch in Tigre between the form of a derivative affix and the element of meaning which it signals. This is different f r o m the Breton e x a m p l e s , s i n c e the m i s m a t c h here is not a structural one; on the contrary, it s e e m s perfectly plausible to a s s u m e that e a c h of the f o r m s in (40) h a s a semantic s t r u c t u r e which is the mirror image of its morphological structure, as, for

Two types of mis ma tch

311

example, in (41). The mismatch here emerges only when two or more nouns are compared; the obvious problem is that the derivative suffixes seem to signal widely divergent meanings from one example to the next. (41)



('little piece of charcoal') [ Semantics of F : where Ρ denotes a set S of proper i-sums, Fsgv(P) denotes tne set of individuals serving as atoms in these i-sums.] This fact presents obvious difficulties for the structure-based approach discussed earlier, which hinges on the possibility of identifying point-by-point matches between terminal elements. A proponent of this approach could, of course, attribute the mismatch to rampant homophony. For instance, one could say that there are actually four different suffixes pronounced as -ay and that one of these is uniformly singulative in interpretation, another is diminutive, still another is pejorative, and the last is augmentative. But this move would surely be wrong. To see why, suppose that there were four different derivative suffixes with the form -ay. If this were the case, then a number of consequences would be expected to follow. One would expect (i) that words might be found in which two or more of the homophonous suffixes contrasted minimally; (ii) that derivative nouns suffixed with -ay might vary in gender, depending on which of the homophonous suffixes they contained; (iii) that the different homophones might vary in their ability to combine with other suffixes; and (iv) that two or more of

312

Gregory Τ. Stump

the homophonous suffixes might show up together in the same word. Careful scrutiny of the Tigre data fails to confirm any of these expectations. Thus, if one attributed the semantic variability of the suffix -ay to four-way homophony, then one would have to view it as pure coincidence that the supposed homophones are in complementary distribution, that they all assign masculine gender, that they all occupy the very same position in the permissible sequence of suffixes, and that they happen never to appear together in the same word. Parallel facts obtain for the other derivative suffixes listed in (39), casting serious doubt on the homophony solution-and more generally, on the structure-based approach to associations involving lexically unlisted structures. Before I propose an alternative account of the Tigre mismatch, a clearer view of the relevant data will be necessary. Careful scrutiny of the forms in (40) reveals several generalizations. First, the basic, unsuffixed nouns in this system are collectives and singulatives. Second, the precise semantic effect of a particular suffix depends on whether it changes or preserves the gender of the base to which it attaches: thus, the suffix by which a singulative is converted into a diminutive always preserves the gender of the singulative; on the other hand, the suffix by which a singulative is converted into a pejorative always produces a form opposite in gender to the singulative. The singulative and augmentative derivatives of a collective noun are likewise opposite in gender. Finally, there are important restrictions on the order and combinability of the derivative suffixes: the suffixes -ät and -ay attach only to basic, unsuffixed nouns, while -it attaches only to nouns suffixed with -ät, the suffix -etay joins with nouns already suffixed with -ät or -ay, but supplants this suffix. These different generalizations are schematically summarized in (42), in which the network of relations between basic nouns and their derivatives can be seen to be quite systematic. With these facts in mind, consider now the possibility of resolving the mismatch between Tigre derivatives and their meanings by means of a paradigm-based system of morphology like the one proposed earlier for Breton. In a system of this sort, the morpholexical rules for Tigre derivatives are trivially easy to state: all that's needed is a rule for each of the four suffixes, as in (43). As with the Breton cases discussed earlier, it is the paradigm rules which do the work. Here, the paradigm rules make it possible to predict the form of a singulative or augmentative noun from the form of the corresponding collective, and to predict the form of a

Two types of mismatch

313

diminutive or pejorative from the form of the corresponding singulative. (42) basic form

(43)

with -ät added

with -it added

Morpholexical rules for Tigre derivatives a. b. c. d.

MLR,([Nx]) MLR 2 ([ N χ ] ) MLR3([Nx]) MLR 4 ([ N x ])

= [N[FEM][Nx]-äf ] = [N[FEM][Nx]-it] = [N[MASC][Nx]-ay] = [ N[MASC] [ N x ] -etay]

More technically, the paradigm rules serve as clauses in a recursive definition of four functions: the functions "singulative-of" and "augmentative-of", whose domain is the set of basic collective nouns; and the functions "diminutive-of" and "pejorative-of", whose domain is the set of singulative nouns. These rules may be stated as in (44); note that these rules encode exactly the network of relations schematized in (42).

314

Gregory Τ. Stump

(44) Paradigm rules for Tigre derivatives: Where α is a singulative noun, χ is a basic member of Class I, and y is a basic member of Class II, PI P2 P3

P4

P5 P6

singulative-of(x) = MLR,(x) and augmentative-of(x) = MLR 3 (x). singulative-of(y) = MLR 3 (y). If α is a basic feminine noun, then diminutive-ofΓα) = MLRjCa), and p e j o r a t i v e - o f ( a ) = MLR3(oO. If a is a basic masculine noun, then diminutive-of(a) = MLR 3 (a), and p e j o r a t i v e - o f ( a ) = MLRjCa). If a = MLRjCx), then diminutive-of(a) = MLR 2 (a) and pejorative-of(a) = MLR 4 (x). If a = MLR 3 (y) and y is a bound stem, then diminutive-of(a) = MLR 4 (y).

As in the Breton cases discussed earlier, the paradigm rules serve three distinct functions. First, they make it possible to maintain the partial listing hypothesis, since they can predict lexically unlisted derivatives from basic collectives and singulatives; thus, they make it possible to assume that the lexical listing for the collective noun fdhäm 'charcoal' omits any explicit listing of the corresponding derivatives, as in (45). (45) fdhäm, Ν, Class I, < SINGULATIVE: _ < AUGMENTATIVE: < DIMINUTIVE: , < PEJORATIVE: Second, the paradigm rules again serve as a filter on the morpholexical rules. Notice, for instance, that if the morpholexical rules in (43) were simply allowed to apply and reapply without restriction, they would massively overgenerate, producing such ungrammatical combinations as those in (46). (46) a. b. c. d. e.

* WD '-it [ = MLR 2 ( wd -) ] *wa'-it-ät [ = MLR!(MLR 2 ( wa'-))] '-ät-ay [ = MLR/MLRj iwa - ) ) ] *wd'-ay-ät [= MLRj(MLR 3 (wa-))] * wo'-ay-it [= MLR 2 (MLR 3 (W9 -))]

Two types of mismatch

But according to the paradigm rules in (44), none (46) qualifies as the singulative, augmentative, pejorative derivative of the basic stem W9'- 'cow'. the forms in (46) can be easily filtered out by means in (47):

315

of the forms in diminutive, or Given this fact, of the principle

(47) An unlisted derivative noun α is well-formed only if for some well-formed expression β, α = diminutive-of (β), pejorative-of(/U singulative-of (β), or augmentative-of(ß.). Finally, the paradigm rules in (44) make it possible to formulate the association rule for derivative forms with the utmost ease, as in (48). (48) Morphology/semantics association rule for Tigre derivatives If

a' J α

is a match, then each of the following is a match:

a. F d i m (a') I diminutive-of (a)

b.

F p e j (a') I pejorative-of(a)

c.

d.

F a u g (a')

F s g v (a')

I I s i ngulat ive-of (a) augmentati ve-of (a) [Semantics of Faug: where Ρ denotes a set of proper i-sums and S is the set of individuals serving as atoms in these i-sums, F aug (P) denotes the set of relatively large members of S. Semantics of F ^ identity function. Pragmatics of F · F p e j (P)(x) conventionally implicates the speaker's discfain for x's referent.] With this rule, there's no real mismatch between the form and the meaning of a Tigre derivative, since no specific semantic content is attributed to any of the derivative suffixes, as it would have to be in the structure-based approach. Here again, the paradigm-based

316

Gregory Τ. Stump

approach provides the most satisfying account of lexically unlisted associations at the morphology/semantics interface.

5. Conclusions In concluding, I should emphasize that under the paradigm-based approach advocated here, the mismatches in Breton and Tigre are not resolved at the interface between morphology and semantics; instead, they are resolved within the morphological component itself. T o that extent, the analyses developed here are not Autolexical, and could in fact be incorporated into a grammatical framework in which the semantic representation of a complex word is built up interpretively. But notwithstanding this fact, the p a r a d i g m - b a s e d conception of m o r p h o l o g y has important consequences for Autolexical theory, since it provides the most workable means of reconciling the maximal w o r d restriction with the partial listing hypothesis. 7

Notes 1.

2.

3.

4.

The data in this section are drawn from fieldwork in Plougastel (extreme northwestern Cornouaille) in 1989-90, from Gros (1984), and from Ernault (1984). This definition of Fdim is based upon the discussion of the semantics of diminutives in Wierzbicka (1980: 53-60); my thanks to Jim McCawley for drawing my attention to her insights. Note that in addition to being semantically different, the base and diminutive forms in (10) are formally divergent as well: in the formation of the diminutive plurals in (10), the nominal bases tok and moc'h take the plural suffix -ed (commonly found with animate nouns) despite the fact that neither has a plural in -ed outside of the diminutive formation. This fact will be returned to below. It might be claimed that the expressions in (13) are not compounds at all, but are instead instances of the Breton syntactic combination [ Ν NP ] exemplified by mamm Mona 'Mona's mother'. There is, however, clear evidence against this position. For instance, loose compounds of the form [ N N ] show a mutation pattern not found in syntactic combinations such as mamm Mona: if the first noun is feminine and singular, the second noun shows initial lenition, as in

Two types of mismatch

5.

6. 7.

317

(13a) and (13b). In addition, loose compounds occasionally undergo processes of derivation, e.g., koc'h-kezeg 'horse manure' — koc'hkezeka 'to gather horse manure'. Thus, I am assuming the process-based conception of morphology advocated by Matthews (1972), Anderson (1982), and Zwicky (1985) rather than the more widely held conception of morphology as wordlevel syntax advocated by Williams (1981), Selkirk (1982), and Pesetsky (1985); I do, however, assume that the expressions generated by morpholexical rules have a hierarchical rather than purely linear structure. The Tigre data discussed here are from Palmer (1962). For a more detailed discussion of the usefulness of the paradigmbased approach to morphology in resolving morphosemantic mismatches of various types, see Stump (1991).

References Anderson, Stephen R. 1982 "Where's morphology?", Linguistic inquiry 13: 571-612. Ernault, Emile 1984 Geriadurig Brezoneg-Galleg [Short Breton-French Dictionary]. (2nd edition.) Brest: Emgleo Breiz. Gros, Jules 1984 Le tresor du breton parle, troisieme partie: Le style populaire. (2nd edition.) Brest: Emgleo Breiz. Huck, Geoffrey J. — Almerindo E. Ojeda (eds.) 1987 Syntax and Semantics, Vol. 20: Discontinuous constituency. San Diego: Academic Press. Lapointe, Steven G. 1987 "Some extensions of the Autolexical approach to structural mismatches", in: Geoffrey J. Huck — Almerindo E. Ojeda (eds.), 152-183. 1988 "Distinguishing types of morphosyntactic cooccurrences: Mismatch resolution, agreement, and government", in: Diane Brentari - Gary Larson - Lynn MacLeod (eds.), Proceedings of the 24th Regional Meeting of the Chicago Linguistic Society, Part 2: Parasession on agreement in grammatical theory. Chicago: Chicago Linguistic Society, 181-201. Link, Godehard 1983 "The logical analysis of plurals and mass terms: A latticetheoretical approach", in: R. Bäuerle — C. Schwarze — A. von Stechow (eds.), Meaning, use, and interpretation of language. Berlin: Walter de Gruyter.

318

Gregory Τ. Stum ρ

Matthews, Peter Η. 1972 Inflectional morphology. Cambridge: Cambridge University Press. Palmer, Frank R. 1962 The morphology of the Tigre noun. London: Oxford University Press. Pesetsky, David 1985 "Morphology and logical form", Linguistic Inquiry 16: 193246. Sadock, Jerrold M. 1985 "Autolexical syntax: a proposal for the treatment of noun incorporation and similar phenomena", Natural Language and Linguistic Theory 3: 379-439. 1987 "Discontinuity in Autolexical and autosemantic syntax", in Geoffrey J. Huck - Almerindo E. Ojeda (eds.), 283-301. 1988 "The Autolexical classification of lexemes", in: Theoretical morphology: Approaches, in: M. Hammond-M. Noonan (eds.), Modern linguistics. New York: Academic Press. Selkirk, Elisabeth 0. 1982 The syntax of words. Cambridge, MA: MIT Press. Stump, Gregory T. 1991 "A paradigm-based theory of morphosemantic mismatches", Language 67: 675-725. Wierzbicka, Anna 1980 Lingua mentalis: The semantics of natural language, Sydney: Academic Press. Williams, Edwin 1981 "On the notions 'lexically related' and 'head of a word"', Linguis tic In q uiry 12: 245-274. Zwicky, Arnold M. 1985 "How to describe inflection", Proceedings of the Berkeley Linguistic Society 11: 372-386.

On restricting the role of morphology in Autolexical Syntax* Anthony C. Woodbury

Introduction In Autolexical Syntax (Sadock 1985, 1988, 1991), syntax, semantics and morphology are taken to form the minimum set of autonomous modules in a parallel grammar. T o them is sometimes added a phonological or morphophonological module, organized along familiar lines into lexical and postlexical levels (Chelliah 1988; Farkas - Sadock 1989). In a framework with so many modules, it is crucial to insure that ( a ) each module is internally simple and consistent, and ( b ) constraints on the mapping from module to module are as stringent as they can be, in keeping with the facts of natural language. Both goals are advanced by recognizing fully the role of an autonomous morphophonological module and by distinguishing it carefully from the module concerned with morphology proper, that is, the categorially-based syntax of elements within words. In particular, I wish to show that by handling the phenomenon of cliticization in morphophonology rather than in morphology (as earlier proposed): (a) Autolexical morphology is made simpler and more internally consistent, and Cb) intramodular constraints are resolved into three types with distinct properties, depending on the interface involved: morphophonology-syntax, morphologysyntax, and morphophonology-morphology. This reorganization within Autolexical Syntax is the simple result of the clitic analysis but has consequences extending to such diverse phenomena as incorporation, classic bracketing paradoxes, multi-word compounding, and productive expressive morphophonology. The same clitic analysis-for it is actually the one most generally a c c e p t e d - d o e s not lead to such a range of consequences in other frameworks. Therefore, in so far as these consequences are desirable ones, they especially recommend the highly autonomous, parallel architecture of Autolexical Syntax that makes them possible.

320

Anthony

C.

Woodbury

1. Clitics in Autolexical Syntax We first turn to the treatment of clitics. In Autolexical Syntax, clitics have been held to show complex autostructural association across the interface of syntax and morphology. In syntax they are simply atoms or terminal elements, while in morphology they combine with inflected words to form "superwords". This is shown more technically in (1), which recapitulates a part of the standard Autolexical morphology. In (1), negative bar levels represent inflected words as X"1, uninflected stems asX"°, and postinflectional superwords as X"2: (1) Autolexical Morphology (based on Sadock 1988, and 1991: section 2.1.4.3) MF1 X"1 = X"°, Y