The Representation of Implicit and Dethematized Subjects 9783110877304, 9783110131147

Citation preview

The Representation of Implicit and Dethematized Subjects

Linguistic Models The publications in this series tackle crucial problems, both empirical and conceptual, within the context of progressive research programs. In particular, Linguistic Models will address the development of formal methods in the study of language with special reference to the interaction of grammatical components Series Editors: Teun Hoekstra Harry van der Hulst Michael Moortgat

Other books in this series: 1 2 3 4 5 6 7 8 9

Michael Moortgat, Harry van der Hulst and Teun Hoekstra (eds.) The Scope of Lexical Rules Harry van der Hulst and Norval Smith (eds.) The Structure of Phonological Representations. Part I. Harry van der Hulst and Norval Smith (eds.) The Structure of Phonological Representations. Part II. Gerald Gazdar, Ewan Klein and Geoffrey K. Pullum (eds.) Order, Concord and Constituency W. de Geest and Y. Putseys (eds.) Sentential Complementation Teun Hoekstra Transitivity. Grammatical Relations in Government-Binding Theory Harry van der Hulst and Norval Smith (eds.) Advances in Nonlinear Phonology Harry van der Hulst Syllable Structure and Stress in Dutch Hans Bennis Gaps and Dummies

The Representation of Implicit and Dethematized Subjects Ian G. Roberts

¥

1987 FORIS PUBLICATIONS Dordrecht - Holland/Providence -U.S.A.

Published by: Foris Publications Holland P.O. Box 509 3300 AM Dordrecht, The Netherlands Sole distributor for the U.S.A. and Canada: Foris Publications USA, Inc. P.O. Box 5904 Providence RI 02903 U.S.A. Sole distributor for Japan: Toppan Company, Ltd. Shufunotomo Bldg. 1-6, Kanda Surugadai Chiyoda-ku Tokyo 101,Japan

CIP-DAT A Roberts, Ian G. The Representation of Implicit and Dethematized Subjects / Ian G. Roberts. - Dordrecht [etc.]: Foris. - (Linguistic Models ; 10) With ref. ISBN 90-6765-301-2 bound ISBN 90-6765-302-0 paper SISO 805.1 U D C 801.5 Subject heading: generative grammar.

ISBN 90 6765 301 2 (Bound) ISBN 90 6765 302 0 (Paper) © 1986 Foris Publications - Dordrecht No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission from the copyright owner. Printed in The Netherlands by ICG Printing, Dordrecht.

I'd like to dedicate this book to my mother and my brother, Paul

Contents

Preface Chapter One Background 1.1. Introduction 1.2. X-bar Theory 1.3. Principles of Thematic Structure 1.3.1. Case Theory 1.3.2. θ-theory and the Projection Principle 1.4. Principles of Quantification Notes to Chapter One

1 1 4 7 7 10 12 14

Chapter Two Passives and Argument Structure 2.1. Passive in GB Theory 2.1.1. The Core Passive 2.1.2. Categorial Properties of Passives 2.1.3. Case Properties of Passives 2.1.4. θ-properties of Passives 2.2. An Alternative Proposal 2.2.1. Passive Morphology 2.2.1.1.En as a clitic 2.2.1.2. Consequences of the Proposal for en 2.2.1.2.1 .Case and Affix-Hopping 2.2.1.2.2.The Representation of the Subject 2.2.1.2.3.The 1-Advancement Exclusiveness Law and the Mirror Principle 2.2.1.3. Affix-Hopping and Auxiliary System 2.2.1.3.1 .Perfect and Passives 2.2.1.3.2.Evidence for syntactic Affix-Hopping: Gapping 2.2.1.4. Conclusion 2.2.2. Subjects and Agents 2.2.3. Lexical and Structural relations: The Projection Principle 2.2.4. Clitics Notes to Chapter Two

17 17 17 19 23 27 32 32 32 34 34 35 36 38 40 43 45 45 51 57 64

viii

The Representation of Implicit and Dethematized

Subjects

Chapter Three The By-Phrase and Implicit Arguments 3.1. Introduction 3.2. Predication and Modification of Implicit Arguments 3.2.1. Adverbs 3.2.1.1. Positional Classes of Adverbs 3.2.1.1.1 .Initial and Final 3.2.1.1.2."Aux position" 3.2.1.2. "Thematic" Properties of Adverbs 3.2.2. Adjectives 3.2.3. Implicit Arguments with PP 3.2.4. Passive and Infinitival Adjuncts 3.2.4.1. Rationale Clauses 3.2.4.1.1.Distribution of Rationale Clauses 3.2.4.1.2.Rationale Clauses and Control 3.2.4.1.3.Rationale Clauses vs. Purpose C l a u s e s . . . 3.2.4.2. Non-purposive Adjuncts 3.3. Referential Properties of Implicit Arguments 3.3.1. Binding Theory and θ-theory 3.3.2. Strong Crossover, Chains and Principle C 3.3.3. The Interpretation of Implicit Arguments 3.3.3.1. Weak Crossover in Passives 3.3.3.2. Implicit Arguments as Quantifiers 3.3.3.3. Implicit Arguments and Extraction 3.3.4. Binding, Coreference and Implicit Arguments 3.4. Conclusion Notes to Chapter Three

69 69 71 71 72 72 74 81 82 96 101 101 101 104 115 120 125 125 133 143 143 145 151 160 172 173

Chapter Four Middles and Nominals 4.1. Properties of Middles 4.2. Middles and Stativity 4.3. Aspectual Classes of Verbs and Transitivity 4.3.1. The Dowty/Vendler Verb Classification 4.3.2. Burzio's Generalization and Aspect 4.4. Middles, Ergatives and Modification 4.4.1. Middles, Ergatives and Unaccusatives 4.4.2. Middles and Predication 4.5. Nominals 4.5.1. GFs in NP 4.5.2. "Passive in N P " 4.5.3. Modification and Implicit Arguments in NP 4.5.4. Determiners as Arguments 4.6. Conclusion Notes to Chapter Four

185 186 195 207 208 217 227 227 232 239 240 245 249 252 256 256

Contents

ix

Chapter Five Impersonal Constructions and the Passive in UG 5.1. Romance si/se 5.2. Middle si/se 5.3. The German Impersonal Passive 5.4. Conclusion Notes to Chapter Five

265 266 273 277 290 290

References

295

Preface

This book is a revised version of my University of Southern California doctoral dissertation, submitted in August 1985. I'd like to thank the many friends and colleagues in various places who have contributed to my work in one way or another, and who helped make the period during which I wrote this book a stimulating and enjoyable time. Space does not permit me to list everyone by name, but those singled out below deserve special mention. Of the people in Southern California I'd especially like to thank Osvaldo Jaeggli, the Chair of my thesis committee, and the other members of that committee past and present: Joseph Aoun, Mürvet Εης, Brian Loar, Bernard Comrie and David Pesetsky. I'd also like to thank the other members of the faculty of the Linguistics Department at USC, in particular Mamoru Saito, and, among the other linguists in the area: Hagit Borer, Nina Hyams, Mohammad Mohammad, Tim Stowell, Eric Wehrli and Ken Wexler. Last but not least, I'm grateful to my friends Gary Gilligan, Heather Holmback and Bonnie Schwartz. The following past and present inhabitants of Cambridge, Massachusetts, deserve special thanks for making my visit to MIT in the Fall of 1983 worthwhile and productive: Mark Baker, Noam Chomsky, Nigel Fabb, Isabelle Ha'ik, Morris Halle, Jim Higginbotham, Kyle Johnson, Juliette Levin, Rita Manzini, Diane Massam, Mario Montalbetti, Tova Rapoport, Luigi Rizzi and Richard Sproat. I'd also like to the acknowledge the help and encouragement provided by my new colleagues here in Geneva: Adriana Belletti, Carol Georgopoulos, Liliane Haegeman, Luigi Rizzi and Sten Vikner. The last word goes to Bonnie Schwartz for all her love, kindness and understanding. Geneva, July 1986

Chapter One

Background

1.1. INTRODUCTION

In this study, we take the central question for linguistic theory to be: what is knowledge of language? The question can be divided into two parts: (i) what does it mean to say of some individual that he/she knows English, French etc.? (ii) where does this knowledge come from? Chomsky (1986) answers these questions as follows: (i) knowledge of a language means the possession of a particular mental faculty, called an I(nternal) language; (ii) the I-language is assumed to be the final state in a process of development from an initial state of language which is a species characteristic of humans present at birth. In other words, the initial state is an innate endowment. The assumption that the initial state is an innate endowment accounts for the rapidity and apparent ease with which children acquire the extraordinary complexities of natural languages without formal instruction and purely on the basis of the degenerate data in the average child's environment. If the central question of theoretical linguistics is as given above, then it is the task of linguists to characterize the properties of the I-language and the initial state. Linguists' theories of given I-languages (for example, the I-language in the mind of individuals commonly referred to as 'English speakers,' 'French speakers,' etc.) are known as grammars (of English, French etc.). The linguist's theory of the initial state is known as Universal Grammar (UG). Developing UG, one is forced to reconcile two seemingly contradictory requirements: on the one hand, UG must be 'rich,' in the sense that it must be able to explain how children develop competence in their native languages so rapidly and easily. This suggests that much of the final state is predetermined by the initial state, so U G must have much in common with the grammars of individual languages. On the other hand, the diversity of the world's languages poses a challenge for any UG; it must be sufficiently 'impoverished' to allow for the attested variation. Much of the intellectual challenge in developing U G resides in reconciling these two requirements. This book is intended as a contribution to UG. In particular, the central question we address is that of accounting for certain realizations of the

2

The Representation of Implicit and Dethematized

Subjects

'semantic roles' of lexical items. The realizations we are interested in are those where a particular semantic role does not appear to have a structural realization. The constructions in question are passive, middle, nominalization and impersonal constructions, exemplified in (1): (1)

a. b. c. d.

The car was washed (by John) - passive. This car washes easily - middle. The washing of the car (was a mistake) - nominalization. Si lava la macchina ogni sabato - (Italian) impersonal. SI washes the car every Saturday 'People wash the car every Saturday.'

Wash/lavare is a transitive Verb, i.e. it has two arguments, a subject NP and an object NP. The simplest specification of the relationship between this particular argument structure and the structural frames in which transitive Verbs appear would be to say that transitive Verbs must always appear with their two arguments realized in phrase structure (e.g. John washed the car). However, in (1), wash/lavare - in their various forms - appear with only one argument, and so this maximally simple view cannot, it seems, be maintained. Three things should be noted in connection with the processes in (1): (i) the processes which relate the various one-argument instances of wash/ lavare to the two-argument instances are general, in that they are not idiosyncratic to this Verb; (ii) we can formalize these operations in terms of relations between 'logical' subjects and objects (semantic roles) and 'structural' subjects and objects (grammatical functions); in these terms, the grammatical functions of subject and object remain constant even though the argument structure changes; and (iii) most importantly for the concerns of this work, in each case the structurally 'missing' semantic role is semantically 'present'; in fact, it is understood as an indefinite of some kind in each case (moreover, in passives and nominals it is optionally structurally realized as a fty-phrase). To sum up observations (i - iii), we could say rather simplistically that the processes illustrated in (1) 'interfere with' the simplest mapping from argument structure (semantic roles) to phrase structure (grammatical functions), and interfere productively. The aim of this book is to provide a characterization of the nature of this 'interference.' Our topic, then, is the theory of argument structure and grammatical functions, where 'argument structure' means the characterization of the semantic role-taking properties of lexical items, and 'grammatical functions' are the structural realizations of these semantic roles as subjects, objects etc. We assume, following Chomsky (1965), that the grammatical functions subject, object etc. are definable in terms of phrase-structure configurations, so we can

Background

3

put our central question as follows: what is the relation between phrase structure and argument structure? Moreover, since the 'missing' semantic role is interpretable as an indefinite of some kind, and since indefinites generally involve quantification, the question can be extended to read: what are the relations between phrase structure, argument structure and the representation of quantification? We will see in what follows that the structures in (1) are an excellent probe into these issues. The proposals we make are developments of recent proposals made in the Government-Binding framework of Chomsky (1981), (1982), (1985), (1986), henceforth GB theory. In GB theory, sentences are taken to have representations at different levels. The levels of representation, and their relations, are shown in (2): (2)

D-Structure (DS) S-Structure (SS) Phonological Form (PF)

Logical Form (LF)

Formally, the mapping between levels is effected by the operation of MoveAlpha, a transformational rule. This rule operates freely, subject to a number of independent constraints which we will introduce in Chapter Two. The leading idea of GB theory concerning the relation between argument structure and phrase structure is that this relation is minimal. The argument structures of lexical items are 'projected' onto a skeleton phrase structure. The 'projection' of argument structure is subject to two fundamental constraints: (i) the θ-criterion, which roughly requires that each argument be assigned only one thematic (Θ) role ('Theme,' 'Agent,' 'Source' etc.) by some predicate and that each predicate assign its thematic roles to a unique argument, and (ii) the Projection Principle, which requires that lexically specified argument structures of lexical items be manifested in syntax. The configuration of the skeleton phrase-structure is determined by X-bar theory, which we will describe in the next section. This 'pure' representation of thematic roles in phrase structure is DS. At SS, a further condition holds: all arguments must be 'visible.' In order to be visible in the required sense, an argument must be assigned abstract Case. Since the positions occupied by θ-role-bearing items are not necessarily visible in this sense, various structural changes can take place between DS and SS. So this means that Move Alpha can alter the configurations of thematic roles, and so alter the grammatical function associated with a given thematic role at SS. The passive is the prime example of this phenomenon. In the mapping from SS to LF, structures are disambiguated in the

4

The Representation of Implicit and Dethematized

Subjects

sense that syntactic properties relevant for interpretation can be represented at LF. One of the most important such processes is the indication of the scope of various operators: quantifiers, negation, tense, mood etc. It has been assumed since May (1977) that quantifiers such as some N', every N' etc., undergo Move Alpha in the mapping to LF, to produce structures where they bind variables, rather in the manner of standard predicate calculus. The quantifier-variable relation is constrained by the Bijection Principle, which rules out both vacuous quantification and free variables by requiring that quantifiers and variables be in a bijective relation. A further condition on variables is that they be free with respect to all arguments - this is a consequence of Principle C of the binding theory of Chomsky (1981). Another important principle that holds at LF is the Principle of Full Interpretation. This principle requires that every element must be able to be interpreted. We will consider the exact form of these principles more closely as we proceed. Argument structure is preserved at SS and LF by the Projection Principle. The Projection Principle requires argument-taking properties of lexical items to be satisfied at all syntactic levels. The correct formal statement of this requirement is one of the central issues we address in this work. The above presentation gives the bare bones of current conceptions of argument structure, phrase structure and quantification. The object of this book is to put flesh on this skeleton through the study of the constructions seen in (1). In the remainder of this chapter, we will present initial formulations of the relevant principles. 1.2. X-BAR T H E O R Y

X-bar theory provides the basic configurations of phrase structure which are relevant for the way in which lexical properties are 'projected' onto phrase structure. The basic configuration for X-bar theory is as in (3): (3) X"

X°

complement to X°

Complements and Specifiers may or may not be present, subject in part to other principles, so the basic requirements of X-bar theory are as in (4):

Background (4)

a. b.

5 Categories must be headed. All X° must have a corresponding Xmax

X° projects its categorial features up to X', and so on. The last projection of a given X° is known as the 'maximal projection' of that head, written X m a \ X° heads Χ\ X1 heads Xi+1, Xmax_1 heads Xmax and so by transitivity X° heads Xmax. The major syntactic categories are Noun (N), Verb (V), Adjective/Adverb (A) and Preposition/Postposition (P). The categories result from the combination of the two categorial features [±V, ±N], as follows: (5)

Ν= V= A= Ρ=

[+N, -V] [-N, +V] [+N, +V] [-N, -V]

The insertion of lexical items into tree structures is free. X-bar theory simply requires all projections of N° to be N n , of V° to be V" etc. All lexical items are members of one of the above four categories. These are the lexical categories. There are also two non-lexical categories: Infi and Comp. The projections of these heads make up the 'S-system,' illustrated in (6): (6)

COMP m a x (= S') Specc,

COMP' COMP^

INFL max (= S) / \ NP

/

INFL

INFL'

\

VP

We will use the symbols CP and IP for Comp max and Infl m a \ following Chomsky (1985). Comp contains complementizers of various kinds and its specifier is the landing-site for wA-movement. Infi contains Tense, Agr - a bundle of features which specify agreement - and modals. The notion of maximal projection allows us to define a relation R, as follows:

6 (7)

The Representation of Implicit and Dethematized

Subjects

R (x,y) iff (VP), Ρ a maximal projection, if Ρ dominates x, then Ρ dominates y. (Aoun & Sportiche (1983:228)).

In terms of R, we can define the relations of government, c-command and R-modification: 1 (8)

a. b. c.

χ c-commands y iff χ and y are maximal projections and χ ¥ y and R(x,y); χ governs y iff χ is P°, y a maximal projection and (R(x,y) & R(y,x)); χ R-modifies y iff χ is a maximal projection, χ does not head Ρ and (R(x,y) & R(y,x)).

We make the standard assumption that R is category-neutral, therefore the structural relations in (8) are category-neutral. The structural relations are illustrated in (9): (9)

a.

C-command

ZP XP

\

XP c-commands YP YP

b.

Government

XP X governs YP YP

c.

R-modification n

Y

Y n+1 /ZP XP modifies Yn, Yn+1 XP

It is important to stress that we are using the term 'R-modification' here in the same way as the terms 'c-command' and 'government.' It denotes a purely structural relation of a particular kind, rather than a semantic relation. In fact, we could consider R-modification as defined here to be the complementary relation to government: the case where R holds between a non-head and a head. We said that the position and presence of complements and Specifiers is forced, at least in part, by principles other than X-bar theory. The

Background

7

complementation properties of a head are a function of lexical properties o f t h a t head, and so the presence of complements is forced by the Projection Principle (although their position is forced by Case Theory, as we shall see in the next two sections). Specifiers, on the other hand, do not seem to appear as a consequence of lexical properties of a head. The most important thing about Specifier position for our purposes is that it is the subject position. This is most clearly seen with IP and NP: (10)

a.

IP Spec I 1 NP I 1 John

/ I ι

\

1

can

VP I

I

V I1 swim

NP

b. Spec NP John's

N'

/

ability

CP to swim

We will develop a proposal about the principles which force subject positions to appear during the course of this work. 1.3 PRINCIPLES OF T H E M A T I C STRUCTURE

It is clear that a very important aspect of argument structure is the distribution of θ-roles in a clause, or the 'thematic structure' of a clause. In this section we introduce the principles governing thematic structure in detail. The subtheories in question are Case Theory and θ-theory. The notion which unifies these subtheories is that of θ-chain. 1.3.1. Θ-Theory and the Projection Principle In this section, we will give the assumptions that we adopt regarding Θtheory: the representation and assignment of θ-roles. Most of the assumptions we make initially are fairly standard.

8

The Representation of Implicit and Dethematized

Subjects

θ-roles are listed in the θ-grids associated with lexical items in the lexicon, as in (11): (11)

donate [ θι, θ 2 , θ 3 ] kiss [ θι, θ 2 ] smile [ θι ]

Also, Verbs have subcategorization frames, as proposed in Chomsky (1965: Chapter Two). Some sample subcategorization frames are given in (12): (12)

donate [ _ N P PP] kiss [ NP] smile [ ]

θ-grids are closely related to subcategorization frames. In fact, every Θrole except at most one in the grid is associated with a subcategorization frame (the θ-role not associated with a subcategorized position is usually assigned to the subject position - for more on the relation between subject arguments, subject position and subcategorization, cf. 2.2.2.). These Θroles correspond to notions such as 'goal-of-action,' 'agent-of-action,' etc. The nature of these notions is unclear, as is their connection with subcategorization. We will not try to shed light on these issues at this point (but cf. 2.2.2. and 4.3.); for the moment we take Θ-roles as given by lexical semantics and consider only their formal properties. Θ-roles are assigned to arguments under government, θ-roles are assigned directly by V to complements occupying subcategorized positions in VP, and indirectly by VP to the subject. The θ-criterion is the main wellformedness condition on the assignment of θ-roles: (13)

Each argument bears one and only one θ-role, and each θ-role is assigned to one and only one argument.

θ-role assignment is technically optional, but forced in many instances by the Projection Principle. The Projection Principle can be stated informally as follows: (14)

Representations at every syntactic level (D-Structure, S-Structure, Logical Form) are projected from the lexicon, in that they observe subcategorization properties of lexical items. (Chomsky (1981:29)).

We stated above that subcategorization entails θ-marking. This has two consequences, given (14): (i) that subcategorized positions must be present

Background

9

at all levels, and (ii) subcategorized arguments are obligatorily θ-marked. It also follows from (ii) that a subcategorized position will 'count as' a θ-marked position at every level. This is of particular importance where Move-Alpha applies and moves material out of that position. A simple case of NP-movement will serve as illustration: (15)

Johni was seen t,.

Here the Projection Principle and the θ-criterion require that an empty category of some kind occupy the DS position of John, namely the object position. This category receives the θ-role assigned to the position. Clearly, the moved NP, John, also receives this θ-role. The mechanism which allows this 'sharing' of a θ-role between two positions is chain formation. Chainformation is defined as follows: (16)

C = (ai...an) is a chain iff a; is the local binder of aj+i.

This definition is adopted from Rizzi (1983:2). We now need to define local binding: (17)

a. b.

a locally binds b iff a binds b and there is no c such that c binds b without binding a. a binds b iff a c-commands b and a is coindexed with b.

The informal statement of the θ-criterion in (13) can now be replaced with (18):2 (18)

Each argument appears in a unique θ-chain containing a unique θ-position P, and each θ-position Ρ is in a unique θ-chain containing a unique argument.

θ-chains are those chains for which (18) is a well-formedness condition. The Projection Principle entails that each subcategorized position will be a member of a θ-chain; this, combined with (18) as a condition on Θchains, has the consequence that movement into subcategorized positions is impossible, as the result will be a θ-chain containing two θ-positions. Subjects are not subcategorized positions (cf. Chomsky (1965:92 - 94) and the fact noted in the previous section that subject positions qua Specifiers do not appear as a consequence of lexical properties of a head), so subjects are not included under (14) as it stands. However, (14) can be restated so as to include selected positions (cf. 2.2.3). In that case, whenever an argument in subject position is selected, i.e. θ-marked, it will fall under (14). As a result, θ-marked subject positions are not possible landing sites for movement.

10

The Representation of Implicit and Dethematized

Subjects

A distinction is made between Α-positions and non-A positions. Apositions are positions which correspond to the grammatical functions subject-of, object-of etc. We define subject-of and object-of as in Chomsky (1965:71): (19)

subject-of: NP/S ("the NP immediately dominated by S") object-of: N P / V P object-of-Preposition: N P / P P

All other positions in clauses are non-A-positions, or A'-positions (on Apositions in NP, see 4.5.). All Α-positions except the subject appear purely as a function of θ-marking and subcategorization. Hence the θ-criterion and the Projection Principle are enough to guarantee their presence, and to take the place of a theory of lexical insertion of the kind proposed in Chomsky (1965, Chapter Two). The subject, however, seems always to appear, even when not filled with an argument of the predicate (in such cases it may be a landing site for movement; if not, it is filled by an expletive element, i.e. it or there). In order to force the presence of subjects independently of θ-marking and selection, Chomsky (1982) proposes the Extended Projection Principle, which adds the statement in (20) to that in (14): (20)

Every clause must have a subject.

In this way, subject positions must always be present but can be landing sites for movement when they are not θ-marked. In general, subjects are the only Α-positions that need not also be θ-positions. This is because the status of subjects with respect to the (Extended) Projection Principle is different to that of complements; subjects must be present even when not θ-marked. 1.3.2. Case Theory Rizzi (1985) distinguishes conditions on the form of categories from conditions on their content. The θ-criterion is a condition on the content of θ-chains, since it requires a biunique relation between θ-roles and Θrole-assigners. Moreover, θ-roles are ultimately semantic entities. The formal condition on θ-chains is the Visibility Condition: (21)

For any θ-chain C, C is visible iff C is Case-marked.

We can in fact incorporate (21) into the θ-criterion, by restating the Θcriterion as follows:

Background (22)

11

Each argument appears in a unique chain containing a unique visible θ-position P, and each θ-position Ρ is visible in a chain containing a unique argument. 3

Case is assigned to positions in chains under government. Any head may assign Case. Typical Case-assignments are as follows: (23)

a. b. c.

I assigns NOM(inative). V,P assign ACC(usative) or DAT(ive). N,A assign GEN(itive).

Case-assigning properties are a function of various features. Infi only assigns Case when it has the feature [+Tense]. For Verbs and Prepositions, Caseassignment depends largely on subcategorization: transitive Verbs and Prepositions have an ACC feature (in languages other than English this may be DAT - see below); ditransitive Verbs, we may assume, have two ACC features; intransitives have no Case feature (this is true also for unaccusative, which may subcategorize for objects - on this class of Verbs, see Belletti (1986)).. Chomsky (1986:270-1) makes a distinction between inherent and structural Case. Structural Case is assigned at SS subject only to the structural requirement of government. Inherent Case is assigned at DS, and realized at SS, subject both to the government requirement and to the Uniformity Condition, which is stated as follows: (24)

If X is an inherent Case-assigner then X Case-marks NP iff X θ-marks the chain headed by NP.

Thus inherent Case is tied to thematic properties of the Case assigner, while structural Case is purely structural. In English, GEN is inherent, while NOM and ACC are structural. Both the Visibility Condition and the requirement that Case is assigned under government are requirements imposed by UG. However, it is possible that in individual languages, the particular inventory of Cases and the direction in which Case is assigned may be specified in particular ways. For example, English lacks a DAT Case and assigns Case rightward, while German has DAT and assigns Case leftward. 4 The direction of Case assignment is an important determinant of word-order in a language (although almost certainly not the only one) - cf. Travis (1984), Coopmans (1985), Koopmann (1985) and Li (1985) on the general issue of wordorder variation from a GB perspective. Also, the specification of which Cases are inherent varies from language to language (cf. 5.3.). Such extra specifications of UG principles in the grammars of individual languages

12

The Representation of Implicit and Dethematized

Subjects

are known as parameters. So we say that the UG requirement that Case be assigned under government has an associated parameter of directionality. Also, the listing of Cases is subject to parametric variation. Parameters are fixed, it is assumed, during the course of acquisition, either on the basis of direct evidence or through interaction with other parameters which are determined on the basis of direct evidence. The question of parametric variation in passive-like constructions will be broached in Chapter Five. 1.4. PRINCIPLES OF QUANTIFICATION

The central principles of quantification concern the representation of quantifiers and variables and the relations between them. As in the previous section, the principles we will adopt initially are fairly standard. Scope relations among quantifiers and other elements (e.g. negation) are represented at LF. In the mapping from SS to LF, Move-Alpha can apply. In one form, this operation moves a quantified NP to an A'-position, from which it binds a variable in the Α-position vacated. The A'-position in question is the position adjoined to IP or VP (although there is also movement-to-Comp - wA-movement - at LF). This variant of Move-Alpha is known as Quantifier Raising, or QR (see May (1977), (1985)). Scope is determined as follows: (25)

X is in the scope of Y iff X is c-commanded by Y.

Consider a sentence containing two quantified NPs, like (26): (26)

Every man loves some woman.

QR can apply to the two NPs here in either order, so a sentence like (26) corresponds to the two LF representations in (27): (27)

a. b.

[ every mani [ some womanj [ ti loves tj]. [ some womanj [ every mani [ ti loves tj].

(27a) corresponds to the reading with a wide-scope universal, and so it can be true if every man loves a different woman, while (27b) corresponds to the reading with a wide-scope existential and so is true only where the same woman is loved by every man. QR is a form of extraction, and extraction involves operator-variable relations. We adopt the theory of these relations of Koopmann & Sportiche (1982). In particular, we adopt their Bijection Principle: (28)

There is a bijective correspondence between variables and operators.

Background

13

We take this principle to hold both at SS and LF. It holds in slightly different forms at the two levels; at SS, it is a condition on representations, while at LF it is a semantic condition on variables. The two forms of the principle are: (29)

a. b.

SS: Every variable must be bound from an A'-position. LF: Every variable must be bound by an operator, and every operator must bind a unique variable or chain of variables.

By "chain of variables" here we mean a sequence of (ai .. a n ) of variables, such that ai locally binds ai+i. Note that it follows from Principle C that each member of the variable chain except the "coda" of the chain must be in an A'-position (if that variable is an empty category; for the case where the variable is a pronoun, see below). There are two kinds of variable: (i) empty categories of a particular type and (ii) pronouns. The empty categories allowed by the theory exhaust the possible values of the features [ ± anaphor, ± pronominal], as follows: (30)

a. b. c. d.

[+anaphor, +pronominal]: PRO, as in He tried ec to leave. [+anaphor, -pronominal]: anaphor, as in He seems ec to be smart. [-anaphor, +pronominal]: pro, as in ec viene "(He) comes." [-anaphor, -pronominal]: variable, as in Who did you see ec?

The features [± anaphor, ± pronominal] encode which principle of the binding theory a given empty category is subject to. The binding principles are as follows (where "A-bound/free" means bound/free with respect to A-positions): (31)

A. B. C.

Anaphors must be Α-bound in their binding domain. Pronominals must be Α-free in their binding domain. Names must be A-free.

A [+anaphor] element obeys Principle A, e.g. NP-trace and PRO, and a [+pronominal], e.g. pro or PRO, obeys Principle B. The contradictory requirements on PRO prevent this element from having a binding domain. Since the definition of binding domain involves government (cf. Chomsky (1981)), it follows that PRO must appear in an ungoverned position. 5 Variables, being [-anaphor, -pronominal], are subject to Principle C. Hence variables cannot be bound by an element in an Α-position. This is what rules out classic cases of strong crossover, like (32):

14

The Representation of Implicit and Dethematized

(32)

*Whoj does hei love ti.

Subjects

Here the variable, ti, is Α-bound by Ae, in violation of Principle C, so the sentence is impossible with this reading. Pronouns can be interpreted as variables. This happens where a pronoun is coindexed with an inherent quantifier - a quantifier such as every or no·, NPs headed by or specified by inherent quantifiers are "not satisfiable by one or more objects in the domain of discourse" (Hai'k (1984:189)). For example, no object satisfies the description no girl, every man, anything etc. For interpretation of the pronoun to be possible, the quantifier must structurally bind the pronoun, i.e. c-command it as well as be coindexed with it. (33) is an example of a pronoun acting as a bound variable: (33)

Everyonei loves hisi mother.

The interpretation of (33) is roughly "every χ loves x's mother," and not "the entity everyone loves its own mother." Note that in the LF for this sentence the pronoun is also bound by the variable bound by everyone, as everyone undergoes QR. This variable occupies an Α-position, but the sentence is grammatical, so Principle C does not hold for pronouns as variables. It is precisely because of this that the pronoun and the variable here can form a variable chain. On the other hand, the Bijection Principle does hold for pronouns as bound variables; this is what underlies Koopmann & Sportiche's account of weak crossover: (34)

Who; does hisi mother love ti?

Here both hisi and ti are variables bound by who. Since they do not ccommand each other, they do not form a variable chain, and hence who binds two variables, violating (29b). Hence (34) is ungrammatical on the interpretation represented. The above sections sketch the principles we adopt initially. Many of them will be refined in the light of data from passive and related constructions that we will consider in the chapters to follow.

NOTES 1. (8a,b) are adapted from Aoun & Sportiche (ibid:228). (8c) is based on the definitions of modification in Zubizarreta (1982) and Marantz (1984). Chomsky (1985:7) introduces "m-command" for Aoun & Sportiche's "R." This paper also introduces a rather different concept of government, relying on m-command and the notion of "barrier." Barriers are introduced in order to unify government and bounding. This is not a central concern here (but cf. 3.3.), so we retain the Aoun/Sportiche definitions.

Background

15

2. This formulation of the θ-criterion differs from that in Chomsky (1986) in certain important respects. First, we add the word "unique" before the first occurrence of the word "θ-chain." This rules out double θ-marking of a single chain, as in (i): (i)

*John seems t to be raining angry.

In fact, we also do not propose that angry θ-marks John in (ii), leading us to an account of predicative Adjectives which is different to Chomsky's (cf. 3.2.3.): (ii)

John left angry.

Secondly, we do not incorporate the Visibility Condition into the statement. This is done in the next section. Third, we explicitly refer to θ-chains, as we allow other kinds of chains for which the θ-criterion does not hold. 3. A major problem for both (21) and (22) is the existence of θ-chains headed by PRO. As PRO must be ungoverned (see 1.4. and fn. 5), it cannot be Case-marked. However, it clearly heads chains: (i) (ii)

PROi to appear fc to have been killed ti would be useful. John tried PROi to get ti signed up ti for the part.

To account for these examples, we could exploit the fact that PRO always needs an antecedent {John in (ii) and some arbitrary element in (i)) and say that PRO forms an extended chain with its antecedent (where an extended chain is defined following Chomsky (1985) as "C = (ai ... a„, b) is an extended chain if (ai, a„) is a chain with index i and b has index /'). The extended chain has two θ-roles, so this approach necessitates keeping (18) and (21) as separate conditions, rather than combining them as in (22). In this way, a θ-chain headed by PRO can meet the Visibility Condition. In cases like (ii), the antecedent will always clearly be Case-marked as it is an argument, hence the Visibility Condition is met through obligatory control. Cases of "arbitrary P R O " like (i) seem to pose more of a problem; however, we will see in 3.2.4. that in at least an important subclass of these cases PRO has an antecedent whose properties are predicted by principles needed to account for the syntax of adjuncts. This antecedent must also be visible, so again the Visibility Condition will be met. On why PRO is able to appear in a chain with two θ-roles, see fn. 5. 4. Cf. Kayne (1984) for the proposal that English lacks D A T while French has this Case, and an investigation of the consequences of this proposal, and Roberts (1985) for an analysis of how D A T was lost in the history of English. In general, it seems that Case Theory is the locus of considerable parametric variation. 5. This is the conclusion of Chomsky (1981). However, it has not gone unchallenged, cf. in particular Bouchard (1984), who argues that the correct generalization is that PRO cannot be Case-marked. More recently, Chomsky (1986) has defined binding domain in terms of Complete Functional Complex (CFC). A C F C for a head is the domain containing chains to which all that head's θ-roles are assigned (cf. 2.2.3.(106)). In these terms, the contradictory requirement on PRO might amount to the requirement that PRO not be part of a single CFC, i.e. in the case of obligatory control that it form a chain with some θ-marked element outside the C F C it structurally appears in (there is a further locality requirement at work here - cf. Manzini (1983a)). Hence PRO must appear in a chain which has two θ-roles. Since each chain must be uniquely Case-marked, it follows, given the remarks in fn. 3, that PRO must appear in a non-Case-marked position. If we are to maintain such an approach to control theory, we must add that the head of each chain and no other position be Casemarked in order to rule out sentences like (i):

16 (i)

The Representation of Implicit and Dethematized

Subjects

It was persuaded Johni that PRO* left.

The suggestions here and in fn. 3 are perhaps candidates for the beginnings of a chainbased theory of control. However, more must clearly be said about deriving the PRO theorem for non-obligatory-control - cf. 3.3. In this matter we follow Clark (1985) in proposing that non-obligatory-control PRO is really a kind of empty operator. Clark proposes that these operators move into the nearest Comp. He also points out that empty operators cannot appear with Case in a Case-marked Comp. We can derive this result if we say that PRO forms an extended chain with the nearest Comp, and this Comp is Case-marked. Hence PRO cannot appear in a Case-marked position. Furthermore, if the links of this chain are subject to a government requirement (cf. 2.2.4. on chains formed by heads), then PRO must appear in subject position as this is the only Α-position governed by Comp. Note that in fact non-obligatory-control PRO always appears in a Case-marked domain: (ii)

a. b.

*[[PRO to shave/shaving each other] to be frowned upon] struck me as oldfashioned. *[[For[PRO to shave/shaving each other] to be frowned upon] struck me as old-fashioned.

Chapter Two

Passive and Argument Structure

We begin this chapter by discussing certain central properties of passives. We then illustrate the explanation in Chomsky (1981:2.7.) for the relation between active and passive clauses. This explanation is inadequate in three main respects: in terms of the categorial, Case and θ-properties of passives. We then propose a new analysis of passives which overcomes these inadequacies. This proposal, whose central claim is that the "understood" subject of a passive is structurally present, has a number of empirical and theoretical consequences. Chapter Three considers the empirical consequences of the proposal at length, while the second part of this Chapter deals with the theoretical aspect. 2.1. PASSIVE IN GB THEORY

2.1.1. The Core Passive Example (1) is an instance of a "core" passive/active alternation with a transitive Verb: (1)

a. b.

John kissed Mary. Mary was kissed (by John).

Kiss is a transitive Verb. Transitive Verbs are two-argument Verbs; therefore they can appear at S-Structure with both a subject and an object. Comparing (1) with (2), we can then see that, while kiss is a transitive Verb, none of the Verbs in (2) are transitive: (2)

a. *John danced Mary. b. *John seemed Mary. c. *John arrived Mary.

These Verbs are intransitives, i.e. they are unable to appear with an object. The passivized form of kiss, as in (lb), seems to pattern like an intransitive Verb in not being able to appear with an object:

18 (3)

The Representation of Implicit and Dethematized

Subjects

a. *Mary was kissed John, b. *John was kissed Mary.

So we can conclude that passivizing a Verb has the effect of rendering it intransitive. Not only does passivizing a Verb make it intransitive, the subject of a passive Verb corresponds to the object of an active Verb. This can be seen from examples of Verbs which impose differing selection restrictions on their objects and subjects: 1 (4)

a. John admires sincerity. b. Sincerity is admired (by John). c. *John is admired (by sincerity).

(5)

a. John built a house. b. A house was built (by John). c. * John was built (by a house).

(6)

a. John saw the mountains. b. The mountains were seen (by John). c. *John was seen (by the mountains).

(7)

a. John learned Spanish. b. Spanish was learned (by John). c. *John was learned (by Spanish).

The subject of the active is optionally realized in a 6^-phrase, as shown in the (b)-examples here. It can also be apparently unrealized. A third feature of passives is that active and passive instances of the same Verb differ in their form. There is considerable variation in the exact form this difference takes, but in general it consists in the attachment of an affix, usually an allomorph of / e d / , as in the examples given so far. In some cases, the affixation is accompanied by a change of stem vowel: buy/bought, sing/sung etc. We will gloss over these morphophonemic differences for the most part, and refer to all of these variants as the result of the attachment of the affix en to the Verb-stem. Fourth, the auxiliary be appears in passives. This auxiliary carries tense and agreement information. A common view in generative grammar has been to regard be...en as a discontinuous passive morpheme (cf. Chomsky (1957)), with en attached to the main Verb by means of a rule of Affix Hopping. No account of the be...en relation is offered in GB theory, so we will have little to say about this until section 2.2., where our alternative is developed.

Passives and Argument

Structure

19

In fact, en also attaches to a much wider class of Verbs than just the class of transitives. The possibilities are illustrated in (8): (8)

a. b. c. d.

John John John John

was was was was

given a b o o k / A book was given (to) John. believed to be a genius. considered smart. spoken about.

(8a) shows that ditransitives can passivize, (8b,c) that Verbs without thematic NP objects which Case-mark the subject of their complement clauses (Exceptional Case-marking (ECM) Verbs) can passivize, and (8d) shows that objects of subcategorized PPs can passivize. The generalization which unites these examples is that the first NP Case-marked by an active Vi cannot appear following V;+en. (This means that spoken and about must be considered as a morphologically "complex Verb" in (8d) - we return to this point in 3.3.2. Cf. also Kayne (1984).) So, we have established that passive Verbs have three properties that distinguish them from their active counterparts: (9)

a. b.

c.

Passive Verbs are Case-intransitive, i.e. they cannot be followed by the NP Case-marked by their active counterparts. The subject of a passive Verb is the Case object of the corresponding active; the subject of the corresponding active is either absent or appears in a 6y-phrase. Passive Verbs are formed from actives by affixation of en.

We next give Chomsky's account of what underlies these three properties. 2.1.2. Categorial Properties of Passives The properties in (9) are all accounted for in the LGB analysis in terms of Case Theory and θ-theory. Chomsky's proposal in LGB (pp. 124 - 7) correlates the properties of passive constructions with the following claims about Case- and θ-role assignment in passives: (10)

a. b.

N P / S is not a θ-position in passives. N P / V P is not assigned Case in passives, for some choice of NP in VP.

Given (10), consider a D-Structure like (11): (11)

e (was) kiss+en Mary.

20

The Representation of Implicit and Dethematized

Subjects

D-Structure is a pure representation of argument structure. We will refer to property (10a) as "subject dethematization." Because of this property, the subject position is not occupied by an argument NP at DS. It is a consequence of (10b) that the NP Mary in (11) is in a position which is not Case-marked by the Verb, although it is governed and Θmarked by the Verb. The S-Structure in (lb) is derived from (11) by MoveAlpha, forced by the Visibility Condition. If movement does not take place, the S-Structure will be (12): (12)

*It was kissed Mary.

(12), however, violates the Visibility Condition, since the chain {Mary) does not contain a Case-marked position.. In this way, movement to N P / S position is "forced," yielding the S-Structure in (13): (13)

Mary; was kissed tj.

We can see, then, that the properties in (10) are enough to derive the core cases of passives. Two questions naturally arise at this point: (i) what exactly are the properties in (10) properties of! (ii) Are they related? The answer to the first of these questions is that the properties in (10) are taken to be properties of passive morphology, the en morpheme which attaches to passive Verbs. This answer also provides us with an account of the role played by the passive morphology en. More precisely, en is taken to be of category [+V], i.e. neutralized between Verb and Adjective (this idea originates in Rouveret & Vergnaud (1980) - on Adjectival passives see 4.4.). It is commonly assumed in work on morphology that the categorial status of a complex word [Stem + Af] is determined by the categorial status of the affix, which is therefore said to head the word (cf. Lieber (1980), Williams (1981)). So, if en is [+V], the derived category of a passive participle, i.e. of [ V + en ], will be [+V] also. 2 This was related to Caseassignment in earlier versions of the theory by the following statement (cf. Chomsky (1980)): (14)

Case assigners are [—N].

Since we allow all categories to assign Case (cf. 1.3.2.(23)) (14) is inconsistent with our assumptions. However, we could retain the spirit of the proposal by saying that structural Case assigners must be [-N]. Adjectives and Nouns are [+N], and these categories assign inherent Case. On the other hand, a transitive Verb is [+V, - N ] and therefore able to assign structural Case. Their passive participles, however, do not have the feature [-N] if en is [+V] only, i.e. unspecified for [-N] (cf. fn. 2). So the assumption that en is [+V] derives (10b).

Passives and Argument

Structure

21

Regarding question (ii) above, consider the following statement, known as "Burzio's generalization:" (15)

If some NP governed by V is assigned no Case, then the VP of which V is the head assigns no θ-role.

We are interested in the following statement, which is entailed by (15):

(16)

If a transitive Verb assigns no Case to its object, then it assigns no θ-role to its subject.

Since en prevents the passive Verb from assigning Case to its object, (15) means that the subject cannot be assigned a θ-role. (15) is a generalization and not a principle, so to the extent that it is a valid generalization, we want to derive it from some principle or the interaction of some principles. Chomsky derives (16) as follows. Assume that (10b) is the fundamental property of passives, derived from the category of en and Case theory, as just described. Then movement from object position must always take place in passives, to avoid violating the Visibility Condition. The only position to which movement is possible is subject position. Because of the interaction of the θ-criterion and the Projection Principle (cf. 1.3.1.), the subject is the only Α-position which need not be a θ-position, hence the subject position is the only position which can be dethematized in the sense of (10a). So subject dethematization is a consequence of (10b); if the subject position were not dethematized when Case is "absorbed" inside VP, the object NP would either remain in its D-Structure position, violating the Visibility Condition at S-Structure, or move to a θ-position, violating the θ-criterion at S-Structure and LF. Thus subject dethematization is entailed by the Case-absorbing property of the affix en, property (10b). (10b) is a consequence of morphological and lexical properties of en. So the properties of passives are deduced from (17): (17)

en is [+V],

The most important step in the deduction of passive from (17) is the assumption in (14). We will now consider these two assumptions in more detail. Here we will show that assumption (17) in conjunction with a wellknown condition on affixation leads to empirical problems, and in the next section we will see that certain V + en complexes must be allowed to assign Case, challenging (14). The evidence that en does not change the category of participles to

22

The Representation of Implicit and Dethematized

Subjects

[+V], and therefore is not itself [+V], comes from the range of Verbs which cannot be passivized. There are three main classes of "nonpassivizable" Verbs in English, shown in (18): (18)

a. b. c.

Raising Verbs: *John was seemed to to be smart by Harry. Intransitive Verbs: *It was danced by everyone. Passive Verbs: *The book was been given by Mary by John.

We can rule out (18c) by means of the No Vacuous Affixation Principle, which we formulate as follows: 3 (19)

No affix can attach to a stem if the attachment does not result in a derived word whose features differ either in kind or value from the features of the stem.

As passive participles are [+V], (19) will prevent en from attaching to the participle in (18c), since this attachment will not result in a derived word whose features differ in kind or value from those of the stem. This is clearly a desirable result, but one which cannot be extended to the other impossible passives of (18). In fact, there is no obvious way to rule out (18a,b), i.e. to prevent attachment of en to intransitive Verbs and raising Verbs. This is unfortunate as the examples seem to form a natural class, and we would therefore like to account for them in a unified way. To distinguish intransitive Verbs and raising Verbs from transitive Verbs, and prevent the first two of these classes from being changed from [+V, - N ] to [+V] by attachment of en, we must make reference to argumenttaking properties of these Verbs. This means that attachment of the category-changing affix en must be sensitive to Case, θ-role-assigning or subcategorization properties of the Verbs. In fact, if we want to rule out double passives in the same way as passives of raising and intransitive Verbs, reference to subcategorization properties is inadequate, as passive Verbs have the subcategorization features of transitives. So the rule attaching en to Verbs must refer to either Case or θ-role-assigning properties of the stems which it attaches to. The category-changing property of this affix is independent in principle of issues of Case and θ-role assignment, so if we maintain that this alone is the core property of en, there seems to be no way of ruling out (18a,b). To rule out (18), then, we must add some specification about the Case or θ-properties either of en or of the Verbs to which it can attach. This suggests that the crucial property of en is directly implicated in argument structure. A good candidate is (20): (20)

en adds the feature [-ACC Case-assigner] to the stem it attaches to.

Passives and Argument

Structure

23

But any such additional specification of en's properties will do all the work of (17), in forcing NP-movement and subject dethematization. So, on general grounds of simplicity, (17) should be abandoned in favor of (20).

Moreover, if the No Vacuous Affixation Principle is retained in the form above, we must abandon (17) for empirical reasons. If en's property were the addition of the two features [-ACC Case-assigner] and [-N], then en could still attach to the Verbs in (18a,b). However, if en only has the property in (20), and if attachment of a negative feature does not differ in kind from absence of that feature for the purposes of (19), then (19) will rule out (18a,b). Hence a negative Case feature not only replaces the categorial feature, but also avoids problems caused by the categorial feature. Let us then tentatively replace (17) with (20) (cf. Marantz (1984), Perlmutter & Postal (1983) for discussion and accounts of this kind of data - the former relying on the No Vacuous Affixation Principle). A further reason for not taking en to be [+V] comes from the theory of categories. Saying that en is [+V] implicitly allows for the neutralization of the categorial features [± Ν, ± V]. This increases the number of major categories from four to eight. As we want to constrain the categorial component as much as possible on general grounds, this seems like an undesirable move, one which we should not make unless we really have to. Also, I know of no good candidates for the other neutralized categories whose existence is entailed by the existence of a [+V] morpheme: [+N], [-N] and [-V]. 4 For this reason too we replace (17) with (20). In the next two sections, we will see that both the Case-theoretic and θ-theoretic aspects of the account of passive given in 2.1.1. need some modification. This will lead us to propose a radically different view of en from that in either (17) or (20). 2.1.3. Case Properties of Passives (20) embodies the idea that attachment of en to a Verb-stem results in a derived Verb which cannot assign Case at all. In this section, we will question this idea. We focus on ditransitive Verbs, e.g. give. In (21), a book receives Case: (21)

John was given a book.

(It is in fact to allow for such examples that (10b) contains the clause "for some NP in VP"). (20) is superficially falsified by examples like (21). Various analyses have recently been proposed for the structure of complements to ditransitive Verbs, notably Stowell (1981) and Kayne (1984). We will only discuss Stowell's account here. 5 Stowell proposes that the structure of double object constructions is as in (22):

24 (22)

The Representation of Implicit and Dethematized

Subjects

[VP [v give - John ] a book ]

In (22), John has undergone NP-Incorporation into the Verb. The existence of passives like (21) appears to falsify such an incorporation analysis, but Stowell suggests that (21) is in fact the passive of the /o-dative construction. NP-movement is forced because to, a dummy Preposition which appears purely for Case-marking purposes, fails to appear. So, on Stowell's account of (21), (20) is irrelevant to Case-absorption. Subject dethematization is nevertheless forced by the θ-criterion (as in the LGB account - see 2.1.2); this suggests that the appearance of a passive participle in (21) is related to subject dethematization in some way independent of Case. So Stowell assigns the S-Structure in (23) to (21): (23)

Johns was [J-+vJ give + en ] a book ti.

Stowell must say that a book is Case-marked by give here. This move then makes (21) a straightforward counterexample to (20),6 and leaves subject dethematization as the only property of en. A possible way out for Stowell would be to claim that the Theme NP, i.e. the book in (23), is Case-marked in some other way (perhaps inherently) here. This account would allow (24): (24)

*It was given a book to John.

Here a book is Case-marked in the same way as in (23). The dummy Preposition to appears and Case-marks John, so this NP does not need to move to subject position for Case reasons. En, as we saw, only dethematizes the subject in (22), and does exactly this in (24) also. So nothing rules out (24). However, the sentence is clearly ungrammatical. Hence this approach cannot be right. An alternative to Stowell's approach is to regard to as subcategorized by give rather than as a dummy Case-marker, and say that (21) is allowed because en takes away exactly one Case from the Verb. So now we reformulate (20) as (25): (25)

En attaches to a Verb-stem with η Cases to assign and produces a derived Verb with n-1 Cases to assign.

This approach predicts that either object of a ditransitive can be passivized, and so (at least for the majority dialect - cf. fn. 6) needs further elaboration. However, the problems noted with Stowell's account are avoided. Further support for such an approach comes from Kayne (1984). Kayne notes the following contrast:

Passives and Argument (26)

Structure

25

a.

John Smith, who I assure you to be the best student in the class,... b. *I assure you John Smith to be...

To account for these facts, Kayne assumes that assure can Case mark an element in the embedded Comp, but not the embedded subject position. This means that (26b), with John Smith in subject position, violates the Visibility Condition. What is relevant to our concerns is the fact that the same contrast exists in the passive counterparts to (26): (27)

a.

John Smith, who I've been assured to be one of the best students in the class... b. *I've been assured John Smith to be....

Kayne concludes: "the passive participle assured must be capable of assigning Case [footnote omitted] (into Comp)" (p. xiii). (25), but not (20), clearly allows this, if Kayne's proposal that assure assigns Case into the lower Comp is retained. In (26) assure assigns two Cases: one to its object and one into Comp. Given (25), we say that in (27a) assured assigns only one Case; the Case into Comp. Consequently, I must move to subject position for Case. In (27b) assure assigns its Case in the same way; here, as in (26b), John Smith, occupying the embedded subject position, does not receive Case (presumably this NP gets Case by forming a chain with the variable in Comp when it occupies the topic position, hence the grammaticality of (26a) and (27a)). So ditransitives of various kinds provide evidence in favor of (25) over (20). (25) retains the result provided by (20) for the examples in (18): if attachment of en reduces the number of Cases a Verb can assign by one, then attachment to a Verb which assigns no Cases at all will be vacuous, and so prevented by (19). However, we still need to explain the contrast in (28): (28)

a. *It was seemed that the earth was flat, b. It was believed that the earth was flat.

To maintain (25), we must assume that believe assigns a Case to its CP complement in the active, while seem does not. Is there some further evidence for this? In fact there is independent evidence for dividing the class of CP-taking Verbs into "transitives" (Case-assigners) and "intransitives" (non-Caseassigners). First, Verbs like believe, which we propose are transitive, have thematic subjects:

26 (29)

The Representation of Implicit and Dethematized

Subjects

a. John believes that the earth is flat, b. *It believes that the earth is flat.

So believe Case-marks its object and θ-marks its subject. Seem, on the other hand, does neither: (30)

a. *John seemed that the earth was flat, b. It seemed that the earth was flat.

So sentential complements obey Burzio's generalization. Exactly those complement-taking Verbs which have thematic subjects can passivize. According to (25), this is because they also assign Case. Further support for this view comes from Chomsky (1986:195-6). Assuming that A'-chains, like other chains, are subject to the Visibility Condition, and that a matrix Comp is not a Case marked position, we expect extraction to be possible only if the extraction site is Case-marked. That this expectation is correct is shown by (31): (31)

*Whoi does it seem ti to leave?

In (31), the extraction site is neither governed by Agr or a Case-assigning Verb, so the sentence is ruled out. Given this evidence, we expect that the CP-taking Verbs which can passivize allow the CP to be questioned in the active because they can Case-mark the extraction site, while those Verbs which cannot passivize will not allow extraction of the CP. (32) shows that this prediction is correct: (32)

a. Whati did you believe ti? b. *Whatj did it seem ti?

Also, the passives of Verbs like believe should not allow extraction: (33)

* Whati was it believed ti?

Presumably, in examples like (28b) and (30b) the CP receives Case by forming a special kind of chain (CHAIN - see 5.3.) with the Case-marked expletive subject it. Extraction of elements in such CHAINs is in general impossible (cf. * Who did there seem to have been killed). In this section, we argued that (25) is the property of en which underlies the properties of passives discussed in 2.1.1. and 2.1.2. This is not our last word on the subject, but for the moment we will put aside Case and turn to the thematic properties of passives.

Passives and Argument

Structure

27

2.1.4. ^-properties of Passives In this section, we reconsider the claim in (10a): subjects of passives are dethematized. The claim that subjects of passives are dethematized means that the θ-role lexically specified as being assigned to the subject is somehow "suppressed." We saw in 2.1.2. that Chomsky proposes in LGB that this is forced by the Visibility Condition. However, there is evidence that the argument assigned this θ-role is present in passives. The clearest evidence comes from the possibility of 6j>-phrases. The Z>j>-phrase has a θ-role which corresponds to the θ-role assigned to the subject of the corresponding active. So, any θ-role that can be assigned to the subject of an active can be assigned to the öy-phrase. This is illustrated by the following examples (the first four of which are taken from Jaeggli (forthcoming^: 13):7 (34)

a. b. c. d. e. f.

Bill was killed by Mary. (Agent). The letter was received by Bill. (Goal). The package was sent by John. (Source). That professor is feared by all his students. (Experiencer) The window was broken by the hammer. (Instrument). The painting was noticed by the thief. (Theme).

Moreover, in a range of quasi-idiomatic cases the 6j-phrase has exactly the θ-role of the subject of the corresponding active, as the following examples show: 8 (35)

a. b. c. d. e.

A baseball was thrown by Fernando. Support was thrown behind the candidate by the CIA. The match was thrown by the prizefighter. The party was thrown by the department. A fit was thrown by the Countess.

(36)

a. b. c. d. e.

A book was taken from the shelf by John. The bus was taken to New York by Mary. A nap was taken by the professor in his office. An aspirin was taken by John for his cold. A letter was taken in shorthand by the secretary.

(37)

a. b. c. d. e.

The The The The The

cockroaches were killed by the fallout. conversation was killed by the linguist. evening was killed by John watching TV. bottle was killed by the wino. audience was killed by the witty repartee.

28 (38)

The Representation of Implicit and Dethematized

Subjects

a. The cat was kicked by John. b. ?The habit was kicked by Mary. c. *The bucket was kicked by Fred.

Noncompositional idioms cannot passivize, as is well-known, and shown in (38c). This is due to the fact that non-compositional idioms are really single words in the lexicon and so cannot be broken up by movement rules without violating the Lexical Integrity Condition of Lapointe (1980). This condition basically prevents lexically entered forms from being broken up by syntactic rules (note that this does not prevent words from being syntactically derived, e.g. by Affix-Hopping). If non-compositional idioms are lexically entered (and note that the lexicon seems to be the only place where their meaning could be specified) then this principle prevents their being broken up by passivization. Furthermore, the ^ - p h r a s e must receive a θ-role, as the ungrammaticality of (39) shows (this example is also taken from Jaeggli (forthcoming,a)): (39)

*The man was killed by it/there.

How is the 6y-phrase θ-marked? Zubizarreta (1985) proposes that the byphrase is an adverbial modifier in passives. She says: it [the fty-phrase - IGR] is interpreted very much like other adverbials. In (15a) the locative adverb under the blanket restricts the meaning of the locative argument in the closet. In the same manner, the agentive, adverb by the police in (15b) restricts the referential value of the agentive argument of kill, which is syntactically unexpressed and is interpreted indefinitely: "John was killed by someone and this someone is the police". (15)

a. b.

John put the box in the closet under the blanket. John was killed by the police.

(p. 254).

However, Zubizarreta must claim that agentive 6^-phrases are distinct from other kinds of agentive adverbial in that her proposed method of interpretation is only available for agentive ^ - p h r a s e s in passives. Compare (40a) with (40b,c): (40)

a. b. c.

John hit Bill by Mary. John deliberately hit Bill. John hit Bill to impress Mary.

(40a), although grammatical, has no agentive interpretation for the byphrase. This ό^-phrase can only be locative. However, the Adverb in (40b) and the purposive in (40c) can both be interpreted exactly as in passives,

Passives and Argument Structure

29

namely as agentive modifiers (cf. 3.2.1.,3.2.4. for a characterization of the notion "agentive modifier"). This casts some doubt on Zubizarreta's proposal for ö^-phrases. To support her proposal, Zubizarreta points out that ^-phrases do not act like arguments with respect to compound formation. Lieber (1983) shows that the argument structure of argument-taking lexical items in compounds is always satisfied either entirely internally to the compound or entirely external to it. This proposal predicts the ungrammaticality of (41): (41)

*candy-handing to children.

Candy and children are both arguments of hand, and so it is impossible for one of them to appear externally and the other internally to the compound. Hence (41) is ungrammatical. .ßy-phrases do not obey this generalization, however: (42)

cigarette-smoking by children revenue-sharing by local communities

However, as Jaeggli (ibid'AS) points out, since these examples do not involve en it is not obvious that they are relevant to the issue of the passive byphrase at all. Instead, the fry-phrases in (42) might be instances of the clearly agentive ύ^-phrase found in nominals, in examples like the destruction of the city by the barbarians, a play by Shakespeare, etc. A characteristic of fry-phrases in nominals is that they are only agentive. Compare (43) with (44) in this respect:9 (43)

a. b. c.

danger was sensed by John. Bill was feared by Marty. Mary was respected by John.

(44)

a. *danger's sense by John b. *Bill's fear by Marty c. *Mary's respect by John

If Jaeggli's suggestion is correct, then we expect a 6y-phrase to be ungrammatical with compounds based on the head Nouns of (44). This is true, as (45) shows:10 (45)

a. *vibration-sensing by psychics b. * spider-fearing by arachnophobes c. *intellectual-respecting by the gullible

30

The Representation of Implicit and Dethematized

Subjects

The contrast between (45) and (43) shows that the properties of fry-phrases in compounds are different from those of 6^-phrases in passive clauses. Hence the argument based on how examples like (42) fail to meet Lieber's condition on compound formation does not tell us that passive ftj'-plirases are not arguments. Another argument of Zubizarreta's that the 6>>-phrase in passives is not an argument centers on the impossibility of coreference between an NP inside the 6y-phrase and some other NP in the sentence. This is illustrated in (46) and (47) (Zubizarreta's (24)): (46)

*HiSj

mother

(47)

*Hisi role is played by John;.

is

loved by John].

Zubizarreta proposes the following constraint on coreference (her (27)): (48)

If X is an argument of Ζ and Y is an adjunct of Z, then X cannot be referentially dependent on Y.

(48) receives independent support from the impossibility of coreference in examples like (49) and (50): (49)

*Mary went to his, house with Johni.

(50)

*Mary sang hisi song to Sally with Johnj.

We will discuss these cases in more detail in 3.3.4. For the moment, we simply note that we are claiming that the 6^-phrase is thematically an argument; it is clearly structurally an adjunct. We will suggest in 3.3.4. that Zubizarreta's constraint in (47) holds and the relation "adjunct-of" is structurally, not thematically, defined. If this is so, then (46) and (47) are not evidence that the ύ^-phrase is a thematic adjunct. Given the evidence in (34 - 38) that the fry-phrase receives exactly the θ-role of the subject of the corresponding active, Zubizarreta's view seems somewhat implausible. We conclude that the simplest proposal for passive 6^-phrases is that they somehow get the θ-role assigned to the subject in actives. We will see how this happens in 2.2.; for the moment the relevant conclusion is that subject dethematization, in the sense of suppression of the subject argument, does not take place obligatorily in passives, at least to the extent that a 6y-phrase can be present. Even where the 6y-phrase is not present, there is evidence that an argument which corresponds to the subject of an active is present. This argument has been referred to as an "implicit argument" (cf. Chomsky

Passives and Argument

Structure

31

(1986), Jaeggli (forthcomings), Roeper (1983), Zubizarreta (1985)). In fact, passives contrast with middles, another construction where subject dethematization is often argued to take place (see for example Keyser & Roeper (1984) and Chapter Four), with regard to the presence of "implicitargument" effects. This contrast is illustrated in (51) and (52): (51)

a. The book was sold deliberately, b. *The book sold deliberately.

(52)

a. The book was sold to make money, b. *The book sold to make money.

We can account for these contrasts if we say that Adverbs like deliberately and rationale clauses require the presence of a nonderived "subject argument" (this idea will be refined and explained at length in Chapter Three). If we say this, then we have to say that such a subject argument is suppressed in middles but not in passives. So it seems that (10a) needs to be reformulated. Instead of being suppressed, it appears that the Θrole assigned to the subject position in actives is somehow "reassigned" in passives either to an overt Zry-phrase or to an empty category in the position occupied by the 67-phrase. On the basis of our discussion so far, we can restate (10) as follows: (53)

a. b.

The θ-role assigned to subject position in actives is reassigned to an implicit argument in passives, En attaches to a Verb stem with η Cases and produces a derived Verb with n-1 Cases.

(53) suggests a way to answer the question posed earlier about the argument status of en. We propose that en in fact is the "implicit" argument. This argument is linked to the öy-phrase in some way, or, given the data of (51) and (52), to an empty position in the position of the £y-phrase when there is no overt ^ - p h r a s e . We expect such an argument to be assigned Case, following the Visibility Condition. This explains (53b): passive morphology takes away one of the Case features of a Verb because it is the structurally present "implicit argument" and, like all other arguments, this argument needs Case. This proposal is consistent with everything we have said about Case-absorption and offers an immediate account of implicit arguments and the Z?^-phrase. This is the leading idea of the approach we will take to the analysis of passives in this thesis. In the next section, we outline our proposal in detail and begin to explore its ramifications.

32

The Representation

of Implicit and Dethematized

Subjects

2.2. AN ALTERNATIVE PROPOSAL

In this section we propose an alternative to Chomsky's analysis of passive described in 2.1. The main feature of our alternative is that we treat passive morphology as an argument; it acts like a clitic, receiving the θ-role assigned to the subject of actives. This proposal derives the properties of passives given in 2.1.(53). The account of passives we propose depends on certain assumptions regarding the Projection Principle, the ways in which θ-roles are assigned to subjects, and the nature of clitics. As we develop our proposals in this section we will make these assumptions explicit. 2.2.1. Passive Morphology Here we first develop our proposal for en, and then consider a number of consequences and apparent drawbacks of this proposal. Under the latter heading we present a sketch of the role of the passive auxiliary. 2.2.1.1. En as a clitic In this section we will elaborate on our proposal for passive morphology, en, and show more explicitly what our proposals for the passive involve. Given the Projection Principle and the θ-criterion, the claim that something is an argument entails that the item in question must bear the θ-role assigned to a particular Α-position at all levels. This can come about in one of two ways: either the item occupies the Α-position at all levels, or it can form a chain with that position. DS is defined as the level at which thematic structure is represented in its pure form. One consequence of this is that chains at DS are taken to have only one member. In other words, at DS every argument occupies the position to which its θ-role is assigned. We propose a view of clitics here which states that the defining property of clitics is that they are exceptions to the generalization that all DS chains have just one member: clitics are those elements that form chains with argument positions at DS. Clitic chains are subject to a stronger requirement than the usual localbinding condition on each "link;" the requirement for clitic chains is that each link govern the next at every level. Note that it also follows f r o m the θ-criterion, the Projection Principle and the definition of DS that each clitic chain contains exactly one Α-position at DS. 11,12 The main motivation for taking en to be a clitic is negative in that the problem with considering en as a nonclitic is that we then have no good way of preventing it f r o m appearing in a range of Α-positions where it seemingly cannot appear. For example, what prevents en from appearing inside VP at DS, as in (54), giving the surface structure (55) with the meaning " J o h n beat someone?"

Passives and Argument

Structure

(54)

John [vp beat en ]

(55)

John was beaten.

33

The only way to prevent (55) having this interpretation is to force en to only appear in the subject position at DS, since it can only be assigned the subject's θ-role. However, this is really a case of an argument requiring a particular θ-role of its Verb. In other words, to maintain that en is an argument in the sense of occupying an Α-position at DS, we have to introduce the possibility of a kind of "reverse selection" into the theory. This is an undesirable move, as such a mechanism has little precedent or independent motivation. Once we claim that en is a clitic, the next question to answer is what position it appears in at DS. By definition of clitics as forming DS Θchains, combined with the θ-criterion, it cannot be in a θ-position. As all θ-positions are Α-positions at DS, except for the subject position, the only possible Α-position is the subject position. However, in this position, en will be unable to govern an Α-position. So en does not appear in an Α-position, but must appear in some position which governs an A position. The two most likely positions are in Infi and attached to V. These two possibilities are illustrated in (56): (56)

a. b.

e [ VP V+en NP ] e en [ VP V NP ]

Let us consider first (56a). Since en is a clitic, it must form a chain with a θ-marked Α-position. We have proposed that the main condition on clitic chains is that each "link" govern the next. We want en to govern the Α-position which is assigned the subject θ-role, i.e. the subject position. In (56a), however, the government domain of en is VP. So we are forced to do two things if we take (56a) to give en's DS position: first, we must propose that passives have an extra VP-internal Α-position to which the subject θ-role is assigned so en can bind it; and second, we have to complicate the statement of how the subject θ-role is assigned in order to account for just this case. Third, this approach does not rule out the problem raised in (54) and (55) in a principled way; there is no obvious way to prevent en from forming a chain with the object position. (Despite these objections we exceptionally allow en to be base-generated on V in perfects of unaccusatives - see pp. 41-2. On the other hand, if we take (56b) as the DS representation for passives, then we need say nothing about extra VP-internal Α-positions at DS or complicate the statement of how the subject θ-role is assigned. Assignment of the subject θ-role proceeds as normal. However, the subject position

34

The Representation of Implicit and Dethematized

Subjects

is governed by Infi, which contains en. The only Α-position that en can form a chain with at DS is the subject position, so (54) and (55) are correctly ruled out. This seems to be the better of the two alternatives in (56). An additional consideration in favor of this view is that en can now be viewed as a typical verbal inflection appearing in Infi, in fact as part of the auxiliary system (see 2.2.1.2.1. and 2.2.1.3.). Assuming (56b), the DS for a typical passive like (lb) (repeated here) will look like (57): (57)

e; [INFL en4]

[VP

kiss Mary ]

In (57), en forms the clitic chain {en, e) with the subject. Kiss assigns its θ-roles in the normal way to both its subject and its object. This is an important point: DS is the pure presentation of argument structure, and we are claiming that active and passive clauses have the same argument structure. This seems to be true, as we argued in the previous section; in fact, this is an intuition that earlier transformational accounts of passive which assigned the same deep structure to passive and active pairs were designed to capture (cf. e.g. Chomsky (1957),(1966)). So our account is superior to others which do not propose that θ-role assignment proceeds alike in both types of clauses. Three questions arise in connection with (57): (i) how is the subject position dethematized? (ii) how does en end up attached to V, as it clearly is on the surface? and, (iii) what kind of empty category is the el We will postpone question (iii) until 2.2.4., while in the next section we answer questions (i) and (ii) and consider the consequences of some of our answers. 2.2.1.2. Consequences of the Proposal for en 2.2.1.2.1. Case and Affix-Hopping We answer questions (i) and (ii) by saying that en "hops" onto V in the derivation from DS to SS. This is an instance of the rule of Affix-Hopping applying to en as it does to other verbal inflections. We will integrate this proposal with a more general view of the English auxiliary system below. After Affix Hopping (AH) has taken place, en is governed by V. Case assignment takes place at SS, and, in order for the chain headed by en to be visible, V assigns it Case. This is only possible if V is a Case assigner, hence passives of intransitives are ruled out by the Visibility Condition; in such examples en receives no Case. 13 Since monotransitive Verbs only assign one Case, the object NP will not be able to meet the Visibility Condition in VP, and so must move to subject position. Hence the effects of Case-absorption come from AH and the Visibility Condition. Why does en have to move? We could say that en must be assigned ACC,

Passives and Argument Structure

35

and for this reason AH is forced, as the only Case en could receive in Infi is ΝΟΜ. However, this proposal gives the wrong result in examples like (58): (58)

believe [ es eni VP ]

Believe is an Exceptional Case-marking Verb. The Case feature of these Verbs is always ACC, as shown in (59): (59)

John believes him/*he to be a genius.

If en somehow required to be assigned ACC, it could meet this requirement in situ in (59) and hence not have to attach to the V of the lower VP. Clearly this does not happen, however. So it seems that AH is the obligatory property, and the Case properties of en, which lead to Case-absorption, are derived from this. In fact, the Case properties of en are just those of any argument in English: en can be assigned either NOM or ACC. However, it is impossible for NOM Case to ever be assigned inside VP in English, so en is effectively forced to be ACC by the fact that it must undergo AH. (Cf. fn. 3, Chapter One on DAT in English, and 5.3. on NOM en in other languages). 2.2.1.2.2. The Representation of the Subject In 2.2.4., we will propose a general theory of argument clitics which postulates that the empty category in (57) is a variable. This point will be developed further there and in 3.3. However, there are further questions about the subject position that we will try to answer here. First, since there is NP movement to subject position, and the subject position is a θ-position at DS, why are the Θcriterion and the Projection Principle not violated? Part of the answer to this question is a reformulation of the Projection Principle so that it does not hold for subject positions. This is the topic of 2.2.3. Assuming then that the Projection Principle does not hold of subject positions, why is the θ-criterion not violated by NP movement to subject position? The answer has to be that the empty category bearing the θ-role moves out of subject position before the object NP moves in. In fact, this is forced by the government condition on links of clitic chains; this condition requires the empty category in DS subject position to move into VP when en does, otherwise it will no longer be governed by en. We will see evidence in the next chapter that this category appears in VP at SS and LF. Corresponding to the DS in (57), then, we propose the SS in (60) for passives:

36 (60)

The Representation of Implicit and Dethematized

Subjects

Maryi was kiss+enj t, ej

The advantages of (60) can be seen when we reconsider what we concluded were the core properties of passives in (53) ((61) is in fact slightly modified from (53) as we now regard en as the implicit argument): (61)

a. b.

The θ-role normally assigned to the subject position is assigned to an implicit argument in passives, Case is assigned to en in passives.

It is immediately apparent that our account captures exactly the intuition expressed in (61a). Also, we are able to maintain that θ-roles are assigned to a subject and an object as in active clauses at DS, as mentioned above. Moreover, we are now able to adduce a set of explicit predictions about implicit arguments. These predictions form the basis of Chapter Three. As we said above, (61b) is now seen as a function of the level at which affix-hopping takes place. The claim that en is an argument which must be made visible by being assigned Case explains why only one Case is "absorbed" in passives of double object Verbs and why passives of intransitives are impossible. 2.1.2.3. The 1AEX and the Mirror Principle Two other results follow from our proposal.14 First, we can derive the effects of the 1-Advancement Exclusiveness Law (1AEX) of Perlmutter & Postal (1983). Perlmutter & Postal notice that passives of ergative/ unaccusative, raising and passive Verbs are impossible: (62)

a. *The bridge was existed under by trolls. b. *I was seemed to to be smart by Harry. c. *A book was been given by John by Mary.

The proposal that en is a clitic which needs to be assigned the subject θ-role will rule out these examples. These Verbs share the property of not assigning θ-roles to their subjects. This can be seen from the fact that these Verbs all allow expletive it or there as subjects (except for passive Verbs with NP objects, because these NPs must move to subject position to receive Case; this is why we use an example of a Verb with a sentential complement): (63)

a. b. c.

There exist many trolls under that bridge. It seems to me that Harry is smart. It was reasoned that grammars are a hallucination.

Passives and Argument

37

Structure

Perlmutter & Postal propose that all "dummies" like it and there are initially (i.e. at DS) in object position and then "advance to 1" (i.e. move to subject position) in the derivation. The 1AEX prevents advancement to 1 from happening twice in a clause, and so rules out (62). In our terms, if en appears with a Verb of this class, it cannot be assigned a θ-role. As a result, the θ-criterion is violated in (62). So our account of passives derives the effects of the 1AEX in full from the properties of en that we have proposed. 15,16 Another advantage of our account is that it is consistent with the Mirror Principle of Baker (1985a). The Mirror Principle constrains the relation between morphology and syntax, as follows: (64)

Morphological derivations must directly reflect syntactic derivations (and vice versa).

The effects of the Mirror Principle are most clearly seen in languages where Verbs take more than one grammatical-function changing affix. For example, this principle predicts that in a language with both passive and "applicative" affixes which attach to Verbs, if the order of affixation is as in (65a) (where N P B E N indicates "benefactive object", N P T H "theme" and N P A G "agent", and the SVO order indicates grammatical functions), we get a clause with the SS mapping of θ-roles to GFs indicated in (65a) with N P B E N as subject. The important point is that the process triggered by the passive affix applies to the output structure of the process triggered by the applicative affix. On the other hand, if the order of affixation is reversed, we would get the mapping of θ-roles to GFs in (65b): (65)

a. b.

NPBEN NPTH

V+appl+pass N P ™ V+pass+appl N P B E N

(NPAG) (NPAG)

The Mirror Principle can be exemplified from Chi-Mwi:ni. In this language, the order of affixes must be applicative-passive. Hence corresponding to (65a) we find (66a), but (66b), corresponding to (65b), is impossible because of the Mirror Principle, given the required order of affixes: (66)

a.

Mwa:limu 0-tet- el- el- a chibu:ku na Nu:ru. teacher SP-bring-appl-asp-pass book by Nuru "The teacher was brought a book by Nuru." b. *Chibu:ku chi-tet- el- el- a mwa:limu na Nu:ru. book SP-bring-appl-asp-pass the teacher by Nuru "The book was brought the teacher by Nuru." (Baker (1985a, (56c,d)).

38

The Representation of Implicit and Dethematized

Subjects

An assumption necessary for the formulation of the Mirror principle is that affixes like passive and agreement here are not attached at DS, but rather attach successively throughout the derivation, possibly triggering movement as they attach. So the proposal that the passive attaches to the Verb during the derivation is consistent with this view.17 Having seen some advantages of the proposal, we now consider an apparent disadvantage. 2.2.1.3. Affix Hopping and the Auxiliary System We have claimed that passive en originates in Infi and undergoes AH onto V. In this section, we will investigate a number of consequences of this claim for the analysis of auxiliaries. First, we reformulate the classical generative account of Affix Hopping. Second, we investigate the relation between passive and perfect morphology and passive and perfect auxiliaries. Finally, we provide evidence from gapping constructions that Affix Hopping takes place in the syntax and not at PF in English. One point is worth noting at the outset, however. Here we treat passive morphology as a property of the auxiliary system. This is not the case in all languages, however. Many languages have a "synthetic" passive formed by passive morphology alone (e.g. Bantu languages, Japanese, Turkish, Semitic, the non-perfect forms in Latin etc.). A piece of evidence for our proposal for passive morphology can be found in those languages which have both a synthetic passive and overt agreement. In such languages there is a full passive agreement paradigm alongside the active one (e.g. Latin amo, amas, amat etc. (active), alongside amor, amaris, amatur etc. (passive)). Passive morphology and agreement morphology are thus represented together as "portmanteau" morphemes. This shows two things: (i) that passive and agreement are base-generated together, presumably in Infi; (ii) that the passive/agreement morphemes undergo AH onto V. Although our attention throughout will be concentrated on languages with periphrastic passives, it is clear that synthetic passives can be incorporated into our theory, and that they can in fact provide extra evidence for it. AH is a property of the verbal inflections that make up the auxiliary system, in the "traditional" accounts of the English auxiliary system (e.g. the one proposed in Chomsky (1957)). In such accounts, each auxiliary is associated with a particular affix, giving base-generated auxiliary sequences like the following: (67)

have en be ing be en V

AH then inverts each affix and the following Verb (see below on the status of have and be as Verbs). In sequences like these, we take the position occupied by the first auxiliary to be Infi. However, the en in this position

Passives and Argument

Structure

39

is "perfect," not passive, and it undergoes AH to attach to the auxiliary be, not to the main Verb. Moreover, the occurrence of en associated with passive does not originate in Infi, and therefore does not govern the subject at DS. In order to resolve these difficulties, we must make a number of assumptions about the Verbs and affixes which make up the auxiliary system: have, be, en (passive and perfect) and ing. These assumptions amount to a partial account of the English auxiliary system (for a fuller account, see Fabb & Roberts (in progress)). We assume that the inflections ing and en are the spell-outs of features of Infi. This is consistent with claiming that en is a clitic; clitics are frequently viewed as spell-outs of various morphological features (cf. Borer (1984)). The auxiliaries have and be are Verbs with no θ-roles, which are syntactically realized as VP modifiers. The DS configuration for a complex sequence like (67) is then (68): (68)

INFL' INFL [+F] [+G]

[+H]

VP. V

VP. V /

VP,

have V

VP

/

be be

V....

(We have nothing to say here on the ordering and (near) non-iteration of have and be.)n We assume that the features in Infi are associated by " A H " with the various V-nodes in the complex VP. The resulting representations are subject to a number of well-formedness conditions, which together rule out the impossible combinations of auxiliary and affix. One of these is that both aspectual and Case features must be "visible." The features are visible in the required sense iff the nodes bearing them are governed by a unique higher Verb. Moreover, certain specific requirements are imposed on individual affixes (for example, ing must be governed by a Verb of a particular class including keep, start, stop, perception Verbs and be in

40

The Representation of Implicit and Dethematized

Subjects

order to be interpreted as progressive). The operation of " A H , " combined with the visibility condition on its output, derives both the nonappearance of unattached affixes (cf. the No Free Affix Principle of Lasnik (1981)) and the non-appearance of doubly affixed Verbs (here again we derive the No Vacuous Affixation Principle (19) - cf. fn. 15). The important point in the present context is that viewing " A H " as taking place freely, subject to well-formedness constraints, in this way allows us to continue to maintain that the passive en (that is, the features associated with en) appears in Infi at DS. 2.2.1.3.1. Perfects and Passives A further question which arises when we view passive en in connection with the rest of the auxiliary system is: what is the connection between passive en and "perfect" en! The combination of either kind of en with a Verb-stem triggers exactly the same phonological form in all instances, including suppletions and lexical gaps, so we might regard the two kinds of en as instances of the same morpheme. We propose that this is in fact the case. Active W-en participles are found only immediately after auxiliary have; in all other contexts where verbal en participles are found, these participles are passive. This is shown below: (69) (70)

a. I have elected him president. PERFECT b. *He has elected president. a. *I was elected him president. b. He was elected president. PASSIVE

Thus active and passive W-en are in complementary distribution. I propose that underlyingly the passive and active W-en are identical and have basically passive properties, and that when immediately preceded by auxiliary have, the passive properties are replaced by active properties. We see from (70) that be does not have this effect on a following W-en participle. So the account of en in perfects reduces to the question of how have requires an active W-en participle following it. We said above that the aspectual auxiliaries have and be are Verbs without θ-roles. They are distinguished from each other as follows: (71)

a. b.

Have is the transitive auxiliary, Be is the unaccusative auxiliary.

What characterizes transitive auxiliaries? Like transitive main Verbs, transitive auxiliaries assign Case. Moreover, consistent with Burzio's generalization, they assign a θ-role to the subject. We cannot claim that

Passives and Argument

Structure

41

the θ-role they assign comes from their own θ-grid, given our general claim about aspectuals (and moreover it is clear that the subject of a perfect has the same θ-role as the subject of a non-perfect active). Instead, we say that have transmits a θ-role to the subject from some other Verb. Unaccusative auxiliaries, on the other hand, do neither of these things, just like unaccusative Verbs generally. When en appears in Infi, either have or be also appears. If be appears, a clitic chain is formed between en and subject position in the way described above. En then undergoes " A H , " and the argument in subject position is forced to move into VP also so that it remains part of a well-formed clitic chain. Otherwise, the underlying subject will not receive a θ-role. Have, on the other hand, will transmit a θ-role to the subject and assign a Case to en. This means that the subject will not need to move into VP with en, and that en will not "take away" one of the main Verb's Cases. In other words, a clause containing have and en will lack the main properties of passive; to all intents and purposes, it will be an active clause. We can regard have as forming a θ-chain with both the subject and en, and say that en satisfies the θ-criterion either directly by forming a chain with an argument it governs or indirectly by being governed by an occurence of have which governs an argument. For the clitic chain to be well-formed, have must govern both en and the subject. How can this happen? If have is in Infi, then it will not govern into VP, while if it appears in VP it will not govern the subject. However, an intrinsic property of perfect have is that it has a particular tense/aspect. We can exploit this and say that have always appears in VP and forms a chain with Infi. Then the chain (subject NP, Infi, have, en) is well-formed; each link of the chain governs the next and en is (indirectly) associated with an argument. Note that if two en's appear, as in a perfect passive, only the first en, the one governed by have, is "de-passivized," while the second en is passive, forming a separate chain. Note further that no more than two ens can appear, as there can only be one aspectual have, given aspectual visibility, and only one subject. This idea has three interesting implications: first, since causative have appears in passive environments, we must regard these constructions as biclausal: (72)

John had Bill arrested.

Here Bill arrested is the clausal complement to have, and of course this clause is passive. Have has all the properties of a transitive Verb here: it assigns Case to Bill, a θ-role to John and presumably a θ-role to the clause. Second, our approach offers an account of "auxiliary selection" in

42

The Representation of Implicit and Dethematized

Subjects

perfects of unaccusatives. In some languages, notably Italian but also German, Dutch, Danish and many others, the perfect auxiliary is the equivalent of be if the participle following is derived from an unaccusative. Cf.: (73)

Gianni e/*ha arrivato. G is/*has arrived. " G has arrived."

In other languages - English, Swedish, Spanish for example - have is the only perfect auxiliary. There is good evidence that unaccusatives are syntactically derived in that their subjects are underlyingly direct objects which are not Case-marked (the locus classicus for such arguments is Burzio (1986)). Moreover, these are single-argument Verbs, so they lack subject θ-roles. We suggested above that the central property of transitive auxiliaries is that they transmit a θ-role to the subject, but in unaccusative structures none is available, hence have cannot appear in constructions with unaccusatives. If we accept this conclusion about have, our proposal that be is the unaccusative auxiliary is supported in these languages because the auxiliary which appears with a perfect form of be (e.g. in a perfect passive) is be, cf.: (74)

Gianni e'/*ha stato arrestato. G is/*has been arrested. " G has been arrested."

If this proposal is correct, what of languages like English, Spanish and Swedish? For these languages we have to say that the requirement on have as a transitive auxiliary is relaxed, so that have is not required to assign a θ-role but only Case (this might amount to Burzio's generalization holding in a different form in these languages - cf. Chapter Four). Languages which select be pose a different problem: how is en licensed? We propose that en is licensed indirectly by being governed by an occurrence of Agr which governs an argument. So en forms a chain with Agr, which itself forms a chain with the DS object for purposes of Case-assignment and agreement. This idea is supported by the fact that en agrees with the underlying object in unaccusative perfects just where be is selected as the auxiliary (if the language has overt agreement). Third, we predict that unaccusative auxiliaries will always be associated with passive. It is observed by Jaeggli (p.c.) that every language which has a passive auxiliary uses a Verb which is otherwise unaccusative for this function. This Verb can be " b e " (English, Russian, Romance), "become" (German, Dutch, Persian), "get" (Celtic, Tzeltal), "go" (Hindi, Persian) - see Keenan (1980:20-24).

Passives and Argument

Structure

43

The above paragraphs show that our proposals for passive morphology can be integrated into a general theory of auxiliaries, in particular passive auxiliaries, in a way which offers an account of several aspects of their behavior. 2.2.1.3.2. Evidence for Syntactic Affix Hopping: Gapping The above sketch of certain parts of the auxiliary system shows how it is possible to regard en as originating in Infi even in complex auxiliary constructions. We have made a further unusual assumption about AH, however: that it applies between DS and SS, rather than in PF, as is more usually assumed. Clearly this assumption is very important for our analysis of passives, and so we need to give some motivation for it. In fact, Siegel (1984) provides just such motivation from Gapping, illustrated in (75): (75)

Ward can't eat caviar and June beans.

(75) is ambiguous between a wide-scope reading for the negated modal, represented in (76a), and a narrow-scope reading, given in (76b): (76)

a. b.

NOT CAN (Ward eat caviar & June beans). NOT CAN (Ward eat caviar) & NOT CAN (June eat beans).

Siegel says: The narrow scope reading is true just in case neither person can eat the food named. The wide scope reading is true just in case it can't be that they both eat the food mentioned, (p. 524).

On a deletion analysis of Gapping, the wide-scope reading has the ungrammatical SS (77): (77)

*Ward can't eat caviar and June eat beans.

But, in all cases except those involving modals, tenseless second conjuncts are impossible, cf.: (78)

*Ward eats caviar and Sue eat beans.

So wide-scope readings for sentential operators under Gapping cannot be derived from grammatical SS's. Therefore, Gapping cannot be deletion. We are assuming an LF like (79) to be appropriate for (77) (here we depart from Siegel's presentation on matters of execution):

44 (79)

The Representation of Implicit and Dethematized

Subjects

Ward can'ti eatj caviar and June [iNFLe], [VP[V e]J beans].

This LF has an empty Infi and an empty V in the second conjunct, whose interpretation is derived from the respective interpretations of the Infi and V of the first conjunct by means of a rule of construal (cf. Fiengo (1974)). The LF corresponding to (78) is (80): (80)

Ward

[INFL

[+present]] eat caviar & June

[INFL

e ] eat beans.

Gapping can rule out (80) and rule in (79) if we assume, as is fairly standard in cases of Gapping and VP-Deletion, that strict syntactic parallelism is required between empty categories and the elements upon which they depend for their interpretation. Then, if AH applies before LF, giving a representation like (81) instead of (80), (78) is ruled out: (81)

Ward eat[+present] caviar and June

CINFL

eat beans.

Hence, to achieve the right results with Gapping, we must assume LF is sensitive to AH. So AH applies at SS. The above argument shows LF to be sensitive to the placement of features realizing sentential operators like negation, tense and modality by AH. However, our concern is more with the aspect and "voice" affixes ing and en. These affixes show only the "narrow-scope" reading under Gapping: (82)

a. b.

Ward hasn't eaten caviar and June beans, Ward isn't eating caviar and June beans.

These examples cannot have the readings in (83) (where, informally, HAS represents perfect and ING progressive): (83)

a. b.

NOT HAS (Ward eat caviar & June beans), NOT ING (Ward eat caviar & June beans).

This absence of wide-scope readings is unsurprising. Perfect and progressive are usually taken to be VP-"operators," in contrast with sentential operators like tense, modality and negation, and we are dealing here with Scoordination, so it is impossible for VP-operators to have clausal scope. In fact, if we maintain the assumption made earlier that these VP-operators are generated in Infi, then we can claim that they must undergo AH to put them to a position with VP-scope by LF. This occurrence of AH must also, therefore, take place in syntax. So we conclude that AH takes place in the syntax.

Passives and Argument

Structure

45

2.2.1.4. Conclusion In this section, we have presented the core of our proposals for the passive. We have provided some support for these proposals and will develop the ideas given here and support them further in the remainder of this chapter and throughout the next. For convenience, at this point we sum up the central proposals so far: (84)

a. b. c.

en is a subject clitic. en appears in INFL at DS. en undergoes AH to adjoin to V at SS.

2.2.2. Subjects and Agents The topic of this section is how θ-roles are assigned to subjects. We will propose a fairly elaborate account of how subjects are θ-marked. The account rests on a distinction between structural and inherent θ-role assignment. The main consequence of this account is that it allows a unified treatment of the processes which result in a "missing subject," and of the relation between subjects and Agents. θ-roles are assigned under government. Thus, in a configuration like (85) V θ-marks NP: (85) V

NP

This is also the configuration in which V subcategorizes for NP. We pointed out earlier that V can have a non-subcategorized argument: the subject. How do we know when a given Verb has a subject θ-role to assign? This follows from subcategorization and a claim we made in 1.3.1.; all arguments except the subject argument are subcategorized. We can display argument structure by linking subcategorization frames to positions in θ-grids (the same approach is adopted by Jaeggli (forthcoming,a), Zubizarreta (1985)); in this way that subject can be identified: 19 (86)

donate θι

θ2

θ3

NP

PP

In (86), θι is marked as the subject argument of donate by default, as it is the only one not associated with a subcategorization frame. An argument appears in the government domain of V iff it is subcategorized. Since the subjects are not subcategorized (cf. 1.3.1.) they cannot be θ-marked in a configuration like (85). Instead, it is usually said that

46

The Representation of Implicit and Dethematized

Subjects

V indirectly θ-marks the subject, while complements are directly θ-marked (cf. 1.3.1.). We propose a rather different typology of θ-marking, which cuts across the complement-subject distinction. There are two kinds of θ-marking: structural and inherent. Structural θ-assignment takes place under R and coindexation, as defined in 1.2.(8); hence XP structurally θ-marks an argument it R-modifies and is coindexed with and a head structurally Θmarks an argument it governs and is coindexed with. So, for example, if an argument is R-modified by the maximal projection of a θ-assigning head, θ-role assignment to that argument takes place via predication in the sense of Williams (1980). Inherent θ-role assignment is a lexical relation between a θ-role assigner and its argument which can be satisfied under R with no coindexation requirement. Which θ-roles are assigned in which ways? We propose that any θ-role may be assigned structurally, but only certain θ-roles are assigned inherently. In other words, some θ-roles are only assigned structurally. These are the structural θ-roles. A structural θ-role assigned by VP is an Agent; a structural θ-role assigned by V is a Theme. We retain the standard view that θ-roles, including Themes and Agents, are lexico-semantic properties of Verbs. However, we add something to this view: in a particular configuration - say where a VP is coindexed with a subject (of the right selectional type - see below), the structural θ-role-reading of Agent can be "overlaid" as it were on the inherent Θrole assigned by the Verb. This arises because the agentivity is read off a structural configuration at LF. This fact underlies the often-noted objection to the buiuniqueness requirement of the θ-criterion that certain arguments intuitively seem to bear two θ-roles. Examples of this with Agents are: (87)

a. b. c. d.

John John John John

received the package. (Goal+Agent) sent the package to Bill. (Source+Agent) opened the door. (Instrument+Agent) worried about Bill. ( E x p e r i e n c e d Agent).

On our view, this is a consequence of the way one θ-role is assigned, since at L F there is coindexation between the subject and the predicate. So the intuition that underlies doubts about the θ-criterion can be captured, while the formal biuniqueness requirement remains in full force. We will say more about the representation of structurally assigned θ-roles in Chapter Four. The idea that Agent θ-roles are structural θ-roles assigned by VP has three consequences: (i) Agents receive compositional θ-roles; (ii) no other kind of subject argument receives a compositional θ-role; (iii) an Agent

Passives and Argument

Structure

47

θ-role could never be assigned to an argument in a subcategorized position. The first of these consequences can be illustrated with examples from Marantz (1984:23 - 31). Marantz uses the following examples to show that subjects receive a compositional θ-role (cf. the earlier passive examples 2.1.4.(35 - 39)): (88)

a. b. c. d. e. f. g· h. i. j· k. 1. m. n. o.

throw a baseball throw support behind a candidate throw a boxing match throw a party throw a fit take a book from the shelf take a bus to New York take a nap take an aspirin for a cold take a letter in shorthand kill a cockroach kill a conversation kill an evening watching TV kill a bottle kill an audience

Marantz says: Although the different predicates formed by adding different direct objects to most simple transitive verbs assign different semantic roles to their subjects, one would be hard pressed to argue that a different predicate implies a different (homophonous) verb in each case. Killing a cockroach involves causing it to become not alive; killing an evening may be accomplished by sitting motionless in front of the TV; and killing a bottle implies causing it to become empty, yet some thread of meaning ties these uses of kill together, a thread we may weave into a single function "kill", from arguments to predicates. (Marantz (1984:25-6)).

In other words, the subject argument here is semantically an argument not just of the Verb, but of the complements of the Verb and the Verb taken together. 20 Note that these are all agentive predicates; the subject is assigned an Agent θ-role (cf. Marantz's glosses using cause). As further support for this view, Marantz points out that there seem to be no idioms in English which consist of subject and Verb with the object slot open, while there are many which consist of a predicate with the subject slot open. So beside the idioms like spill the beans, let the cat out of the bag, there are none like a soccer player kick NP or the cat let NP. This indicates that predicates, i.e. Verbs plus complements, are formed independently of subjects. It is also quite clear in these cases

48

The Representation of Implicit and Dethematized

Subjects

that the subject is the compositional argument of VP. These idioms are in general understood with agentive subjects. Regarding consequence (ii) above, I know of no cases where a nonAgent subject allows the range of quasi-idiomatic readings that Marantz's examples show. Compositional Themes, Sources, Experiencers etc. do not seem to exist.21 The third consequence of our proposal, that subcategorized Agents are not found, seems to hold in general. Marantz (1984) points out that this is at least the unmarked situation in the world's languages (except for those he calls "syntactically ergative" languages). He also gives evidence that children acquiring language will analyze transitive Verbs as having agentive subjects and, if presented with an invented Verb which assigns an Agent θ-role to its complement (his example is moak, as in The table moaks my elbow "I pound the table with my elbow"), they will have difficulties. Cf. also Pinker (1984:39ff.) on the "semantic bootstrapping" hypothesis in acquisition; this is the idea that acquisition of syntax takes place by means of an initial assumed semantics-syntax mapping which enables syntactic categories and relations to be identified in terms of cognitive/semantic primitives. One aspect of this hypothesis is that syntactic subjects are initially assumed to be semantic Agents. If Agents are absolutely prohibited from appearing in subcategorized positions, and the argument in the ^ - p h r a s e is the compositional argument of the predicate (cf. 2.1.4.(35-39)), then the fry-phrase must have moved into VP from subject position. So our approach to θ-role assignment, which has a number of independent advantages, requires us to adopt a theory of passive like the one outlined in 2.2.1., where the underlying subject argument is in subject position at DS. Two other phenomena which the above account of agentivity allows an account of are (i) secondary agentivity and (ii) "intentional" be. These are illustrated in (89):

(89)

a. b.

Mary was deliberately seduced, John was being a fool.

(89a) is an example of secondary agentivity where deliberately is understood as a property of Mary, so we could gloss this example as "Mary deliberately got herself seduced." (89b) is an example of intentional be. What the sentences of (89) share is a θ-role-like relation between the predicate and an otherwise "nonthematic" subject position: the subject of a passive in (89a) and the subject of the unaccusative be in (89b). One piece of evidence that the relation is θ-role-like comes from the fact that selection is involved; as is typical with Agents, inanimates lead to anomaly:

Passives and Argument Structure (90)

49

a. The kitchen was deliberately painted, b. ??The concept was being a pain.

(90a) only allows the reading where deliberately is associated with the implicit Agent; this is because kitchens cannot be deliberate (as is typical with selection violations, sufficient imagination lets (90a) in). To account for secondary agentivity we propose that the conditions for the assignment of a structural θ-role can be met at SS and LF where there is not necessarily structural assignment of an inherent θ-role to the subject position at DS (or where the argument assigned a θ-role in subject position at DS has moved out of that position, as in passives). Nevertheless, since the LF configuration is one of structural θ-role assignment to the derived subject, this argument is interpreted as agentive (and again, this interpretation is "overlaid" on the interpretation of the inherent θ-role, if there is one). In terms of our proposal for Agent θ-role assignment, then, the phenomenon of secondary agentivity has a natural explanation; it exploits the independence of the two modes of θ-role assignment. Furthermore, we correctly predict that there is no phenomenon of "secondary Sourcehood, Experiencer" etc., as inherent θ-roles correspond directly to lexical θ-roles assigned at DS. Themes, too, cannot be θ-marked in this way as they are structurally θ-marked inside VP, so any argument occupying Theme position must have received its θ-role in that position, given the Projection Principle (cf. 1.3.1. and the section to follow). In (89b) and (90b) - examples involving intentional be - the situation is similar. Be assigns no θ-role, and yet the predicate named by (the progressive occurrence of) be here clearly assigns a θ-role to the subject. Intentional be in general appears when be is a "main Verb," i.e. not functioning as part of the auxiliary system (cf. Williams (1984) for a detailed discussion of the environments in which this kind of be appears). In other words, intentional be is the occurrence of be as head of a predicate. This predicate can be coindexed with the subject, and hence the structural configuration for Agent θ-role assignment obtains. Hence the agentive (or intentional) reading for these predicates. As a final point, it is often observed that agentivity entails eventhood. We can integrate this observation fairly simply into our proposals for agentivity. We propose that the coindexing between VP and the subject is mediated by Infi. VP is coindexed with Infi, and Infi is coindexed with the subject. This is illustrated in (91):

50

The Representation of Implicit and Dethematized

Subjects

IP

(91) NP,

Γ VP, / V,

Where the Verb assigns an Agent θ-role to the subject, the subject Θrole is first assigned to Infi, and then Infi transmits it to the subject. In this case V shares the index with VP, as the θ-role originates as a lexical property of V. This proposal means that Infi is always coindexed with both VP and the subject where the subject is Agent. Infi contributes a tense/aspect specification to the compositional θ-role assigned the subject (we assume this is so even for infinitival Infi - cf. Stowell (1982), En? (1985)). This is what gives rise to the event interpretation. Clearly more needs to be said about this, but we will defer discussion of the general relation between tense, aspect and θ-roles until Chapter Four. The proposal for Infl's contribution to θ-role assignment was implicit in what was said about auxiliaries in the previous section; the presence of have and be influences the transmission of a θ-role to the subject. We saw in 2.2.3.1. that it is possible to construct an account of "auxiliary selection" in passives and perfects, and in particular perfects of unaccusatives, in these terms. θ-role assignment to the subject can thus "break d o w n " in various ways, giving rise to the appearance of a "suppressed" subject. First, some element may appear in Infi and act as an argument requiring the θ-role - this is what happens in passives (and perfects) when en appears and in impersonal and middle constructions with si. Second, since coindexation between VP and Infi is optional, it may fail to hold: this is what happens in middles (see Chapter Four). Third, if Infi is not present, θ-role assignment is expected to proceed in a rather different way - this is the core of the peculiarities of nominals (see 4.5.). In this section we have proposed that arguments, in particular arguments in subject position, can be θ-marked either structurally or inherently. Some θ-roles are assigned either way, some only structurally: the latter class consists of Agent and Theme. Structurally θ-marked subjects can have an agentive reading "overlaid" on the core inherent θ-role at LF; this possibility is available for arguments which are not DS subjects and for predicates headed by non-0-assigners (intentional be). Structural θ-marking is a relation between VP, Infi and the subject, and therefore is compositional (the contribution of the rest of VP) and eventive (the contribution of Infl's

Passives and Argument Structure

51

tense specification). We next turn to the most important constraint on θ-role assignment in general: the Projection Principle. 2.2.3. Lexical and Structural Relations: the Projection Principle Our proposals for the passive mean that we must reformulate the Projection Principle so that it does not hold for subject positions. In 1.3.1 we gave the following informal version of the Projection Principle: (92)

Representations at each syntactic level (i.e. LF, and D- and SStructure) are projected from the lexicon, in that they observe the subcategorization properties of lexical items.

In this section, we will consider the more formal version of this principle given in LGB, pp. 36 - 40. We then propose a reformulation which allows for movement of arguments from subject position into VP. Since this proposal leaves open the question of what forces subject positions, we also introduce the Principle of Full Interpretation (FI) to explain this. The formulation of the Projection Principle that we propose is not sufficiently restrictive in certain ways. To deal with this problem we propose a condition on θ-role assignment which does some of the work of the LGB Projection Principle. Because of this, it would perhaps be more accurate to say that we are splitting rather than weakening the Projection Principle. The split that we make leaves a residue of cases unconstrained either by the Projection Principle or by our condition on θ-role assignment. One of these cases is that of movement of an external argument into VP, as proposed in our analysis of the passive. Before going on, we must clarify two assumptions that we have been making about θ-role assignment: (93)

a. b.

θ-role assignment is optional. θ-roles are assigned to categories occupying structural positions.

The importance of the Projection Principle can be seen when we consider the case of a VP containing a Verb V and some XP an argument of V. We are concerned with a configuration like (94): (94)

VP V

XP

What are the relations between V and XP here? Clearly V governs XP. Also, V strictly subcategorizes for XP (this in fact follows from the original proposals for subcategorization in Chomsky (1965:Chapter Two)). Furt-

52

The Representation of Implicit and Dethematized

Subjects

hermore, as we have seen, the following requirement holds: (95)

If V subcategorizes for XP, then V θ-marks the category occupying X.

(95) does not hold from right to left. V is not subcategorized for its subject, but subjects can be θ-marked by V, in the manner described in the previous section. So (95) is a one-way implication only. Next, Chomsky introduces the following condition: (96)

If X θ-marks Y, then X selects Y. (Chomsky (1981:38)).

Again, (96) is a one-way implication. We exploit this fact in Chapter Three when we define modification as selection without θ-marking. Given what was said in the previous section, (96) means that both subject and object arguments are selected by the predicate. Also, in (94), V selects XP. Now the Projection Principle is given as (97): (97)

i. if X subcategorizes Y at Li, then X θ-marks Y. ii. if X selects Y as lexical property, then X selects Y at Li. iii. if X selects Y at Li, then X selects Y at Lj.

(97i) is a restatement of (95). (97ii,iii) express the idea that thematic properties must hold at all syntactic levels. In other words, (97) still expresses the basic idea that syntax is a projection of the lexicon. Other things are predicted by (97), in conjunction with (96) and the θ-criterion: (98)

a. b. c.

Movement in a derivation to a position at Lj which is θ-marked at Li is impossible. Positions that are θ-marked at Li are structurally present at U (i>j). Complements are always θ-marked by heads.

The statements in (98) together have the result that movement to a Θposition is impossible, as we saw in 1.3.1. Also, "tree pruning" (at least of thematic branches) is impossible since θ-marked positions are θ-marked at all levels. Effectively, then, the Projection Principle combined with (96) makes the θ-criterion hold at all levels for the positions occupied by the arguments that are subject to it. Put another way, the Projection Principle underlies the existence and formation of θ-chains.

Passives and Argument

Structure

53

However, although the Projection Principle as stated in (97) guarantees the presence of lexically specified (selected and/or subcategorized) complement positions, not all subject positions are guaranteed. Only θ-marked (i.e. selected) subjects are forced to be present. It is a property of subject positions, as distinct from complements, that they can be nonthematic. Nonthematic subjects appear in the following examples: (99)

a. b. c. d.

For it to seem that John was intelligent was amazing. There arose some unpleasant odors. It was reasoned that grammars are a hallucination. It is unbelievable for there to be unicorns in the garden.

To account for the presence of subjects in examples like (99), Chomsky (1982) proposes the Extended Projection Principle (EPP). The EPP is basically (97) with (100) added: (100)

Every clause must have a subject position.

It seems that (100) and (97) do not form a natural class of statements. (100) is essentially a stipulation purely about structures, while (97) is intended to capture the relation between lexical properties and structures. As noted in 1.3.1., there seems to be a principled distinction between subject and complement positions in that complement positions only appear as a function of lexical properties of V while subject positions always appear in the clause. This distinction is in fact captured by the theory of subcategorization combined with PS rules for the expansion of S (as in Chomsky (1965)), but not by the EPP. A different approach is suggested by some of Chomsky's more recent proposals. Chomsky (1986) proposes that subjects are forced as a result of the Principle of Full Interpretation (FI). We formulate FI as follows: (101)

The content of every category must be recoverable.

This principle holds at LF. The instantiation of FI for categories which are predicates is: (102)

Every predicate must be in relation R with a subject. (Cf. Rothstein (1983)).

We refer to (102) as the Predication Principle. This is a condition on the content of predicates, not on that of subjects. No requirement is imposed on the subject beyond that the position be present in the government domain of a predicate. We retain (102) as a

54

The Representation of Implicit and Dethematized Subjects

stipulation for now; in Chapter Four, we will suggest a way in which it can be derived from principles of temporal/aspectual interpretation. (102) is redundant with respect to (97) to the extent that (97) forces thematic subject positions. Accordingly, we restrict (97) as follows: (103)

For X a head and Y a subcategorized argument of X: i.X θ-marks Y iff X selects Y. ii.X selects Y as a lexical property iff X selects Y at all levels.

(103) is our restatement of the Projection Principle. Different predictions are made by (102) and (103) in combination, as compared to the combination of (96) and (97). The most important predictions concern the status of subjects. The important thing about our approach is that, instead of requiring subjects as a consequence of the Projection Principle, which, given (97 = 103i), forces subject positions as a function of selectional properties and so makes the θ-criterion hold for subjects (or fails to, as in (99)), we impose the requirement in (102). (102) is a requirement on the content of predicates which requires them to be in a purely structural relation with a subject and so it does not interact with the θ-criterion in any way. In principle, our approach can have three consequences: (104)

a. b. c.

Movement in a derivation to a subject position at Lj which is θ-marked at Li may be possible. (Cf. (99a)). Subject positions need not contain arguments, even where the Verb is specified as having a subject argument. Subject positions that are θ-marked at Li need not be structurally present at Lj. (Cf. (99b)).

The situation in (104a) arises in passives, we have proposed. The θ-criterion will rule out a derivation like that described in (104a) unless the argument that bears the θ-role assigned to that position at Li is moved out of that position. If the element θ-marked at Li is moved out, it is not forced to form a chain with an empty category in the position, and because of the θ-criterion it is not able to form a chain with that position, so nothing prevents some other argument from moving into the position. (104b) holds as we are assuming that θ-role assignment is not itself obligatory (cf. (93)); θ-roles are only forced to be assigned to arguments that are forced to be present by selection/subcategorization. Given the formulation of the Projection Principle in (103), this is the case only with complement positions. Nothing forces subject arguments to appear in subject positions. As we mentioned above, (102) forces the presence of the subject position, but nothing we have said requires an argument of

Passives and Argument Structure

55

the Verb to occupy this position, where the Verb is lexically specified as having a subject argument. We leave open precisely this possibility: the subject position may sometimes not be filled by the NP assigned the Verb's subject θ-role because VP might fail to coindex with Infi (cf. 2.2.2.). This is what happens in middles, as we will see in Chapter Four. When this happens, we find that the aspectual interpretation of the clause is affected in a particular way. Formulating these aspectual constraints and their interaction with θ-role assignment is one of the topics of Chapter Four. (104c) indicates that this is certainly too unconstrained a view of subject positions as it stands, since arguments in subject position can freely disappear between levels. To prevent this, we introduce a condition on θ-role assignment. Subject arguments could also be inserted at Lj. However, subject arguments must be present at DS as DS is defined as the pure representation of argument structure. On our approach, though, nothing prevents subjects from disappearing between DS and SS and reappearing at LF (as long as the CFCC can be satisfied - see below). The condition in question is the Complete Functional Complex Condition on θ-role assignment (CFCC): 22 (105)

If X is a head and YP is a maximal projection selected by X, then X θ-marks YP iff YP is a CFC.

CFCs are defined as follows: (106)

CFC(X) = the smallest maximal projection containing all positions θ-marked by X where XP satisfies (102).

The CFCC interacts with the Projection Principle so as to prevent subjects from disappearing between levels. Although neither the θ-criterion nor the Projection Principle forces θ-role assignment to subjects at all levels, if a clause contains a Verb which does not θ-mark its selected subject, then that clause will not be a CFC. If the clause is an embedded clause, then if it is not a CFC, it will not be able to be θ-marked, and so the Projection Principle will be violated because the item in the higher clause which should θ-mark it will be unable to do so. If the clause is a root clause, we can assume that it will be ruled out by the following requirement: (107)

Root clauses must be CFCs.

To the extent that (107) holds and Verbs select for complement clauses which are subject to the Projection Principle, the CFCC will prevent subjects from disappearing between levels. Another entailment of (105) is that no

56

The Representation of Implicit and Dethematized

Subjects

projection which assigns a θ-role to a subject argument can itself be an argument (cf. the remarks on functionals in fn. 22). We are assuming (icontra Williams (1981)) that NPs never have "external arguments," i.e. that all arguments of Ν are inside NP (see 4.5). The same is true of clauses. VPs, on the other hand, typically do not contain all their arguments, and moreover are predicates subject to (102) so cannot be assigned θ-roles in general. When a VP receives a θ-role it must form a small-clause to contain all its arguments and satisfy (102); this happens in causative and perception complements in English. Our statement of the Projection Principle allows us to define a notion Pred(X) ("the predicate headed by X"), as follows: (108)

Pred(X) is the smallest category containing X and all X's arguments required by the Projection Principle.

In terms of this definition, we can in turn define "external argument" and "internal argument:" (109)

a. b.

A chain C is the external argument of X iff C is selected by X° and governed by Pred(X). A chain C is an internal argument of X iff C is θ-marked by X° and not governed by Pred(X).

Pred(X) is subject to the Predication Principle. In conjunction with the CFCC, these principles and definitions allow the following generalizations: (110)

a. b.

If XP is a CFC and if XP is headed by a 0-assigner then XP contains a Specifier, If XP is not a CFC, then XP will have a predication relation with its subject.

(110) captures a fundamental difference between predicates headed by V and predicates headed by N, where V and Ν have exactly the same lexical argument structure, as we shall see in 4.5. To summarize, we have seen how the Projection Principle makes the θ-criterion apply to the positions occupied by arguments, in addition to applying to the arguments in those positions themselves. We proposed that the Projection Principle, thought of this way, should not hold for subject positions. Instead, subject positions are forced by the requirement on predicates in (102). This requirement is purely structural, so the presence of the subject position is not a function of selectional properties, or therefore, of the θ-Criterion. So the possibilities in (104) hold whether the subject position is occupied by an argument or not. Some of the possibilities in

Passives and Argument

Structure

57

(104) are desirable; we have already seen how our analysis of passives favors (104a), and in Chapter Four we will see what happens when the optional coindexing between V and Infi does not take place, meaning that they do not θ-mark their subjects, fulfilling the possibility in (104b). (104c) is effectively neutralized by the CFCC. 2.2.4. Clitics In this section we will elaborate on the assumptions we have adopted concerning the nature of clitics. These assumptions are clearly central to our theory of passive, given our proposal that en is a clitic. Our view of clitics is something of an amalgam of the theories of Aoun (1979a, 1981b), Borer (1984), Jaeggli (1982,forthcoming^) but with some departures from the views of those authors (cf. also Everett (1985)). We first reconsider in a little more detail the assumptions we have made about clitics so far. In 2.2.1. we said that clitics have the properties in (111):

(111)

a. Clitics form DS θ-chains with A-positions. b. Clitics are the spell-out of certain features of that chain.

We continue to assume (111). A further statement which is frequently made about clitics (e.g. in Borer (1984)) is (112): (112)

Clitics are part of the feature matrix of heads.

Whether the clitic is "part o f ' the head in this sense or has autonomous features which just happen to be spelt out on the head is a question we will deal with in the discussion of clitic doubling below. We will see there that (112) holds for some clitics but not for others. The condition on clitic chains can be stated as follows: (113)

In a clitic chain C = (ai ... a n ), at must govern ai+i.

In general, local binding is the condition on chain links (cf. 1.3.1. (16,17)). Why should clitic chains be subject to the more stringent requirement in (113)? A possibility is that R as in 1.2(8) is the general structural condition on chain-formation; if this is so then local binding and binding in general will be defined in erms of government where the binder is a head and in terms of c-command where the binder is a maximal projection. In the case of chain-formation, heads will form chains each of whose links must govern the next, while maximal projections will form chains each of whose links must c-command the next. For this proposal to work, we must further state that the X-bar status of the head of a chain C = (ai ... a n ) determines the structural requirement which must hold between ai and ai+i.

58

The Representation of Implicit and Dethematized

Subjects

Clitics always appear attached to a head, so they govern the complements to that head; in fact, they govern everything the head governs. So a clitic in Infi governs the subject but not the object. Since clitic chains are subject to the condition that each link must govern the next, this has the desirable consequence that en can form a chain with the subject but not with the object, as we saw in 2.2.1. According to (111a), clitics do not themselves appear in A-positions, but must be coindexed with an argument in an Α-position which they govern. Since they are not in Α-positions, clitics cannot antecede pronouns or anaphors (cf. the presentation of the Binding Theory in 1.4.). Note that the Α-position in question does not have to contain a category Θmarked by the head the clitic attaches to; we have to say this since Infi clearly does not θ-mark the subject (whatever Infl's involvement in Θtransmission to the subject - see 2.2.2.) and yet en appears in Infi. There is some independent motivation for this proposal from certain instances of Exceptional-Case-marking like phenomena in French, for example: (114)

Je lei crois t, un genie I him believe a genius "I believe him (to be) a genius."

Here the clitic le is attached to croire, a Verb which Case-marks but does not θ-mark a category occupying the Α-position to which the clitic is linked. So it seems that this possibility must be allowed in any case. In the majority of cases, clitics will be associated with a θ-marked complement of the head they are attached to because of the government condition on cliticchain formation (cf. also fn. 12). Let us consider how what we have said up to now applies to passives. We saw evidence in examples (34 - 38) of 2.1.4. that the &y-phrase receives the subject θ-role from the VP in passives. To the extent that this θ-role is agentive and compositional, the theory of θ-role-assignment put forward in 2.2.2. forces us to say that the öy-phrase originates in subject position. So we regard the £y-phrase as the overt realization of the "implicit argument" of agentless passives. Hence the D-Structure of a passive with a Z>y-phrase is as in (115): (115)

(by) Johnj en* kiss Mary.

Here the subject θ-role is assigned to the chain C = {en, John). When en attaches to V in the syntax, John is forced to move into VP as a consequence of the government requirement on links of clitic chains given in (113) (unless have appears to form a chain linking the NP in subject position and en - see 2.2.1.3.1.). At SS, en is assigned ACC by V, so Mary moves to subject position

Passives and Argument Structure

59

to be assigned Case. We thus arrive at an SS representation like (116): (116)

Maryj kiss+eni tj by Johni.

This analysis treats passives with overt ύ^-phrases as cases of clitic-doubling. In other words, we are claiming that en and John in (116) have the same relation as the pair (clitic, NP) in a language which allows clitic-doubling fairly freely, e.g. River Plate Spanish (also Rumanian - cf. Steriade (1980), Borer (1984), Dobrovie-Sorin (1985)). Example (117) exemplifies the phenomenon in River Plate Spanish: (117)

Loi vimos a Guillei. Him we-saw to G "We saw G . " (cf. Jaeggli (1982)).

So en and the ύ^-phrase in (116) have the same relationship as lo and Guille in (117). En and RP Spanish lo are examples of clitics which allow doubling. We propose to account for the possibility of clitic doubling in terms of a very general condition on binding: (118)

If X binds Y, then X and Y must not conflict in features.

(118) rules out examples like (119): (119) *Johni likes herselfi. Strictly speaking, (119) does not violate Principle A of the binding theory, as the anaphor herself is bound in its binding domain. However, there is conflict in features, since John is (inherently) masculine and herself is feminine, so condition (118) is violated. Alternatively, we can regard (118) as a well formedness condition on binding relations; if (118) does not hold (i.e. if there is conflict in features) then binding is not possible. In this view, (118) is a violation of Principle A as herself cannot have John as antecedent. Since (118) is a condition on binding, it is also a condition on chains. In the case of clitic chains, this has the consequence that there must not be conflict in features between the clitic and the doubled NP. (120), for example, violates this condition: (120) *Laj vimos a Guillei. Her we-saw to G.

60

The Representation of Implicit and Dethematized

Subjects

(120) is ruled out because the clitic la is not a member of a well-formed chain. As la is an argument clitic, this is ultimately a violation of the θ-criterion. With an auxiliary assumption, this account also extends to Kayne's generalization: (121)

An object NP may be doubled by a clitic only if this NP is preceded by a Preposition. (cf. Jaeggli (1982:20).

Kayne's generalization holds both for RP Spanish object-doubling and for the passive (by being the Preposition in question). This generalization is the instantiation of the constraint in (118) for Case features. The clitic chain must be assigned Case, following the Visibility Condition. Overt NPs always carry a Case feature, which may or may not be overt, so (118) requires the Case feature of the clitic and the Case feature of the doubled NP to not be in conflict. If the features are in conflict, the clitic chain will be ill-formed. We need to make an auxiliary assumption about doubled NPs, however: they must be overtly Case-marked. In languages with no inflectional case morphology on Nouns, overt manifestation of Case is effected by government by a Preposition. Ths last statement allows us to derive Kayne's generalization in full from (118). We assume therefore that both a and by can be ACC-assigners. 23 An important issue concerning clitic-doubling is that of the extractability ^ of doubled NPs. Extraction of clitic-doubled direct objects is in general impossible in dialects where the doubling clitic is lo: (122) i A quien lo viste? Ρ who cl you-saw "Who did you see?"

However, extraction of clitic-doubled indirect objects is possible: (123) i A quien le han regalado ese libro? Ρ who cl they-have given that book "To whom have they given that book?"

This contrast sheds further light on the exact form of the "Case agreement" condition on clitic chains. Dobrovie-Sorin (1985) accounts for the contrast between extraction of doubled indirect objects and extraction of doubled direct objects in terms of inherent Case-assignment. We will adopt this idea here and extend it slightly.

Passives and Argument

Structure

61

The crucial thing about extraction of doubled NPs is that the variable occupying the extraction site is a member of two chains: the A'-chain and the clitic chain. The variable only needs to be part of the A'-chain to be well-formed. However, the clitic must still be identified as part of the clitic chain or the θ-criterion will be violated. In other words, for the clitic to be well-formed as an argument clitic, the variable must be identified as part of the clitic chain. This, we propose, can take place in one of two ways: (i) the position occupied by the variable can agree in Case with the clitic, or (ii) the variable has a θ-role which it could only have received by being part of the clitic chain. The first way of identifying variables amounts to the following condition on Case-manifestation for doubled NPs: they must be Case-marked both overtly and in situ. This condition explains the contrast between (122) and (123). An NP can be inherently Case-marked in situ at DS. Movement in syntax will not affect this, hence (123) is allowed. On the other hand, structural Case is assigned at SS, so if an ACC NP is extracted in the syntax it cannot meet the condition that its Case agreement be manifested in situ. Confirmation for this idea comes from the leista dialects of Spanish where the analog of (122) is possible, but the form of the clitic is DAT le not ACC lo (see Borer (1984), Jaeggli (forthcoming,b)). These dialects exceptionally allow a direct object clitic chain to receive inherent DAT rather than structural ACC (this may be facilitated by the ambivalent status of a - see fn. 23).24 How does our proposed case of clitic-doubling with the chain (en, byphrase) compare with RP Spanish? Clearly, extraction of &y-phrases is possible: (124)

By whonii was John seen ti?

The second way of identifying variables is relevant here. Note that the first way could not be relevant as, given the assumptions above about Case-agreement, we know that the ft^-phrase is ACC because en is assigned ACC by the Verb. However, since the A'-chain here is assigned the subject θ-role (as will be clearest where the subject is a compositional Agent, given the proposals in 2.2.2., but will be clear in any case owing to the Verb's selectional properties), the variable will be identified as a member of the clitic chain headed by en. This accounts for the widespread possibility of extraction of öj'-phrases, in contrast with the constraints on extraction of clitic-doubled complements noted above. Many clitics do not allow doubling. The paradigm example is the object clitics of French and Italian. The Α-position associated with these clitics is obligatorily phonologically null, as the following paradigms show:

62

The Representation of Implicit and Dethematized

(125)

a. Jean voit Marie. b. Jean la voit. c. *Jean la voit Marie. "John sees Mary/her."

(126)

a. Gianni vede Maria. b. Gianni la vede. c. *Gianni la vede Maria. "John sees Mary/her."

Subjects

(In Castilian Spanish, too, clitic doubling is impossible with objects, in contrast with (117).) To account for this variation among clitics, we exploit an ambivalence in what we have said about clitics in general so far. Clitics attach to a head, and may be either "part o f ' that head or not. Let us construe "being part of a head" as being essentially an affix. The difference between an affix and a clitic, on this view, is that an affix shares its features with the stem it attaches to (in the case of affixes which are heads - e.g. derivational affixes - the affix's features determine those of the derived word, cf. Lieber (1980), Williams (1981) and 2.1.2.), while a clitic has autonomous features with respect to the head. A general condition on affixes is the No Vacuous Affixation Principle, given in 2.1.3(19) (cf. also fn. 3). This principle, whatever its exact scope and formulation, certainly rules out redundant specification of features, e.g. no element can be associated with fifteen occurrences of [+masculine]. Clitics which form part of a head, if they are truly affix-like, are subject to this principle. Suppose moreover that the whole clitic chain formed by such clitics is subject to the No Vacuous Affixation Principle. In that case, no features of the clitic could be replicated on the doubled NP. But since all overt NPs have some feature specification, this prevents any doubled NP from appearing. This is our account of the impossibility of clitic-doubling in examples like (125), (126); in these cases, the clitic is essentially like an affix. Clitics vary with respect to doubling because of their ambivalent status as a class between being affixes and being autonomous elements. The clitics we are most interested in here are the passive en and the various si morphemes and their cognates (see Chapter Five). We saw above that en allows clitic doubling quite freely. Si, on the other hand, whether reflexive, middle or impersonal, does not. So, in terms of the above proposals, en is a true clitic and si is an affix-like clitic. Another issue which relates to clitic chains is the status of the empty category in the Α-position when there is no doubling. Of the four empty categories, PRO is ruled out by the government condition on clitic chains

Passives and Argument

Structure

63

combined with the fact that PRO cannot appear in a governed position, and anaphors are ruled out by Principle A (if the clitic is the only antecedent). This leaves pro and variable as the only possibilities. The licensing condition for pro, following Rizzi (1985), has two clauses: a condition on form and a condition on content, as follows: (127)

a. b.

pro is licensed by a head X which locally binds it; pro has the grammatical specification of the features on X. (adapted from Rizzi (1985:28-29)).

In our terms, a locally binding head must be a governing head. This said, pro's licensing condition clearly allows it to be part of a clitic chain in most of the cases we have discussed up to now. (127) is also relevant for the null-subject parameter. In languages with a sufficiently "rich" AGR specification in terms of person and number, AGR can license pro as subject, if governed and coindexed with it. This is the case for example in Spanish and Italian. So the pro-drop parameter reduces to the possibility of AGR being identified as a clitic (this view is proposed in Rizzi (forthcoming)). What should we say for English passives? Can en license pro? This is not the case, since this morpheme lacks any overt indication of person and number features, and these are the features which are relevant in (127).25 One could point out that en is assigned Case, and maybe this is enough for (127). However, if we propose that Case features are enough to identify pro, the only explanation we would be able to offer for the fact that English does not allow null subjects would be to stipulate the absence of coindexing of AGR and the subject (and it is possible that this coindexing is needed for NOM assignment in English, in which case this stipulation would not be available). Since this seems like an unsatisfactory account of the absence of null subjects in English, we conclude that (127) prevents pro from appearing in English (at least in passives). If the empty category linked to en is not pro, then it must be a variable. Clitics A'-bind the position with which they are coindexed and which they c-command (or govern, if that is the relevant relation here - see above). So an empty category in that position may be a variable at SS. At LF, the status of this empty category as a variable will depend on whether there is an operator which can locally bind it. We will see in 3.3. that there exists a mechanism of existential closure - the introduction of an existential quantifier coindexed with the free variable. This operation licenses the variable bound by en at LF. So the LF representation for passives is roughly as follows:

The Representation of Implicit and Dethematized

64

(128)

Subjects

3 x j N P j V + e n tj χι

This means that we expect to find scope interactions between the implicit argument of passives and other quantifiers, negation, modality etc. These possibilities are explored in 3.3.3. We can summarize our proposals for the passive by briefly specifying how en fits into the view of clitics we have just outlined. At DS, en appears in Infi as a bundle of features, much like other types of verbal inflection. It seems to have just one feature: namely, a Case feature, as it is an argument. As a clitic, it must form a chain with an argument that it governs. The only candidate is the subject argument, if there is one. Because it appears in Infi, en cannot be associated with any other argument at DS. In the derivation to SS, en undergoes Affix Hopping, like other verbal inflections. The argument bound to en must also move into VP at SS, unless have appears in Infi to form a chain linking the VP-internal en and the subject argument. The Projection Principle, as we have formulated it, neither forces nor prevents this movement. Also at SS, en is assigned ACC Case by V. This prevents V from assigning Case to its object, which then moves to subject position in order to receive Case. At LF, the original DS subject, if it is an empty category, is interpreted under existential closure. In this way, both central properties of passives are derived: subject dethematization and Case absorption. These properties are essentially the consequence of the fact that en is an argument. In the next Chapter, we will investigate the implications of these proposals for "implicit arguments." NOTES 1. (4c) and (6c) have grammatical readings. In (4c) admired can be an Adjective (cf. John seems admired by everyone, John is the most admired man in town). This reading is unavailable if the fty-phrase is included, forcing the passive reading. (6c) has a grammatical reading even where the 6y-phrase is included, if the /^-phrase is locative (cf. John was seen next to the mountains). H o w e v e r , o n this reading, another, passive 6^-phrase can be added: John was seen by the supervisor by the mountains. 2. Strictly speaking, if we f o l l o w Lieber's feature-percolation conventions, for a passive participle to be [ + V ] as a result of en's being [ + V ] , we have to take en as being [ + V , ON] - i.e. unspecified for [ + N ] - so that the affix's value for [ ± N ] will percolate to the derived word. Otherwise, the passive participle should still be [ + V , - N ] , with [ - N ] being contributed by the stem. On the problem o f neutralized features, see b e l o w and fn. 4. 3. This principle is formulated by Marantz (1984), as follows: "For a certain class of features F, an [alpha Fi] affix may only attach to a [-alpha Fj] root." (p. 128). This formulation expresses the same intuition as ours, namely that affixes only appear where they are non-redundant (cf. the Principle o f Full Interpretation o f 2.2.3). Our formulation differs f r o m Marantz's because we need to allow for the case where an affix neutralizes

Passives and Argument

Structure

65

a feature of a stem. This possibility does not arise in the framework developed by Marantz. 4. Although cf. Aoun (1981a) for the suggestion that some languages may have only neutralized categories. Note that allowing neutralization as a parameter in this way is more restrictive than allowing every language the possibility of having eight major categories. 5. We neglect Kayne's proposals because he has a different approach to passives which does not involve Case-absorption, therefore his theory of double-object constructions does not take this matter into account. 6. Stowell has a different account for the grammaticality of (i) in some dialects: (i)

A book was given John.

He suggests that (i) is ruled out in the majority dialect because NP-Incorporation only applies to Verb stems, so cannot apply to morphologically passive Verbs. For those of us who allow (i), Stowell says "some dialects allow a marginal extension of NP Incorporation so as to apply to participial stems" (p. 330). 7. Up to now, we have disavowed any significance of the terms "Agent," "Theme," "Source," etc. For the purposes of this still somewhat expository section, these terms can be taken intuitively. In 2.2.2. we will develop a theory of θ-role assignment which distinguishes Agent and Theme from other θ-roles in terms of their mode of assignment, and in Chapter Four, we will attempt a characterization of what it means for a Verb to have a particular kind of θ-grid. 8. Examples (35 - 37) are adapted from Marantz (1984). 9. These examples were originally pointed out in Hornstein (1975). 10. It is important to take sense as the nominalization of a true perception Verb in (45a). In this reading, it cannot, for example, occur as the complement of finish·. (i)

*The psychic finished sensing the ghost.

In (i), sense can only have an agentive reading. Cf. the contrast between see and look at in complements to finish. 11. Actually this is the defining property of argument clitics only. There is also a class of clitics which are associated with adjuncts of various kinds, e.g. French en, y, Italian ci (cf. Elliot (forthcoming) on French, Burzio (1986), Hyams (1983) on ci). There is also a class of ethical-dative clitics in Spanish, which are not associated with an argument position (cf. Jaeggli (1982:18 - 20, forthcoming,b:8 - 14). 12. Causative constructions in Romance pose problems for the view of clitics as forming DS chains under government. Consider the following examples: (i)

Lai hice limpiar ei por Pedro. CL he-made clean by Ρ

(ii)

La; hice limpiar e, a Pedro. CL he-made clean to Ρ "He made Pedro clean it."

For these examples we must propose DS reanalysis of the causative Verb with the lower Verb in order to make it possible for the clitic to govern the object of the lower Verb. This poses no problem in (i), but Burzio (1986) and Zubizarreta (1982) both provide evidence that the complement in (ii) is sentential at S-Structure, and it is therefore impossible for the clitic to govern the object of the Verb. One way to resolve this dilemma seems to be to propose DS reanalysis followed by "unreanalysis" in the derivation to SS, a possibility

66

The Representation of Implicit and Dethematized

Subjects

we will not go into here. Cf. also the proposals in 2.2.1.3.1. for chain-formation between en and have. Possibly a similar relation can hold in these constructions between to the causative Verb and the lower Verb, allowing something attached the causative Verb to form a chain with the lower Verb's object (cf. Rouveret & Vergnaud (1980) for a similar proposal, worked out in full detail). 13. This clearly cannot be true in languages which allow intransitive Verbs to passivize, like German and Dutch. See 5.3. 14. The implications of these proposals were brought to my attention by Mark Baker and Kyle Johnson and form the subject of joint work in progress. 15. As we mentioned above, Marantz (1984) derives the 1AEX from the No Vacuous Affixation Principle, stated as in fn. 3. However, it may be the case that this principle itself can be derived from deeper principles, once the properties of affixes are properly understood. See 2.2.4. for more on this principle and its relation to clitics and affixes. 16. Comparing (62) with (18) we can see that neither the 1AEX nor our derivation of it will rule out passives of (unergative) intransitives. These Verbs assign subject θ-roles, as (i) shows: (i)

*It/there danced.

However, passives of intransitives can be ruled out by Case theory. En in these environments is unable to get (ACC) Case. The existence of passives of intransitives in languages like Dutch and German then comes down to the question of (NOM) Case assignment into VP. See 5.3. 17. The Mirror Principle appears to pose problems for the view commonly held in lexical phonology that all word-formation takes place in the lexicon, including that involving inflection (see in particular Lieber (1980)). The kinds of derivation proposed in (61 - 65) are in apparent contradiction with this idea, as here word-formation is clearly syntactic. However, we can propose some form of recursion into the lexicon after the attachment of each affix. An alternative approach to word-formation, consistent with the Mirror Principle, is proposed in Fabb (in progress). 18. Although English has such an impoverished morphological system that it is hard to find direct evidence for the Mirror Principle, one might ask in the present context if the "morphosyntactic" relation of the Mirror Principle goes any deeper; in particular we might reformulate it as a relation between the core predicate-argument structure of a clause (as represented in pure form at DS) and the other elements of a clause: those which specify tense, aspect, mood etc. and those which "alter grammatical functions." Looked at in this way, the Mirror Principle would govern the ordering of auxiliaries and of affixes: passive affects V (through its argument structure), and so must appear innermost; perfect and progressive affect VP; and tense and mood affect S. The fact that perfect precedes progressive is presumably a consequence of have's property of "retransmitting" its θ-role to the subject - c f . 2.2.1.3.1. 19. Pesetsky (1982) proposes that subcategorization in the Aspects sense subsumes two separate processes: s-selection and c-selection. S-selection ("semantic selection") encodes the semantic argument-taking requirements of predicates, e.g. wonder requires a question argument, say requires a prepositional argument, etc. (cf. Grimshaw (1979)). C-selection ("categorial selection") encodes structural requirements for the frame in which a lexical item can be inserted, e.g. that wonder and know require a CP complement. Such a view is probably correct, and our proposals for lexical structure could be reformulated in these terms; however, we retain subctegorization in our exposition for convenience and simplicity. For some more developed proposals about the lexical representation of Verbs, see Chapter Four. 20. It does not seem necessary to take the view that Verbs are functions in order to maintain Marantz's insight.

Passives and Argument

Structure

67

21. An apparent exception to the claim that idioms with an open subject slot is kick the bucket (cf. (38c)). The θ-role assigned by this predicate is clearly compositional, and equally clearly not agentive. However, kick the bucket differs from the other idioms discussed in being completely impossible to break up by any syntactic operation as illustrated in (35 - 38). As suggested on p. 24, we regard such idioms as lexically entered as single Verbs, so the issue of compositionality does not arise. This complex Verb has lexically determined its θ-roles just like other Verbs. In this case, its argument structure is just like die; its one θ-role is a Theme. 22 This principle is very similar to - and to some extent inspired by - (i) as recently proposed by Higginbotham (1985:562): (i)

Every argument is saturated.

Higginbotham points out that It is an immediate consequence of [(i)] that functionals [functions whose arguments are other functions - IGR] are not to be found in the grammars of languages where [(i)] holds; the arguments of a functional must be unsaturated, and this is just what [(i)] forbids. In terms of the CFCC, we could say that functionals are not forbidden; rather they cannot assign θ-roles. One of the most likely examples of a class of functionals in natural language is the class of auxiliary Verbs, and we have already pointed out that these do not assign θ-roles. Note also that for the correct working of the CFCC, we must take names, pronouns and empty categories to be CFCs. This is unproblematic if we say that the C F C C is satisfied vacuously for categories not headed by a θ-assigner or transmitter. 23. The Preposition a has quite complex Case-assigning properties. In all dialects of Spanish a doubles a dative clitic: (i)

le entregue el libro al profesor cl-DAT I-gave the book to-the professor "I gave the book to the professor."

Also, a must appear quite independently of the presence of a clitic where there is a specific, animate direct object, as in (ii): (ii)

a.

b.

Vimos a Guille. (*Vimos Guille). we-saw Ρ G "We saw Guille." Vimos la casa. (*Vimos a la casa). "We saw the house."

We can account for this by saying two things about a: (i) a is an optional Case-assigner, and optionally assigns ACC or D A T Case, subject to the feature-agreement condition on clitic chains; (ii) a has autonomous semantic properties. 24. Unfortumately, the situation is still more complex than the text suggests. According to Hurtado (1984), certain dialects of Latin American Spanish allow extraction of the doubled NP, i.e. they allow examples like (122). Moreover, the extraction site does not act like a variable with respect to weak crossover and licensing parasitic gaps. Hence we find the following data:

68

The Representation of Implicit and Dethematized

(i)

LA. quierii loi acuso sui complice Xi? Ρ whom cl he-accused his accomplice "Who did his accomplice accuse?"

(ii)

*LA. quien lo nombraste antes de conocer t bien? Ρ whom cl you-nominated before to know well "Who did you nominate before knowing well?"

Subjects

(i) shows that weak crossover effects do not hold, and (ii) that parasitic gaps cannot be licensed by the empty category occupying the position of the extracted doubled NP. If lo is absent, the grammaticality judgements are reversed. These facts suggest that the extraction site is occupied by a non-variable occurrence of pro. Such an occurrence of pro must be able to meet the Case manifestation requirement on the clitic-doubled position. See below on licensing pro. This further implies that, as pointed out by Hurtado, the quantifiers in question are not inherent quantifiers, given the theory of variables proposed in 1.4. 25. As is well-known, past participles in Romance languages agree with the subject under certain conditions (cf. 2.2.1.3.): (i)

Maria e stata vista. Μ is been(f) seen(f) "Maria has been seen."

In theory, the underlying subject could be pro here. Whether this empty category is pro or a variable, it acts like a variable at LF. This is consistent with our assumptions regarding pronouns and bound variables - cf. 1.4. and 3.3.3.

Chapter Three

The By-Phrase and Implicit Arguments

3.1. INTRODUCTION

In Chapter Two, we proposed that passive clauses have D-Structures as in (1) which, by movement first of en, then of the empty category in subject position, then of the object NP, give S-Structures as in (2) and surface structures (via a few morphological adjustments) as in (3): (1)

ei en, kiss Mary.

(2)

Maryj kiss+enj tj e*.

(3)

Mary was kissed.

The main advantage of this approach is that the D-Structure closely resembles the D-Structure of an active clause. Since D-Structure is the representation of argument structure, (1) states that active and passive clauses do not differ fundamentally in argument structure. The differing positional realizations of the arguments at S-Structure are the result of the presence of the clitic en, the requirements on clitics outlined in the previous chapter, and of Case theory. En, we proposed, forms a clitic chain with the subject position in (1). Because of conditions on clitic chains, the argument in subject position moves into VP when en does. This argument may be an overt NP, in which case it is realized in a 6y-phrase, or an empty category, a so-called implicit argument. In the context of this theory of passive the term "implicit argument" is a misnomer, as we consider these arguments to be explicitly structurally represented. The goal of this Chapter is to support that view. As a matter of notational convenience, however, we will refer to the structurally present non-overt logical subjects of passives as IMPs henceforth. The purpose of this Chapter is primarily to consider the properties of IMPs and to derive these properties from independent factors and/or from the proposal for passive presented above. We will also pay attention to 6y-phrases when this is relevant. The IMP in the chain (en, IMP) in (1) and (2) is an empty category

70

The Representation of Implicit and Dethematized

Subjects

bearing the subject θ-role. It represents a specific hypothesis about the status of the "suppressed" subject. Chomsky (1986) distinguishes this instance of a suppressed subject from other such instances, PRO and the logical subjects of middle Verbs, by six tests. In contrast with the nonrealized subject arguments of middles, this empty category can be modified by certain Adverbs, and they can control PRO in Rationale clauses (or RatC - cf. Jones (1985)): (4)

a. The book was sold voluntarily, b. *The book sold voluntarily.

(5)

a. The book was sold PRO to make money, b. *The book sold PRO to make money.

On the other hand IMPs are claimed to differ from PRO in four main ways: (6)

IMPs cannot bind anaphors1 a. *?Letters were sent to themselves. b. They; expected PRO; to send letters to each others

(7)

IMPs cannot be predicated with/control into adjuncts a. *The book was read without putting it down. b. They expected to read the the book without putting it down.

(8)

IMPs cannot be controlled a. They expected malicious rumors to be spread. b. They; expected PROi to spread malicious rumors

(9)

IMPs cannot be the objects of predication2 a. *The room was left sad. b. Theyi expected PROi to leave sad.

In (8a) they cannot be interpreted as the spreaders of rumors. In (8b), of course, only this interpretation is available. (9a) cannot mean that those who left the room were sad (although it does have an interpretation with predication of sad of the room where the room is interpreted as "those in the room"). There are two main sets of facts concerning IMPs: those involving predication and modification seen in (4,5,7,9) and those related to determining their reference and referential dependencies, i.e. relating broadly to issues covered by binding theory; the facts of (6) in particular, among many others, fall under this heading. We will deal with each set

The By-Phrase and Implicit

Arguments

71

of facts in turn, considering also, where relevant, how the overt &>>-phrase patterns. We propose a fairly simple theory of predication and modification - a slight extension of some of the proposals of the previous chapter which accounts for the properties of IMPs seen in (4,5,7,9) straightforwardly. Then we investigate the referential properties of IMPs. These are fairly surprising, but can be made to follow from certain independently motivated conditions on chains and a version of the innerisland constraint on extraction of Ross (1984). Given these assumptions, the central properties of IMPs are predicted by the analysis of passive given in the previous section. The conclusion of this Chapter will be, then, that the central property of passive constructions is a "redistribution" of the arguments of a clause triggered by the appearance of en, a particular bundle of features, in Infi. With this conclusion in mind, we will proceed to the analysis of middles and nominals in Chapter Four. 3.2. PREDICATION AND MODIFICATION OF IMPLICIT ARGUMENTS

In this section we will review the predication and modification properties of IMPs. We will outline a theory of predication and modification in line with the proposals for θ-role assignment in 2.2.2. and the Principle of Full Interpretation (FI) of Chapter Two (101). After giving this theory some motivation, we show how it accounts for the modification properties of IMPs seen in (4,5,7,9). Since this account is fairly straightforward, we take this section as a whole as an argument in favor of the treatment of IMP as an empty category occupying a structural position. 3.2.1. Adverbs The purpose of this section is to discuss the contrast noted in example (4). A natural account for this is in terms of selection restrictions: Adverbs like voluntarily select for arguments of a certain type: the right type of argument is not present in (4b) and is present in (4a)3. But what does "the right type of argument" mean here? In (4a) the Verb assigns an Agent θ-role to the subject. Following what we said in 2.2.2., this means that the clause has an event reading. (4b), on the other hand, is neither agentive nor eventive (see 4.2). That both an Agent and an event are required by the Adverb is shown by the ungrammaticality of (10): (10)

a. *The ice deliberately melted. b. *John intentionally knew the answer.

(10a) has an event but no Agent; (10b) could be construed as having an Agent but no event, and both examples are bad. (10b) further shows that

72

The Representation of Implicit and Dethematized Subjects

selection is not the whole story because the selection requirement of the Adverb is presumably met here. This indicates that these Adverbs must appear in a configuration of Agent θ-role assignment, i.e. where V is coindexed with Infi and Infi is coindexed with the subject position (cf. 2.2.2.). The fact that these Adverbs are precisely those which can bring about readings of secondary agentivity (cf. 2.2.2.(89),(90)) is consistent with this proposal (we will say more about this phenomenon in 3.2.1.2.). In terms of our analysis of passives, the claim is that structural θ-role assignment to the chain C = {en, IMP) meets the requirement of the Adverb in (4a). This configuration does not hold in (4b), so this structure does not meet the Adverb's requirement, and so the sentence is ungrammatical. This section aims to develop this analysis of the Adverbs of (4) and (10) in detail. We will make explicit the relation between these Adverbs and IMPs. We adopt certain insights about "adverb-orientation" from Jackendoff (1972) and some about modification from Zubizarreta (1982). We begin with a detailed proposal for the syntax of Adverbs. 3.2.1.1. Positional Classes of Adverbs 3.2.1.1.1. Initial and Final According to Jackendoff, Adverbs can appear in three positions in a clause: initial, "auxiliary" (see below for an explicit representation of this position; we will refer to it as "aux position" for now) or final. Adverbs fall into five distributional classes according to their ability to appear in these three positions, as follows: (11)

Class I: appear in all three positions, but with varying meanings: cleverly, clumsily, carefully, deliberately, voluntarily, intentionally... Class II: appear in all three positions with no meaning change: quickly, slowly, quietly, frequently.. Class III: only appear in initial and aux position: evidently, probably, unbelievably.. Class IV: only appear in aux position and final position: completely, easily, totally.. Class V: final position only: hard, well, more.. Class VI: aux position only: merely, virtually..

We will disregard Classes V and VI. Classes I - IV are exemplified in their various positions in (12 - 15): (12)

Class I a. Cleverly, John kissed Mary. b. John cleverly kissed Mary. c. John kissed Mary cleverly.

The By-Phrase and Implicit

Arguments

(13)

Class II a. Quickly, John kissed Mary. b. John quickly kissed Mary.

(14)

Class III a. Evidently, John kissed Mary. b. John evidently kissed Mary.

(15)

Class IV a. John easily ate his lunch. b. John ate his lunch easily.

73

Intuitively, the Adverb in (12) modifies John and the predicate; while in (13) just the predicate is modified. In (14), the Adverb modifies the clause, and in (15) the Adverb again modifies the predicate, but here it seems to have a closer relation to the Verb than in (12) and (13). Three factors interact to capture these intuitive correlations: (16)

a. b. c.

Adverbs are predicates. Adverbs have differing inherent lexical properties; this is what underlies the different classes in (11 - 15). The "adjacency condition on Case assignment."

(16a) means that Adverbs must appear in a relation of R-modification with a subject, following the Predication Principle of Chapter Two (102). Recall that the definition of R-modification is effectively government by non-head, as in (17) (repeated from Chapter One (8c)): (17)

X R-modifies Y iff X is a maximal projection, X is not the head of Ρ, Ρ the maximal projection dominating X and Y, and (R(x,y) &R(y,x)).

(17) represents the structural requirement on the semantic predication relation between an Adverb (X) and its "subject" (Y). This is a necessary but not a sufficient requirement for well-formed structures with Adverbs. The sufficient requirement is that the Adverb's selection requirement be met. We can regard these requirements as restrictions imposed by individual Adverbs on the kind of subject which allows them to satisfy the Predication Principle. Suppose further that the selection requirement is satisfied by coindexation with a subject with which the Adverb is in an R-modification relation. (16b) states that Adverbs have differing selectional properties: some are clausal modifiers, some are VP modifiers, etc. In fact, VP-Adverbs distinguish selectionally between VPs of different

74

The Representation of Implicit and Dethematized

Subjects

types - see (19) and the discussion below. The conditions (16a,b) also hold for θ-role assignment by VP to subject position. So adverbs are just like VPs except that they do not assign θ-roles. As in the case of VP, we will refer to the relation between an Adverb and its subject as predication. These conditions mean that a given Adverb cannot be interpreted in all the available positions. Examples like (18) demonstrate the conditions in action: (18)

a. *John ran away evidently. b. *Completely, John ate his lunch.

Evidently requires Infi as a "subject" as we shall argue below. In its VPadjoined position here, it does not R-modify Infi so the requirement is violated; ultimately the Predication Principle cannot be satisfied here. Similarly, in (18b) completely does not R-modify V, as is required. To distinguish among the VP-Adverbs - Classes I, II, and IV - we can use the notions Agent and event as defined in 2.2.2. Class I Adverbs require both an Agent and an event, as we have just seen. Class II Adverbs require just an Agent, and Class IV require neither. To see these distinctions, compare (19) with (10): (19)

a. b.

The ice quickly/easily melted. John *quickly/easily knew the answer.

Our account of these distinctions is that Class I Adverbs are predicated with Infi or V and an Agent; Class II with Infi or V (recall that the fact that these classes require an eventive predicate means that Infi and V will always be coindexed where they appear) and Class IV with V. We will justify the Class I vs. Class II distinction in 3.2.1.2., and the characterization of Class IV in 4.4. Consider next (16c). We could claim that the Adjacency Condition on Case Assignment prevents Adverbs from appearing in all but the most peripheral positions, and positions where there is no Case-assignment (such as between an NP and a PP as in John spoke quickly to Bill). However, it is not clear what the scope of the Adjacency Condition on Case Assignment is. For example, Adverbs appear between the subject NP and Infi, where there is Case-assignment: (20)

a. b.

It evidently must be raining. John intentionally didn't visit the man.

Perhaps the Adjacency Condition (or at least a "maximal closeness") condition holds for inherent/lexical θ-role assignment. Whatever the exact

The By-Phrase and Implicit

Arguments

75

condition involved, this is what forces Adverbs to appear in peripheral positions. Given (17), Adverbs in initial position appear in the configuration (21): (21)

ADV

NP

Class I Adverbs are predicated with the subject and Infi from this position, Class II and Class III with just Infi. Members of this last class typically express notional modality. Infi is the locus of modality in the clause, so Class III Adverbs are predicated with that position. This is particularly clear in cases where a Class III Adverb occurs with a modal. Compare: (22)

a. b.

It must be raining outside. It must evidently be raining outside.

The epistemic modality in the two examples is subtly different. We attribute this effect to predication between Infi and the Adverb. That Class IV Adverbs cannot appear initially, as in (18b), indicates that they appear inside VP, predicated with a subpart of the VP, not the whole VP. If they were predicated with the whole VP, they should be able to appear initially. Independent motivation for this theory of Adverbs comes from the following examples: 4 (23)

a. *Evidently, who left? b. *I don't know evidently who left. c. *Definitely is he coming?

(23a,b) show that Adverbs that otherwise appear in initial position cannot precede a filled Comp. This is explained by our proposal that Class III Adverbs are not predicated with CP but instead with Infi. Because of the structural condition on predication (R-modification), predication with Infi is impossible from a position external to CP. In (23b) evidently cannot be predicated with the matrix Infi as it is contained in VP. (23c) falls under this generalization too on the assumption that subject-aux inversion involves movement of the auxiliary into Comp. 5 If Adverbs in final position are part of VP, then our proposals would

76

The Representation of Implicit and Dethematized

Subjects

explain the impossibility of purely Infl-modifying Adverbs - Class III in this position (unless there is comma intonation, which presumably marks the Adverb as adjoined to IP; this is consistent with (17)). There are two possible configurations for VP-internal final-position Adverbs, shown in (24): (24)

a.

IP NP

b. Γ

I

NP VP

Vn

IP Γ I

ADV

VP VP

ADV

Well-known constituency tests for VP-hood - VP-Fronting, VP-Deletion, though VP-movement, pseudoclefting and Right Node Raising - distinguish these two structures. If these operations can leave an Adverb behind, then that Adverb must appear in structure (24b) and the lower VP undergoes movement or deletion. If an Adverb cannot be left behind, then it appears in (24a) To be consistent with the proposals so far, we expect Class IV Adverbs to appear in (24a) and Classes I and II in either configuration. In other words, the main prediction made by our proposal for Class IV Adverbs is that they must undergo RNR etc. as members of VP. This prediction is confirmed by the data in (25 - 29), where we contrast examples where the Adverb is taken with the VP (the (a) examples) with those where it is not (the (b) examples): (25)

a. VP-Deletion I Bob kissed Sue intentionally and Joe did too. II Bob kissed Sue quickly and Joe did too. IV Bob kissed Sue easily and Joe did too. b. I Bob kissed Sue intentionally and Joe did deliberately. II Bob kissed Sue quickly and Joe did slowly. IV?*Bob ate his soup totally and Joe did easily.

(26)

a. I

VP-Fronting Bob wanted to kiss Sue intentionally - and kiss her intentionally he did. II Bob wanted to kiss Sue quickly - and kiss her quickly he did. IV Bob wanted to kiss Sue easily - and kiss her easily he did.

The By-Phrase and Implicit Arguments

77

b. I ??Bob wanted to kiss Sue - and kiss her he did intentionally. II ??Bob wanted to kiss Sue - and kiss her he did quickly. IV *Bob wanted to eat his soup - and eat his soup he did easily. (27)

a. I

(28)

a. I II IV b. I II IV

(29)

RNR Bob wanted to kiss Sue - and Joe tried to kiss Kim intentionally. II Bob wanted to kiss Sue - and Joe tried to kiss Kim - quickly. IV?*John tried to eat his soup - and Joe tried to wash his bike - totally. b. I Bob wanted to kiss Sue intentionally - and Joe tried to kiss Kim deliberately - t o win the bet. II Bob wanted to kiss Sue quickly - and Joe tried to kiss Kim slowly - to win the bet. IV Bob wanted to eat his soup easily - and Joe tried to wash his bike totally - to win the bet. Though VP-movement Kiss Sue intentionally though Bob may ... Kiss Sue quickly though Bob may ... Eat his soup totally though Bob may ... ?Kiss Sue though Bob may intentionally ... ?Kiss Sue though Bob may quickly ... *Eat his soup though John may totally ...

a. Pseudoclefts I What Bob did was kiss Sue intentionally. II What Bob did was kiss Sue quickly. IV What Joe did was wash his bike easily, b. I What Bob did intentionally was kiss Sue. II What Bob did quickly was kiss Sue. IV?*What Joe did easily was wash his bike.

Although Classes I and II are not perfect in all these examples (notably (26b) and (28b) - these are the cases where there is clearly VP-movement, which suggests that one explanation for this could be that these Adverbs cannot modify VP-traces; this idea requires that (25, 27) be analyzed as having a pro-VP rather than a VP-trace, a matter we cannot go into here), the important point for our proposal is that Class IV Adverbs cannot

78

The Representation of Implicit and Dethematized

Subjects

be "left behind" by these rules. This is shown by the ungrammatically of the (b IV) examples. These examples provide support for what we have proposed about the structural properties of Adverbs. As a final point on the initial and final position of Adverbs, what can we say about multiple sequences of Adverbs? Andrews (1983) gives some evidence that, given a principle like (17), we should not assume a flat structure for multiple-Adverb sequences. Consider the following pair: (30)

a. b.

John knocked on the door intentionally twice, John knocked on the door twice intentionally.

(30a), Andrews points out, involves two events of intentional knocking, while (30b) involves a single intentional event of knocking twice. To account for this, Andrews proposes that the structure of the VPs in these examples is as in (31):6 (31) a. [ V P [ V P knocked on the door] intentionally ] twice ]. b. [VP [VP knocked on the door] twice ] intentionally ]. We can adopt this approach here, as it is consistent with our assumptions. Andrews also points out that these Adverbs can occur on either side of VP: (32)

a. b.

John knocked twice on the door intentionally, John knocked intentionally twice on the door.

(32a) allows either reading in (31) fairly indifferently, while (32b) shows a preference for the reading with wide-scope intentionally, in (31b). Neither we nor Andrews have an account for these preferences. Jackendoff (1972:87) points out that many examples of stacked Adverbs are impossible: (33)

a. *Evidently carefully John left the room. b. Evidently John carefully left the room. c. ?*John evidently carefully left the room. d. John evidently has carefully been concealing the truth. e. *John has left the room evidently carefully.

(33a,c,e) can be explained if we adopt the constraint that no two Adverbs can occur adjacent to each other. However, this only holds for cooccurrences of certain Adverb classes. In particular,this is true for a sequence of Class III and Class I Adverbs, as in (33a,c,e), but not true for a sequence of Class II and Class I adverbs, except in "aux position," as (30) and (32b) show (assuming that twice is a Class II Adverb).

79

The By -Phrase and Implicit Arguments

We can explain the impossibility of adjacent sequences of Class I and Class III Adverbs in terms of (17) and their selection requirements. Assuming that Adverbs do not occur together in flat structures, we have the following two possibilities for (33a), for example: (34)

a.

b.

IP

evidently IP

carefully

NP

I carefully

Γ

evidently

VP

IP

NP

Γ

VP

In (34a) the selection requirement of evidently (Infi) cannot be met, and in (34b) that of carefully cannot be met, hence such sequences are impossible. However, stacking of a Class I with a Class II Adverb does not pose the same problems, as we saw with reference to (31). Given this account for (33a,c,e), two other possibilities are correctly predicted to be impossible: (35)

a. *Carefully evidently John left the room, b. *John left the room carefully evidently.

(Compare John left the room carefullly, evidently with (35b)). Also, we correctly predict the awkwardness of two Class III Adverbs cooccurring: (36) ??Evidently John probably left. In (33b,d) we assume that carefully is inside VP, modifying Vn. This possibility is also available for (33c), explaining the marginal possibility of this example. In this section we have suggested a way of deriving many properties of the distribution of Adverbs. Although many intricacies remain unexplained, the distribution of the core classes of Adverbs can be accounted for, it seems. 3.2.1.1.2. "Aux position" Adverbs in aux position pose some interesting questions. "Aux position" corresponds to one of the three structural positions in (37):

The Representation of Implicit and Dethematized

80 (37) a.

IP

Λ /i\ ADV I VP NP I'

b.

IP

Λ / W I ADV

Subjects c.

NP I'

IP NP

VP

/

I VP ADV V"

(There is actually a fourth possibility, given I': that where ADV adjoint to I'. This possibility is hard to distinguish from that represented in (37a) so we will ignore it here.) Given what we have said so far, Class I and II Adverbs should be possible in any of these configurations, while Class III should be impossible in (37c), and Class IV in (37a,b). We can test these predictions by seeing how Abverbs interact with auxiliaries. Note that (37a) is equivalent to (37b) in terms of (17). For example, if there is an overt element in Infi, (37a) will be distinguishable from (37b) and (37c). Also Class IV Adverbs, which can only modify V n < m a x , appear in (37c), so they cannot precede an overt Infi. Modals form a class of overt elements able to appear in Infi (cf. Roberts (1983)), so we can test this prediction as follows: (38)

*John totally will eat his lunch.

(38) is clearly ungrammatical. This is explained by our approach in terms of (17) and the lexical property of totally that it selects V. Recall that we assume that the aspectual auxiliaries have and be appear in syntax with VP-"complements" (cf. 2.2.1.3.). Thus the structure of a complex auxiliary sequence is as in (39) (cf. 2.2.1.3. (68)): (39)

VP

A

V, VP2 have V2 VP3 be Class IV Adverbs must appear in a position modifying the lowest VP, as (40) shows: (40)

a. ?*George will completely [VP have [VP read the book ]]. b. *John will be rapidly [VP being [vp beaten by Bill ]].

The By -Phrase and Implicit

Arguments

81

This is explained if Class IV Adverbs are taken to modify V n < m a x . on the other hand, is grammatical, as we predict: (41)

John evidently will kiss Mary.

This is because Class III Adverbs select Infi. They cannot appear in VP, as we will be able to test if we can distinguish (37b) from (37c). In fact, we can distinguish (37b) from (37c) by interspersing Adverbs in sequences of aspectual auxiliaries. In configurations like (38), then, Class III Adverbs should be impossible. This is illustrated in (42): (42)

?John will have probably been kissing Mary.

(42) is at best awkward (although not quite as bad as expected). Again, this prediction derives from (17) and the lexical properties of probably.1 3.2.1.2. "Thematic" Properties of Adverbs In this section we deal with the contrast between Classes I and II. Both classes can appear in all three positions discussed in the last section, but the difference between them is that Class I Adverbs select a configuration of Agent θ-role assignment and Infi, while Class II Adverbs, on the other hand, select just Infi. This contrast is shown by the different behavior of these Adverb classes with different kinds of predicate. The following examples show that Class II Adverbs may be good or bad with predicates of different types (for reasons which we go into in more detail in 4.4.) independently of whether the predicate is agentive, while Class I Adverbs are good iff the predicate is agentive. In each of the following examples, the (a) and (b) sentences have nonagentive predicates, while the (c) and (d) sentences have agentive predicates: (43)

a. *The crystal deliberately dissolved. b. John deliberately built a house. c. The crystal quickly dissolved. d. John quickly built a house.

(44)

a. *John was deliberately asleep. b. *John was quickly asleep. c. John was deliberately polite. d. *John was quickly polite.

(45)

a. *The box deliberately rolled down the hill, b. The box quickly rolled down the hill.

82

The Representation of Implicit and Dethematized c. d.

(46)

Subjects

John deliberately walked. John quickly walked.

a. *John deliberately noticed the stolen painting. b. John quickly noticed the stolen painting. c. John deliberately killed. d. ?John quickly killed.

So Class I Adverbs always require an agentive NP in the clause in which they appear ((44b,d) confirm our earlier proposal that Class II Adverbs require an event predicate). Zubizarreta (1982:32 - 49) proposes that Class I Adverbs assign a Θrole to either the structural subject or the thematic Agent of the clause. Since the Agent NP is always an argument of the main predicate, the θ-role assigned by the Adverb cannot be subject to the θ-criterion. Zubizarreta proposes that this θ-role is an "adjunct" θ-role. Adjunct Θrole are optionally assigned at LF. When assigned, adjunct θ-roles must be assigned to arguments, i.e. must be assigned to positions occupied by elements which are assigned a main θ-role. This is Zubizarreta's Adjunct θ-criterion ((69), p.45), which rules out (47): (47)

*There deliberately seemed to be a riot.

In the context of the proposals we have already made for Adverbs in general, we can simply regard Class I Adverbs as two place predicates; in addition to selecting an event, a property which is shared with Class II Adverbs, Class I Adverbs select for an Agent NP (the semantic property of Class I Adverbs seems to be, roughly speaking, that they specify a relation between an individual and an event). So Class I Adverbs are coindexed with an Infi coindexed with V (or with V itself), and with an NP in a configuration which allows an agentive interpretation for that NP (through coindexation with Infi and possibly VP - cf. 2.2.2., especially the discussion of (89) and (90)). Since the NP selected by the Adverb must allow an agentive interpretation, it will always be θ-marked, since an agentive interpretation results from a structural θ-marking configuration, and such a configuration must involve an argument. This analysis derives the effect of the Adjunct θ-criterion. So we retain the idea that Adverbs do not assign θ-roles, although they have other properties of predicates (selection, being subject to the Predication Principle). Like other Adverbs, Class I Adverbs must R-modify the subject they select. In some cases, however, a Class I Adverb appears inside VP and yet selects the subject (this is the case in examples like (31) above). This happens where the Class I Adverb is coindexed with V and V with Infi

The By-Phrase and Implicit Arguments

83

and the subject. This is why we say that Class I Adverbs must appear in a configuration of Agent θ-role assignment: this requirement can be met in any one of three ways: (i) by coindexation both with Infi and the subject; (ii) by coindexation with Infi; (iii) by coindexation with a V which is coindexed with an Agent. As both Zubizarreta and Jackendoff point out, and as we pointed out in 2.2.2., surface subjects of passives can be modified by Class I Adverbs. Jackendoffs example is (48) (cf. Jackendoff {op cit: (3.14), p. 82)): (48)

Mary was intentionally seduced by John.

In (48), the intention of seduction is ambiguously Mary's or John's. Here intentionally selects as Agent either Mary or John in by John. On the former reading, selecting Mary, the sentence has the representation in (49): (49)

IP

PP intentionally

seduce+enj

*

by

ΔJohnj

Here the Adverb coindexes with Infi and the subject. This creates a configuration of structural θ-role assignment to Mary, and so allows this NP to be interpreted as an Agent (cf. 2.2.2.). On the other hand, where the Adverb selects John, the structure is as in (50): (50)

IP

Maryi was intentionally

by

Johnj

In (50), the Adverb coindexes with V and with John. In both of these examples the selection requirement of the Adverb is met: the Adverb

84

The Representation of Implicit and Dethematized

Subjects

coindexes with an Infi or V coindexed with an Agent. This account makes explicit what was left implicit in 2.2.2., namely that these Adverbs coindex with their "arguments" at SS and LF, and it is these levels of representation that are relevant to interpreting their relations with the other elements in the clauses in which they appear. An important further point is that the above representations would be the same if the example contained an IMP instead of an overt fry-phrase. In general, sentences with middles are non-agentive and non-eventive (see 4.2.). This is why Class I Adverbs cannot appear in them. This is shown in the following examples: (51)

a. *The bureaucrats bribed voluntarily, b. *The slave sold deliberately.

However, secondary agentivity is marginally possible here also. Cf.: (52)

?Bureaucrats bribe (easily) voluntarily.

As in the passive (48), the surface subject here can be selected by the Adverb. In fact, as no IMP is present in middles, this is the only way of interpreting the Adverb here. To the extent that Class I Adverbs are possible with middles, middles lose both their non-eventive and their nonagentive reading (cf. the detailed discussion of this point in 4.2.). This is predicted by our proposals for agentivity and events and for secondary agentivity. (52) will have a representation exactly as in (49), except for the absence of en and the IMP. As (48) shows, a Class I Adverb in aux position can modify either the surface subject or the deep Agent. But note that adding auxiliaries so as to disambiguate the structure of (48) with respect to the possibilities in (37) alters the available readings. We indicate in parenthesis after each example who has the intention: (53)

a. b. c.

Mary was intentionally being seduced by John. (Mary) Mary was seduced intentionally (by John). (John/IMP) Mary was seduced (by John) intentionally. (John/IMP)

In (53a), the presence of being after intentionally shows that the Adverb is not inside the lowest VP (was is in Infi). So intentionally cannot modify the fry-phrase; instead, it modifies the argument in subject position. So Mary unambiguously has the intention in this example. In (53b,c) intentionally is unambiguously inside VP and can only be interpreted as modifying the fry-phrase, so John unambiguously has the intention. The same holds when there is no by-phrase and the Adverb modifies IMP in VP.

The By-Phrase and Implicit Arguments

85

In embedded contexts we find the same correlation: (54)

a. b.

They expected John to intentionally have been seduced (by Mary). John is believed/seems to intentionally have been seduced (by Mary).

In all these cases, the Adverb modifies John. However, we find that the 6y-phrase/IMP only is modified if the Adverb follows been: (55)

a. b. c.

They expected John to have been intentionally seduced (by Mary). John is believed to have been intentionally seduced (by Mary). John seems to have been cleverly tricked (by Mary).

If we put the Adverb in between the two auxiliaries, we seem to have either choice: (56)

a. b. c.

They expected John to have intentionally been seduced (by Mary). John is believed to have intentionally been seduced (by Mary). John seems to have intentionally been seduced (by Mary).

Again, these results are consistent with an approach based on the insights of Jackendoff and Zubizarreta, like the one adopted here. Another insight of Jackendoff s is that root modals pattern like Class I Adverbs. Root modals have, roughly speaking, the following senses: (57)

can may must will

-

ability permission obligation volition/tendency

In these readings, then (as opposed to the epistemic readings, which are like the Class III Adverbs), the modals select an Agent (Zubizarreta op. cit. suggests that root modals assign adjunct θ-roles, in the same way as Class I Adverbs). Modals appear in Infi. We can see this if we assume that negation attaches not to Infi and inversion moves Infi and adjoins it to IP. Modals are the only Verbs which must appear to the left of sentential not and must invert: (58)

a.

John must not leave.

86

The Representation of Implicit and Dethematized

Subjects

b. *John does not must leave. (59)

a. Must John leave? b. *Does John must leave?

Also, modals have the same distribution as Tense, which is typically assumed to appear in Infi. Modals cannot appear in infinitival clauses: (60)

a. b. c. d. e.

*To must leave now is a shame. *John tried to must leave. *I have the book to must read. *I bought the book to must read it. *John is easy to must see.

Clearly then, root modals lack the mobility of Adverbs. Nevertheless, like Class I Adverbs, root modals can modify either the subject or the Agent: (61)

a. b.

Students may be taught Marxism in this seminary, Visitors must be served dinner at 8.

Permission and obligation, respectively, can be understood as predicated either of the subject or of IMP here. Obviously, the modal is in no sense a sister of the empty Z>j>-phrase in these examples. A possible explanation for the IMP-modification readings of (61) is that either DS or SS is the level at which the selection properties of root modals are interpreted (this makes modals more "Verb-like" than Adverbs in their properties in that satisfaction of lexical requirements at DS is a general property of Verbs). This would account for the fact that the IMP in (61) is modified by must: IMP-modification takes place at DS; subject-modification at SS. Root readings are unavailable with middles, again like Class I Adverbs: (62)

* Arabic can read easily.

(63)

*Dinner must serve best at 8.

Given the proposal that modals can satisfy their selection requirements at DS, these examples show that middles have no Agent argument even at DS. The modals in (62) and (63) must be epistemic (meaning "possibility" and "necessity" respectively - see 4.4. on the relation between epistemic modals and middles). The two main claims we have made about Adverbs here are: (i) that they select a subject in the configuration of R-modification (17) and (ii) that different Adverbs have different lexical selection properties which,

The By-Phrase and Implicit

Arguments

87

in conjunction with (17), account for many aspects of their distribution. Note how in this account of Adverbs the surface distribution is guaranteed by lexical properties, by (17) and the Predication Principle in a similar way to how the surface distribution of argument NPs is guaranteed by subcategorization properties, the definition of government and the Projection Principle. The Class I subgroup is particularly relevant to the concerns of this book, as these Adverbs also modify an argument NP. This NP must appear in a configuration of Agent θ-role assignment by being coindexed either with Infi or V (or both). These claims together predict the interaction of the readings and distribution of Class I Adverbs. More importantly, they also interact with the properties of the öy-phrase/IMP so as to give an account of "implicit argument" phenomena with Adverbs. The behavior of Class I Adverbs in passives shows that passive Verbs assign an Agent θ-role. The last claim supports the analysis of the passive proposed in Chapter Two. 3.2.2. Adjectives Unlike Class I Adverbs, Adjectives do not appear to be predicable with IMPs. Contrast (64) and (65): (64) (65)

Theyi left the room sad;. *The room was left

IMPi

sadi.

On the other hand, (66) shows that Adjectives can be predicated of the surface subjects of passives (this is in fact marginally possible in (65)): (66)

a. b.

They hammered the metali flati. The metali was hammered ti flat;.

Let us assume that Adjectives, just like Adverbs, are predicates which select a subject and must be predicated of that subject following the Predication Principle, i.e. under R-modification. This requirement is met in (64) by the subject and in (66)-by the object. It is not met in (65). The purpose of this section is to specify the different ways in which the Predication Principle is or is not met by Adjectives (although we ignore premodifying Adjectives in NP throughout - cf. 4.5. for a brief consideration of these). Clearly a general account of this should explain the ungrammaticality of (65), and such an explanation will be independently motivated. The first point to note is that there is a clear difference in meaning

88

The Representation of Implicit and Dethematized

Subjects

between sad in (64) and flat in (66): sad denotes a state which persisted throughout the action described by the predicate, while flat denotes the result of the action described by the predicate. 8 In other words, in (64) the modifiee of sad can be in a state of sadness before the time of the action described by the predicate; in (66) the modifiee of flat is in a state of flatness precisely from the time of the action of the predicate. We call Adjectives like sad in (64) "temporally independent" (i.e. with respect to the temporal interpretation of the clause); those like flat in (66) "temporally dependent." We can deduce the following generalization from the examples in (64 -66):

(67)

In their temporally dependent readings, Adjectives modify the object; they modify the subject on their temporally independent reading.

To be consistent with our definition of predication via R-modification, then, temporally independent Adjectives should appear outside VP and temporally dependent ones inside VP. In that case, temporally independent Adjectives will only R-modify the subject and temporally dependent Adjectives will only R-modify complements (we put non-object complements to one side for the moment - see below). So we assign the following representations to (64) and (66a), respectively: (68)

[IP Theys [VP left the room] sadi].

(69)

[ip They [VP hammered the metali flat,]].

We can test for whether a given Adjective is in VP using the familiar tests for VP-constituency: (70)

a.

VP-Deletion John met Mary sad and Bill did too. ?John met Mary sad and Bill did happy. b. VP-Fronting John wanted to meet Mary happy - and meet her happy he did. *John wanted to meet Mary - and meet her he did happy. c. RNR John wanted to meet Mary happy - and Fred wanted to meet Sue sad - in order to win the bet. ??John wanted to meet Mary - and Fred wanted to see Sue - sad.

The By-Phrase and Implicit

Arguments

89

d.

Though-movement Meet Mary happy though John may ... *Meet Mary though John may happy ... e. Pseudoclefts What John did was meet Mary happy ... *What John did happy was meet Mary ... These examples (with the possible exception of the VP-Deletion case in (70a) which seems marginally possible on the temporally independent reading) indicate that the Adjective must be in VP, no matter which reading it has (cf. Andrews (1982). Binding theory gives a further reason to think that predicative Adjectives are always in VP. Consider the following examples with respect to Principles Β and C: (71)

a. *Mary met himi angry about John,. b. *Sue saw him; proud of Fredj. c. *WhOi did you consider ti angry at t,?

The names in (71a,b) cannot be coreferent with the pronouns. This is impossible whether the Adjectives are read as temporally dependent or independent. This is not due to Principle B, as the names here do not c-command the pronouns, since they are contained in AP. If the pronouns c-command the names, Principle C will rule out these examples; but for this to be so we must assume that the names, and therefore the APs which contain them, are VP-internal. On the same assumption, (71c) contains a variable which is locally Α-bound by another variable, so this example can be ruled out in the same way as (71a,b) (for more on parasitic gaps cf. 3.2.4.1.2.). The facts in (71) give a further indication that the APs we are concerned with here are always VP-internal. (This argument would receive compelling support if *Mary met hisj mother proud of Johni were grammatical. However, in 3.3.4. we will propose a constraint that independently rules out such examples; the constraint is a general one on coreference relations between (parts of) arguments and (parts of) adjuncts). We take the evidence in (70) and (71) - especially the latter - as conclusive, and regard predicative APs as VP internal in all the examples we have considered. In that case, we have to replace the representation of (64) given in (68) with one where the Adjective is inside VP and account for how the Adjective is predicated with the subject, given our theory of predication. We propose that, rather like Class I Adverbs (cf. 3.2.1.1.), predicative Adjectives can satisfy the Predication Principle by being coindexed with

90

The Representation of Implicit and Dethematized

Subjects

a Verb that they R-modify. So we propose (72) as the representation of (64), in place of (68): (72)

[IP They; [VP left; the room sad,]].

(72) is clearly consistent with the data in (70) and (71). We assume that the interpretation of predication with the subject is the result of the fact that V is coindexed with the subject for θ-role-assignment. This account really states that the Verb and Adjective form a complex predicate, predicated with the subject. Suppose that this is required of Adjectives which are not predicated with the object. Suppose further that government is a condition on complex-predicate formation. In that case, the Adjective must appear in VP in order to be governed by V in order to form a complex predicate with V; if it appeared outside VP, the Adjective would be unable either to predicate with the object or form complex predicate with V. This then explains the facts of (70) and (71). We also have to reformulate our account of the contrast between temporally-dependent and temporally-independent Adjectives in light of the data in (70) and (71) and the representation for (64) given in (72). To do this, we elaborate the notion of "temporal dependency." The intuitive idea is that just as certain arguments are referentially dependent on certain other arguments for their interpretation, so certain elements (usually predicates, but see below) may be dependent on other elements for a temporal interpretation. Temporal dependencies are expressed by means of coindexation in a local domain. It seems that the relevant locality restriction is one of local government: where Y is temporally dependent on Χ, X must govern Y and be coindexed with Y, and there must be no Ζ which governs Y and does not governs X and is coindexed with Y. Given our remarks about clitic chains in 2.2.4., this locality restriction is thus a local binding condition for a "chain" of binders headed by a head. The head in question is usually Infi. The relevant kind of R-relation for the binding relations is thus government. To account for the difference in temporal dependency between (64) and (66), for example, we can say that predicative Adjectives are optionally temporally dependent on V ( b u t cf. fn. 8). These Adjectives are interpreted as predicated of an object iff dependent and as predicated of the subject otherwise. In fact, rather than predication with an object, it appears to be predication with the Theme that leads to the temporally dependent reading, as shown by the following examples: (73)

a. *They gave the book to John* happyi. b. *We gave Johni a book happyi.

The By -Phrase and Implicit Arguments

91

c. *John chewed at the meat; raws. d. *It impressed me; happyi that.... e. *John baked Maryi a cake drunki. (The judgements apply to the temporally dependent reading only: (73e), at least, is grammatical on the independent reading, where the Adjective is construed with the subject). We proposed in 2.2.2. that Themes are structurally θ-marked. This means that they receive their θ-roles by coindexation under government by a head. Suppose that this coindexation also makes the Theme temporally dependent on the Verb (this captures the idea that the Theme is that argument which undergoes a change of state - cf. 4.3.). Then an Adjective coindexed with the Theme will also be temporally dependent on V. So Adjectives are either temporally dependent on V via coindexation with the Theme - as in (66) - or they form a complex predicate with the Verb, as in (64). A natural assumption is that the same coindexing relations "count for" both temporal dependencies and predication/structural θ-role assignment. In order to maintain such an assumption, we have to distinguish the subject-predicated and object-predicated Adjectives, even though both are coindexed with the Verb, and therefore, if the Verb is coindexed with Infi, with Infi. This suggests that coindexation is not the best mechanism for temporal dependencies; instead, we should use the "linking" notation of Higginbotham (1983). In terms of linking, the two kinds of predication differ as follows:

V

A

(Complex V-A predicate).

A

(Object predication).

In (73f) the Adjective is dependent on Tense, as is the Verb, but neither is dependent on the other. In (73g), on the other hand, the Adjective is dependent on the Verb, which is itself dependent on Infi. The linking notation captures this distinction in a way which cannot be achieved by coindexation. The coindexing for temporal dependence/structural θ-marking seems to be relevant for certain extraction contrasts. Temporally dependent Adjectives are extractable, but temporally independent ones are not: (74)

a. *How sadi did you leave the room t,? b. How flati did you hammer the metal ti?

Contrast these examples with short adjunct movement:

92 (75)

The Representation of Implicit and Dethematized a. b.

Subjects

Why did John leave the room? How did John hammer the metal?

These contrasts are accounted for by the Empty Category principle (ECP). The Empty Category Principle (ECP) requires non-pronominal empty categories to be properly governed. This requirement is satisfied by an empty category if it is a complement (lexically governed) or if it is antecedentgoverned. An empty category is antecedent-governed if there is no barrier between that category and the next link of its chain. Roughly speaking, a barrier is a maximal projection which is neither a complement nor IP (see Chomsky (1985) for more elaborated discussion and definitions - in particular, Chomsky assumes that VPs are barriers, whose barrierhood can be voided by means of adjunction. For the moment, we will put this assumption of Chomsky's to one side and assume that VP is a barrier for antecedent-government). In these terms, the traces of how and why are antecedent governed by their operators in (75), as no non-IP ηοη-θ-marked category intervenes between the operator and the trace (these adjuncts are base-generated outside VP). If we interpose more material, for example a complex NP or a clause whose Comp contains another wh, the sentences become bad: (76)

a. *Why did you believe the rumor that John left the room t? b. *How did you wonder whether John hammered the metal t?

On the other hand, the barrier VP intervenes between the trace and the wA-phrase in (74). Therefore the grammaticality of (74b) is what needs to be explained. We could simply propose that VPs containing Themes are not barriers. However, at least two facts suggest that this is not the right solution: first, long extraction of temporally-dependent Adjectives is much better than (76): (77)

?How flati did you wonder whether/believe the story that they had hammered the metal ti?

Second, if the Theme is extracted, the Adjective cannot be extracted: (78)

a. *How flati did you wonder which metalj they had extracted tj ti? b. ?Which metali did you wonder whether they had hammered ti flat?

(77) suggests that antecedent-government is not at issue. Long extraction here only produces a wA-island violation, a weaker effect than the ECP

The By-Phrase and Implicit

Arguments

93

violation of (76). Compare long extraction of complements with (76) and (77): (79)

?What did you wonder whether/believe the story that John stole t?

(77) is violation like (79) rather than like (76). This suggests that, with respect to long extraction, temporally-dependent Adjectives act like complements rather than like adjuncts. (78) shows that the Theme NP must be in situ for long extraction of the AP to be possible. Putting these two observations together, we arrive at the conclusion that traces of temporally-dependent APs are lexicallygoverned by the Theme NP. For this reason the barrierhood of VP is irrelevant in (74b) and the AP acts like a complement in (77). (78) indicates that lexical governors must be in situ for lexical government of complements to work, a matter which has been little discussed in the literature since the standard lexical governors are Verbs and wA-movement of Verbs is not attested in the world's languages (but cf. Koopmann (1985)). So where an AP-trace is governed by a Theme NP which it is temporallydependent on it is lexically governed by that Theme NP, and therefore can be extracted without violating the ECP. This conclusion suggests that coindexation for structural θ-marking/temporal dependence is closely related to lexical government, an idea which is consistent with the intuition that direct θ-marking is what underlies lexical government, and that the relation between this and antecedent-government is that both are forms of coindexation (cf. Stowell (1981), Chomsky (1985, Section 11)). The contrast in (74) suggests that - at least for AP - temporal dependence is relevant to the ECP. However, given our proposals about complexpredicate formation above, (74a) may be ruled out by the Lexical Integrity Condition, since the Adjective is a part of the same word as the Verb. So far we have proposed a distinction between temporally dependent and temporally-independent VP-internal APs. We still have not specified what underlies the impossibility of (65) and (73). Comparing Adjectives with Adverbs, the following generalization suggests itself: (80)

a. b.

Adjectives appear in VP and may be sensitive to Theme Θroles. Adverbs appear in I P / C P and may be sensitive to the Agent θ-role.

We could summarize (80) by saying that Adverbs are primarily Infi or Comp modifiers: these are the heads of which they select particular properties, and if they select a θ-role it will thus be a VP-external θ-role,

94

The Representation of Implicit and Dethematized

Subjects

the Agent. Adjectives on the other hand are primarily V-modifiers, and so they select properties of V and the canonical VP-internal θ-role, the Theme (this idea about Adjectives could be extended to NP-internal Adjectives in a fairly obvious way). We should add two statements to (80): (81)

a. b.

APs (either Adjective or Adverb) are only associated with structurally-O-marked NPs. The selection/predication requirements of APs are met at SS/ LF, not at DS.

(81a) derives the impossibility of the readings in (73). (81b) allows us to explain the ungrammaticality of (65): Adjectives can only be predicated with surface subjects (via predicate-formation with V), or with Themes via temporal binding. So these proposals will explain the contrast found with IMPs, illustrated in (82): (82)

a. The book was sold voluntarily, b. *The book was sold angry.

The same account carries over to 6j>-phrases. This seems to be correct, as the following examples show (although Zubizarreta (1985) claims that ^-phrases can be construed with predicative Adjectives): (83)

a. b. c. d.

*John was seen by Mary; drunk;. *Fred was kissed by Suei happy;. *Tom was met by Bill, angry;. *Sue was arrested by the police; drunki.

Two more phenomena should be pointed out in order to complete our account of predicative Adjectives. First, there is class of selected Adjectives. These are illustrated in (84): (84)

a. b. c.

They saw the child; sad;. John made Mary; scared;. We consider this boy; smart;.

The matrix Verbs here select Adjectival complements. The CFCC (Chapter Two (106)) therefore requires that the Adjective be predicated of its subject inside a maximal projection θ-marked by the Verb. This forces the Adjective to project a VP-internal small-clause constituent. Temporal dependence in the small-clause constructions is optional, it seems to be triggered as a lexical property of the matrix Verb; for example,

The By-Phrase and Implicit

Arguments

95

make requires temporal dependency in (84b), while consider does not allow it in (84c). This is unrelated to the possibility of extraction, which seems to be possible in all cases: (85)

a. b. c.

How sad did they see the child t? How scared did John make Mary t? How smart do you consider this boy t?

This is expected; the AP is lexically governed as a complement of the matrix Verb. 9 Second, Adjectives can be predicated of objects without being temporally dependent on them. The following examples are cases of temporally independent Adjectives predicated of objects: (86)

a. b. c.

We left the committee! unhappy;. John met Fredi drunki. John chewed the meat, rawi.

In all these examples, the property named by the Adjective can hold of the NP the Adjective is predicated with at a time prior to (i.e. independent of) the time specified by the tense of the Verb. Note that extraction is not possible in these cases: (87)

a. ??How unhappy did you leave the committee t? b. *How drunk did you meet Fred t? c. *How raw did John chew the meat t?

If raw is understood dependently in (87c), the extraction becomes good (even though this reading is anomalous, as it is impossible to make meat raw by chewing it). (87a) may be marginally possible to the extent that leave can have a AP small-clause complement. These examples show two things: (i) temporal dependencies are relevant for extraction possibilities; (ii) we can retain the link between Themehood and temporal dependency by assuming that the objects here are not Themes. The latter move is made at the cost of qualifying the generalization in (81a). For the moment, we will make this move and regard the instances of predication with nonTheme objects in (86) as exceptions to the general tendency, which is correctly summed up in (81a). So we have proposed three kinds of structures for predicative Adjectives: (88)

a. b. c.

NP NP NP

V, (NP) APi] ( = (64)) [VP Vi Ν Pi APi ] ( = (66)) [VP [X NPi APi ]] ( = (83)) [VP

96

The Representation of Implicit and Dethematized

Subjects

These are all in fact distinct in various ways. First, (86b) requires a temporally dependent reading of AP. Second, only (86a) has a nonextractable AP. Third, only (86c) features a selected AP. 10 In this section we have provided a general account of predicative Adjectives which distinguishes them from Adverbs in fairly natural way, explains why IMPs are not predicated of them and is independently motivated. We have also introduced a notion of temporal dependency, and related it to predication, structural θ-marking and possibly lexical government. The notion of temporal dependency will be central in our account of middles in Chapter Four. 3.2.3 Implicit Arguments with PP In this section we briefly discuss another class of cases which may be relevant to determining the properties of IMPs: the "subject-oriented PPs" of Keenan (1979) (see also Keenan & Faltz (1978)). In this section we will briefly consider the facts and attempt to see how the generalizations we arrived at in the previous section for predicative Adjectives apply to the relevant PPs. We will see that the generalizations fail to yield the correct predictions, so we conclude, with supporting evidence from the ECP, that these PPs are optional arguments of V, and the putative implicitargument effects are the consequence of compositional θ-role assignment to the subject, combined with inherent properties of Themes. The basic contrast is between "causatives of verbs of motion" and "action-at-a-distance" Verbs, as Keenan calls them. If a locative PP appears with a member of the former class, the object of the Verb is understood to be in the location specified by the PP. Hence (89a) entails (89b): (89)

a. b.

John dragged/led/tugged/brought Bill from the attic, Bill was in the attic.

With "action-at-a-distance" Verbs, on the other hand, the subject of the Verb is understood to be in the location specified by the PP. So we have an entailment between (90a) and (90b): (90)

a. b.

John saw/attacked/signalled Bill from the attic, John is in the attic.

We could assume, in line with our proposals for Adjectives in the previous section, that PP optionally coindexes with the Theme, to give the reading in (89), or with V and hence the subject, to give the reading in (90). Two predictions follow if PP is coindexed with the Theme: (i) PP is interpreted as temporally dependent on V in (89) (but in (90) it is not), (ii) PP is lexically governed in (89) (but in (90) it is not).

The By-Phrase and Implicit Arguments

97

The first of these predictions appears to be correct. Note that we expressed the entailments (89b) and (90b) in different tenses. This is intended to capture the fact that (89) is understood to mean that Bill was in the attic up to the time specified by the tense of the Verb, but not afterwards, while (90) is understood to mean that John's presence in the attic lasted for a period independent of the time specified by the Verb. Hence these PPs seem to be temporally dependent or independent, in a way similar to the Adjectives of the previous section (there is a difference, however: with temporally dependent Adjectives the property described by the Adjective lasted from the time specified by the Verb, while with temporally dependent PPs the location specified by Ρ lasts up to the time specified by the Verb). Coindexing PP and the Theme in (89) but not in (90) will account for this difference. Also, it accounts for the different meaning of from in the two examples. In (89) from has a "motion" reading, while in (90) it does not. Since Prepositions specify relations, in the simplest cases - like those under discussion - between physical locations, from, when part of a dependent PP, specifies a relation of motion between locations, and the motion is understood to take place at the time specified by the Verb. When part of an independent PP, on the other hand, from specifies a static relation between two locations. We will turn the second prediction below. The above suggests that PPs are just like APs in their predication possibilities. This is not the case, however, as the passive of (90) shows: (91)

Bill was seen/attacked/signalled from the attic.

This example no longer entails that the subject is in the attic. Instead, it entails that the referent of the IMP is in the attic. The same is true with overt 6^-phrases. Compare: (92)

Bill was seen/attacked/signalled by John from the attic.

Here, clearly, John, not Bill, is in the attic. We could account for this by allowing PPs to be interpreted as predicated of the IMP/6y-phrase, in addition to the Theme or V. Note that the PP in (91) and (92) is not interpreted as dependent. In other words, from specifies a static relation between locations, not a movement relation. The obvious way to allow for this is to say that PPs differ from APs in that their predication properties can be satisfied at DS. Consider next extraction. Extraction of the PPs in (89), (90) and (91) is possible, regardless of temporal dependency:

98 (93)

The Representation of Implicit and Dethematized a. b. c.

Subjects

Which attic did John drag/lead/tug/bring Bill from? Which attic did John see/attack/signal Bill from? Which attic was Bill signalled (by John) from?

(Note that the choice of stranding or pied-piping the Preposition makes no difference here. We will ignore this choice in what follows). Long extraction produces only the relatively mild subjacency effects: (94)

a. ?Which attic did you wonder who John dragged from? b. ?Which attic did you believe the story that John saw Bill from? c. ?Which attic did you forget whether Bill was signalled from?

To assimilate these PPs to the APs discussed in the previous section, we would have to claim that the Theme NP is what lexically governs the PP in (93a), and that this PP is not lexically governed in (93b,c). The grammaticality of the latter examples suggests that this is the wrong approach, and this suggestion is borne out by the long-extraction facts in (94). Notice in particular that in (94a) the putative lexical governor of the PP is also extracted. In the last section, we gave evidence that this leads to ECP violations, as lexical governors must be in situ. In (94a), however, the violation is no stronger than a typical wA-island violation. So extraction facts suggest that the approach sketched up to now, which assimilates these PPs to APs by treating them as adjuncts, is wrong. An alternative suggests itself immediately. We just noted that (91) and (92) are evidence that the predication properties of PPs can be satisfied at DS. Compare also the passive of (89): (95)

Bill was dragged/led/tugged/brought from the attic.

Here too the Theme of the Verb is understood to be in the attic. The constancy of these relations suggests that the entailments in question are a DS matter, related to θ-role assignment. The alternative to the above account is to treat the PPs as arguments of V, and consider the entailments as a function of lexical properties of V: Verbs of certain classes require their Theme to be understood as "predicated o f ' a locative argument, while Verbs of other classes require their Agent to be. So the differing properties of the PPs we have seen are a function of properties of the Verb that selects them, not of the PP itself. There are several advantages to treating these PPs as arguments of V, aside from the obvious one that it allows us to retain the theory of adjunct-predication proposed in the preceding sections. First, if the PP is an argument of V, it is lexically governed by V. This accounts for the extraction data in (93) and (94). Second, since this account treats the "predication" properties of PPs as

The By -Phrase and Implicit

Arguments

99

induced by the Verb, we can say that PPs are always the CFC(P) Prepositions do not have subjects - so there is no small clause in any of the examples we have considered. The relative acceptability of (94a) is evidence against a small clause analysis of these examples, given our assumptions about lexical government by predicates in situ. Third, since the predication properties of PPs are lexically determined requirements, they are not affected by syntactic operations like passive. This is shown by (91), (92) and (95). So we adopt the idea that these PPs are optionally subcategorized arguments of V. There is still further evidence for this view if we consider a wider class of Verbs, and further extraction behavior of the PPs. After reviewing this evidence, we will return to the only matter left open by taking these PPs to be arguments of V: the determination of the temporal readings of PP as dependent or independent. Consider next, then, the class of "state locatives." The paradigm we are interested in is given in (96): (96)

a. b. c.

John saw/found/killed Bill at the meeting. John criticized/cursed/mentioned Bill at the meeting. Bill was criticized at the meeting.

(96a) entails that Bill was at the meeting, (96b) that John was at the meeting, and (96c) that the referent of the IMP was at the meeting. These entailments parallel (89), (90) and (91). However, note that there is no difference in the reading of the PP; in all three examples, at maintains a static meaning (which it does not have inherently, cf. John threw the book at Bill). So the correlation noted in the earlier examples between predication with the Theme and a change-of-state reading for the Preposition breaks down here. This casts further doubt on the idea that these PPs are adjuncts with predication properties like those of APs. Again, the simpler proposal seems to be that the PPs are subcategorized by V and the semantic relations between PP and the other arguments of V are lexically specified. Turning now to the question of extraction, we noted above that the predicted extraction contrast between dependent and independent PPs fails to hold. In fact, all these PPs act consistently as though lexically governed. Further evidence for this comes from the fact that we can construct a "McCloskey-Chung paradigm" (cf. McCloskey & Chung (1983)) with these PPs. Long extraction of PP from a wA-island is best if the subject only undergoes short movement, more marginal if there is short movement of an object, and very bad if there is long movement of the subject. These facts are illustrated in (97): (97)

a. ?This is the attic from which you wondered which people were attacked.

100

The Representation of Implicit and Dethematized

Subjects

b. ?*This is the meeting at which you wondered which people Bill was criticized in front of. c. *This is the attic from which you wondered which people the police said were attacked. (We have chosen the pied-piping option here in order to preserve parallelism.) If the PPs were not arguments here, the empty categories left behind by extraction would have to be antecedent-governed. In all the examples in (97), antecedent-government is impossible, so they should be equally ruled out by the ECP. The fact that they are not, and that we find the kind of gradation typical of subjacency violations with lexicallygoverned complements, indicates that the PPs are arguments. We take1 these facts - in conjunction with those noted in (93) and (94) - as strong evidence for the argument status of subject- and object-oriented PPs. Note that the extraction properties of the APs discussed in the previous section are quite different (cf. in particular fns. 7 - 10). So we conclude that both subject- and object-oriented PPs are optional arguments of the Verbs they cooccur with. The "predication" properties of these PPs result from idiosyncratic properties of the Verbs of which they are arguments. Note that these properties fall into subregularities determined by the semantic class of the Verb, something we expect of lexical properties. The final issue regarding these PPs concerns the temporally dependent readings we observed in (89). How should we account for these readings in terms of our general lexical account of the PPs? One possibility would be to state the generalization that dependent readings of PPs only arise where a Theme argument is present. In other words, if V requires the optional PP argument to be interpreted as if predicated of the Theme, the PP has a dependent reading. Otherwise it does not have a dependent reading. This is consistent with the examples above. Further evidence for this comes from Verbs whose only NP argument is a Theme. Cf.: (98)

a. b.

John fell/leapt from the attic. John came/ran/travelled from London.

In these examples the PP clearly has a dependent reading, in the sense outlined above. Compare now (99), where the one NP argument of V is in every case an Agent: (99)

John spoke/sang/shouted/whistled from the roof.

Here PP is independent. So the correlation between presence of a Theme argument and dependent reading of PP holds up. For the moment we

The By-Phrase and Implicit

Arguments

101

will leave it as an observation, although it is natural to try to conflate this observation with the point made in the previous section that adjunct APs predicated with the Theme are temporally dependent and attempt to make some generalization regarding Themes and the notion changeof-state. Since we pointed out in 2.2.2. that Agents entail eventhood this suggests the further generalization that structural θ-role assignment is intimately connected to the tense/aspect of the clause. We will develop this point further in Chapter Four. 3.2.4. Passives and Infinitival

Adjuncts

In this section, we deal with the relation between infinitival adjuncts and passives. We treat infinitival adjuncts in the same way as the APs of 3.2.1. and 3.2.2., namely as predicates with particular selectional properties which restrict the way in which they satisfy the Predication Principle. The selection restrictions must be met by an argument R-modified and coindexed with the adjunct. In fact, infinitival adjuncts resemble Adverbs more closely than they do Adjectives, in that (i) their domain of predication is C P / IP, rather than VP or NP (cf. 3.2.2.(80)), and (ii) they fall into selectional classes which closely parallel those observed by Jackendoff for Adverbs (cf. 3.2.1. (11)). The differences between infinitival adjuncts and Adverbs can mostly be traced to the fact that the infinitival adjuncts have more internal structure than Adverbs. In particular, infinitival adjuncts contain PRO, an element which imposes its own well-formedness conditions on the structure in which it appears. 3.2.4.1. Rationale Clauses We first discuss the type of infinitival adjunct which is most relevant to our theory of the passive; Rationale Clauses. We refer to Rationale Clauses as "RatC" throughout, pace Jones (1985). RatC are the infinitival adjuncts which correspond to Class I Adverbs. 3.2.4.1.1. Distribution of Rationale Clauses As we saw earlier, IMPs seem to be able to control the PRO of a RatC. As middles lack IMPs, we have the contrast of (5), repeated here as (100): (100)

a. The book was sold [PRO to make money] b. *The book sold [PRO to make money]

We propose exactly the same account as for the parallel passive/middle contrast with Class I Adverbs (cf. (4)): RatC select for an agentive, eventive clause. Passives are agentive and eventive, and middles are not. Hence

102

The Representation of Implicit and Dethematized

Subjects

middles cannot appear with RatC. What imposes the selectional requirement? To answer this question, we assign the structure [P CP] to RatC, where Ρ can be either in order or 0 . This Preposition imposes the selectional restriction in question. We will spell out the relation between Ρ and PRO in detail below. Our general proposal, then, is that RatC resemble Class I Adverbs in their external syntax, i.e. in their relationship to the rest of the clause. So we extend our proposals for predication to infinitival adjuncts. We will see in 3.2.4.1.4. and 3.2.4.2. that other kinds of infinitival adjuncts share selectional properties with other classes of Adverbs. We will also propose below that predication is relevant for the "internal syntax" of infinitival adjuncts, in licensing and accounting for the reference of PRO. RatC resemble Class I Adverbs in two notable ways: selection and transportability (in the sense of Keyser (1968)). The contrast in (100) indicates that RatC, like Class I Adverbs, select an Agent and an event. This is confirmed in (101): (101)

a. *The crystal dissolved [PRO to annoy the physicist], b. *John knew the answer [PRO to impress everyone].

The ungrammaticality of these examples is due to the selectional requirement of the (empty) preposition heading the RatC not being met. A comparison with 3.2.1.(10) shows that this selectional requirement is identical to that of Class I Adverbs: structural 0-role assignment by VP is what is required. We saw in 2.2.2. that structural θ-role assignment by VP leads to an agentive, eventive clause. RatC coindex with Infi and the subject, or with V and the IMP, in configurations of structural 0role assignment. They therefore require a θ-marked argument. This explains the ungrammaticality of (102): (102) T h e r e seemed to be a riot PRO to scare everyone. Also, the argument coindexed with the RatC must denote an entity capable of being agentive. This explains the ungrammaticality of (103): (103) *The key opened the door PRO to get inside. It is not enough to rule out (100b) in this way, as shown by (101b) and the following: (104)

a. *The slavei sold [PROi to get the money], b. *The sailori sank [PROi to save the world].

The By-Phrase and Implicit Arguments

103

For these cases, it is the fact that middles are non-eventive (see Chapter Four) which prevents predication with the RatC. "Transportability" refers to the property of appearing in any "peripheral" position in the clause: initial, between the subject and Infi, between Infi and VP, and adjoined to the left or right Infi. We saw in 3.2.1. how these positional possibilities for Adverbs interact with predication in ways predicted by the definition of R-modification given in (17). The same is true for RatC. RatC can appear in any of the positions just mentioned. In (100a) we saw a RatC in final position; (105) illustrates the other positions: (105)

a. To make money, John sold the books. b. John, to make money, sold the books. c. ?John has, to make money, sold the books. d. ??The books have been, to make money, sold by John.

(Also, RatC and Class I Adverbs can appear in topic position - cf. fn. 7). (105c,d) are somewhat awkward, but compare the following examples, which involve purpose clauses (PCs - see 3.2.4.1.3.; the examples are from Jones (1985:42)): (106)

a. *They will, to talk to, bring John along. b. *They will, to show them the way, bring John along.

Further, as Jones notes, RatC are able to act as members of VP with respect to VP-constituency tests: (107) I

a. b.

John dove from the cliff [(in order) to impress Mary] and then Fred did, too, [(in order) to get away from her], John dove from the cliff [(in order) to impress Mary], and then Fred did, too. (Jones' (67), p. 29).

II

a. ?John said that he'd jump off the cliff in order to impress Mary, and jump off the cliff in order to impress Mary he did. b. ?John said he'd jump off the cliff in order to impress Mary, and jump off the cliff he did in order to impress Mary. (Jones' (70), p. 30).

Ill

a. b.

Though John may dive off the cliff to impress Mary he'll still never be as handsome as Bob. Dive off the cliff to impress Mary though John may, he'll still never be as handsome as Bob.

104

IV

The Representation of Implicit and Dethematized

Subjects

c.

Dive off the cliff though John may in order to impress Mary, he'll still never be as handsome as Bob. (Jones' (73), p. 31).

a. b.

What John did was dive off the cliff in order to impress Mary, What John did in order to impress Mary was dive off the cliff. (Jones' (76), p. 31).

In all these respects - selection, transportability and optional domination by VP (as opposed to obligatory domination, which Jones shows to be the case for PC) - RatC parallel Class I Adverbs. So we propose the following representations for active and passive clauses containing RatC: (108)

a. b.

Johni Infi kissedi Mary [PPI 0I [CP PROi to impress Bill] Maryj was kiss+eni tj IMPi [pp 0, [CP PROi to impress Bill].

The RatC coindexes with either Infi and the subject (108a) or V and the IMP (108b). The coindexation is a consequence of the selection property of P, hence Ρ has the index and percolates it to PP. RatC R-modifies the subject in (108a) and the IMP in (108b). All of this is the same as with Class I Adverbs. However, it is clear that more is going on in RatC than with Adverbs. We have to specify how PRO in these clauses gets an index, i.e. how it is controlled. This is the topic of the next section. 3.2.4.1.2. Rationale Clauses and Control The idea that the IMP is related to PRO in (108a) by a rule of control in the accepted sense (e.g. as in Chomsky (1980)) is challenged by Jaeggli (forthcoming,a). Jaeggli points out a number of differences between the IMP-PRO relation and the relation between Bill and PRO in examples like (109) (although Jaeggli does not assume the existence of a structurally present IMP): (109)

a. b. c.

Bill tried PRO to leave. They persuaded Bill PRO to leave. Bill promised them PRO to leave.

The IMP-PRO relation in (108b) is called "thematic control" (0-control) by Jaeggli, and the relation Bill-PRO in (109) "argument" (A-) control. There are three differences between θ-control and A-control:

The By -Phrase and Implicit

Arguments

105

(110)

Α-control requires c-command; Θ-control doesn't a. *They were promised [by Billi] PROi to win. b. The pricej was increased by the government! [PROj/.j to suit the rich].

(111)

Α-control allows long-distance control; Q-control doesn't a. Theyi thought I had suggested that [[PRO, feeding each othen] would help]. b. John, was told I M P / b y his friendsj that PROj/.j to clean the house in order P R O ^ to impress the guests was foolish.

(112)

Α-control goes into passive complements; ^-control doesn't a. Johni wants PROi to be loved by Mary. b. *The gifts were brought IMPi/by Maryi [PROi to be loved by the Indians],

What underlies the two different kinds of control? In Chapter One, fns. 3 and 5, we sketched a theory of PRO whose main idea is that PRO must, because of the Visibility Condition, form an extended chain with a Case-marked element. Here we will pursue this approach a little further. Suppose that Α-control is the case of extended-chain formation where the argument with which PRO forms an extended chain is specified by the matrix Verb: promise and try require that PRO form a chain with their subject, persuade requires that the head of the extended chain be the object etc. (these remarks are intended to be sketchy as our purpose here is not to propose a detailed theory of Α-control). With θ-control, on the other hand, extended-chain formation is independent of the lexical properties of the matrix Verb. The purpose of this section is to show that the properties which distinguish θ-control and Α-control in (110 112) can be derived on our account of RatC as predicates with the external properties of Class I Adverbs. How is PRO assigned an index in (108)? Assume the following: (113)

PRO in RatC is coindexed with the Ρ head of the RatC.

Instead of receiving directly the index of some argument, (113) states that PRO receives the index of the head of the predicate it forms part of: Ρ the head of RatC = PP. Ρ must be coindexed with some argument following the Predication Principle, and more precisely with a particular type of argument it selects under R-modification, so PRO forms an extended chain with the argument predicated with the RatC. Since RatC must R-modify the argument they are predicated with, (113) is a way of limiting the possibilities of controlling PRO in RatC: the

106

The Representation of Implicit and Dethematized

Subjects

controller of PRO must be R-modified by the RatC. Selection restrictions also limit these possibilities in two ways: (i) because the head of the RatC imposes a selection requirement on the argument it is predicated of; (ii) because PRO must meet its own selection restrictions as subject of the RatC. Furthermore, even if both requirements are met, they cannot be contradictory or extended-chain formation with PRO will lead to anomaly. With these conditions in mind, we will now see how the properties of θ-control are derived. Consider first (110b). This property of θ-control is simply explained. The representation of this example is just like (108b) in all relevant respects: Ρ in the RatC is predicated with the agentive ^ - p h r a s e ; PRO receives the index of the ^ - p h r a s e - i.e. the index of the DS subject - via P. To account for the impossibility of Α-control in (110a), we need to assume a further condition on Α-control: that the controller must occupy a DS position θ-marked by the control Verb. This condition guarantees that IMPs/^-phrases cannot be controllers, as they do not occupy DSpositions. 11 Our account of (110b) entails that the RatC is inside VP. This idea can be tested in two ways: (i) by parasitic gaps, (ii) by VP-constituency tests. We will deal with parasitic gaps first. It has been argued (cf. Engdahl (1983), Chomsky (1982)) that a real gap must not c-command a parasitic gap. Chomsky (1982) argues that, since c-command of a parasitic gap by a real gap would produce an instance of a locally Α-bound variable, this anti-c-command requirement follows from Principle C of the Binding Theory. Hence the adjunct containing the parasitic gap must be VP-external. In general, parasitic gaps can appear inside RatC: (114)

Which painting did John steal t [in order to forge t]?

However, we predict that parasitic gaps should be impossible in (110b), since the RatC must be in VP to R-modify its argument. The prediction is tested in (115): (115)

Which book was John given t to impress the author of t ?

(115) is as good as any parasitic-gap sentence. Following Chomsky's account of the anti-c-command requirement above, the RatC in (115) must be VPexternal. This conclusion contradicts our proposal for (110b). However, we can resolve this contradiction following a proposal made by Mamoru Saito (class lectures, Fall 1984). Saito proposes that adjuncts containing parasitic gaps are base generated inside VP and move rightward to adjoin to VP during the derivation. This would give (116) as the SStructure of (115):

The By-Phrase and Implicit (116)

Arguments

107

Which booki was Johnj given tj ti tk IMP k [PPk Ok [CP PRO k to impress the author of ti]

Since the two ti are not in a c-command relation in (116), Principle C is not violated. The trace of the RatC tk modifies IMP k allowing PRO to be coindexed with IMP k . Hence, Saito's proposal explains why both parasitic gaps and VP-internal modification are possible with RatC. The independent motivation for Saito's account comes from several sources. First, Contreras (1984) noticed that if we said that the position of a real gap could c-command a parasitic gap, then the ungrammaticality of (117) would be straightforwardly explained: (117) *John filed themi without reading Mary's articles;. Second, both Engdahl (1983) and Chomsky (1985) point out that parasitic gaps are marginally possible in complements: (118)

Which candidates; did you convince t, that you were going to support ti.

Finally, (119) can be accounted for as an instance of strong crossover if themt c-commands tr. (119) *The candidatesi that you convinced them; that you were going to support ti. All of these facts can be accounted for if we say that the phrase containing the parasitic gap optionally extraposes. This extraposition is rendered obligatory in regular parasitic-gap cases and in these examples by the need to avoid a violation of Principle C. This proposal leads to a problem with examples like (120), where coindexation of him\ and John\ should be possible but isn't: (120) *You convinced him; that you were going to support John;. However, Saito suggests that these examples involve reconstruction of the extraposed structures. Reconstruction leads to a violation of Principle C for the name in (120). This analysis is supported by an effect noted by Reinhart (1976) and van Riemsdijk & Williams (1981): reconstruction avoids the effects of Principle C if the name is deeply enough embedded. For the case at hand, this is shown in (121): (121)

You convinced himi that you were going to vote for the man who Johni supported.

108

The Representation of Implicit and Dethematized

Subjects

So these facts actually support the extraposition analysis of parasitic gaps. They also lead to the conclusion that there are two variants of Principle C: one relevant for names, the other for variables. The effects of the variant relevant for names are attenuated if the name is sufficiently distant from its putative antecedent. Chomsky (1986) in fact proposes the following formulation of Principle C: (122)

a. b.

A name must be Α-free in the domain of its operator, A name must be A-free.

The two parts of this principle apply disjunctively, i.e. (122b) applies if (122a) cannot. Clearly (122a) is relevant to variables, given the Bijection Principle. In the light of the reconstruction data above, some qualification to (122b) seems necessary, although we will leave its exact nature open here. We return to the question of the correct formulation of Principle C in 3.3. A final piece of evidence in favor of the extraposition analysis of parasitic gaps is provided by Andy Barss. Barss notes that Adverbs of certain types (mainly Class IV) cannot be understood as predicated with the matrix Verb, if they appear following an adjunct containing a parasitic gap. Compare the following: (123)

a. *Which book did you file without reading quickly? b. Which book did you file quickly without reading? c. You filed this book without reading it quickly.

These examples also show that Adverbial predication takes place at a level before reconstruction, presumably, as we have been assuming all along, SS. So we adopt Saito's extraposition proposal for parasitic gaps. This allows us to maintain our account of (110b). Such an extraposition and.reconstruction analysis of parasitic gaps can carry over to VP-movement and deletion. The prediction of our analysis of RatC is that RatC should behave as VP-internal when modifying an IMP, as they must appear VP internally in order to R-modify the IMP. Consider the data in (124 - 6): (124)

a. ?The books were sold to Bill to make money and the records were to help him get hip. b. The books were sold to Bill to make money and the records were, too.

(125)

a. ?John wanted the books to be sold to make money and sold they were to make money.

The By-Phrase and Implicit b.

(126)

Arguments

109

John wanted the books to be sold to make money and sold to make money they were.

a. ?Sold though the books may be to make money ... b. Sold to make money though the books may be ...

It appears from these examples that there is preference for RatC with passive matrix clauses to be analyzed as part of VP by these rules. However, the contrast between these and the examples with an active matrix VP in (103 - 107) is not as strong as predicted. This can be attributed to the fact that the RatC is extraposed before VP-movement or deletion takes place. How is PRO in the RatC able to be controlled by the IMP in the extraposed structure? This question is especially important as we took the ungrammaticality of (123a) to show that Adverbial predication takes place before reconstruction, so we cannot invoke reconstruction to account for this. Instead, we propose that traces of RatC can be predicated with the argument selected by the head of the RatC, and therefore that control of PRO in the RatC takes place as normal. This proposal for RatC does not affect the status of (123a), as this example is ruled out because of the impossibility of the Adverb predicating with V under R-modification - no proposal involving traces can save this example. In fact, we have already invoked this possibility for both Adverbs and RatC for whmovement (see fn. 7). So we can account for all the facts discussed so far. Consider next (111). It is clear that (113) will give us an immediate account of the impossibility of long-distance θ-control. In (111b), neither PRO can be coindexed with IMP because both are inside sentential subjects. The RatC therefore does not R-modify IMP. Compare (111b) with another example of Jaeggli's: (127)

Johni was told IMPj [PROi to clean the house] [in order PRO^ to impress the guests].

As indicated, the first PRO is obligatorily (A-) controlled by John. The second PRO can be coindexed either with the first PRO or IMP. These two possibilities reflect the following possible structures for (127):

110 (128)

The Representation of Implicit and Dethematized a.

Subjects

ip

RatC

b.

John,

was

tel

In (128a), the RatC is VP-external and so R-modifies only John, so PROi is controlled by John. In (128b), the RatC can modify IMP so PRO can be coindexed with IMP. Alternatively, in (128b) PRO can be controlled by John because it can modify the trace of John inside VP. These results are consistent with our proposals for the predication properties of RatC. So long-distance control of PRO in RatC is impossible because PRO can only be controlled by the argument with which RatC is predicated, and predication takes place under R-modification. It is unclear why A-control is possible here, as in (128a).12 Consider next (112). We will enlarge the data base to consider the possible combinations of passive, middle and active clauses. (The possibilities of Α-control in these examples depend on a number of factors - cf. fn. 11 on passives and 4.2. on infinitival middles). Our first set of examples has active main clauses: (129)

a. Theyi sold the books [PROi to make money]. b. Theyi wrote the books [PROi to be given tenure]. c. *They sold the books to read easily.

The By -Phrase and Implicit

Arguments

111

(129a) is, of course, unproblematic. (129b) is also unproblematic: they is the Agent, and the RatC is predicated with and PRO is controlled by they. (129c) is ungrammatical because they is predicated with the RatC, but coindexing they and PRO produces selectional incompatibility; things that can be Agent of sell books cannot be Theme of read normally. Note that if we force a purely intransitive, nonmiddle interpretation of read, PRO will be the Agent of read and then the selectional incompatibility disappears: (130)

They bought the books [PRO to read better],

(i.e. to practice their reading skills). This shows that (129c) is not ruled out on structural grounds. The next triple has middle main clauses: (131)

a. *The books sold [PRO to make money]. b. *The books sold [PRO to be read]. c. *The books sold [PRO to read easily].

All these examples are ruled out because the selection requirement of the RatC is not met. As we suggested earlier, middle predicates are non-agentive and non-eventive (cf. 3.2.1.). The third set of examples has passive main clauses: (132)

a. The books were sold IMPj [PROj to make money]. b. ??The books were sold IMPj [PROj to be read]. c. *The books were sold IMP [PRO to read easily],

(132a) is equivalent to (101b); here, the RatC is predicated with IMP and so PRO is coindexed with IMP. (132b) features the same kind of selectional incompatibility that we saw in (129c): things which can be Agents of sell cannot usually be Theme of read (except in the figurative sense of "read an author," meaning read the author's work - in that sense, (132b) is OK; (129c) does not allow this reading as "author" is not the Theme argument of read and surface subjects of middles must be Themes - cf. Chapter Four). Note that this example has another reading; it can be understood to mean "the books were sold by someone for someone (else) to read them." Presumably, the books coindexes with PRO to produce this reading. This coindexation poses a problem for (113), as (113) requires PRO to receive the index of the modifiee of the RatC (via the Prepositional head of the RatC). However, on this reading the example is arguably a subject purpose clause (SPC, see Jones (1985) and 3.2.4.1.3.). The most important evidence that this

112

The Representation of Implicit and Dethematized

Subjects

is an SPC comes from transportability. SPCs are in general not transportable, while RatCs are. If we move the adjunct in (132b) to other peripheral positions in the clause, we find that only the RatC reading (the one with selectional incompatibility where IMP controls PRO - i.e. the reading where the author sells his books so that he, the author, can be read) is available. Cf.: (133)

a. ?*To be read, the books were sold. b. *The books, to be read, were sold. c. *The books were, to be read, sold.

So we conclude that (132b) allows an SPC reading which is irrelevant to our concerns here. In the next section we will discuss the similarities and differences between RatCs and other kinds of purpose clauses in detail. (132c) is ruled out in the same way as (129c). The RatC predicates with IMP, but PRO cannot be coindexed with IMP due to selectional incompatibility. We have now derived all the properties of θ-control from properties of RatC. With the exception of (113) we have added nothing to what was said about Adverbial modification. (113) is really part of the theory of control and is inevitable in that we have to say something about assigning an index to PRO in RatC. Moreover, as part of a potential theory of control, (113) is independent in principle of implicit arguments. Two more points remain to be made in this section concerning RatC. First, RatC allow overt subjects, with for in Comp as a Case-assigner (cf. Jones (1985) for arguments that this for phrase is the subject of the RatC and not a matrix benefactive indirect object). So we also need to explain the coreference possibilities in examples like (134): (134)

The boat was sunk by the Navy [for the President to get reelected].

Unlike PRO, overt NPs do not need an antecedent, so coindexing can be left entirely optional, subject to Principle C. Hence the RatC can be predicated of an argument distinct from its subject, as shown in (134). On the other hand, the overt subject of a RatC must be disjoint from the IMP predicated with the RatC, as shown in (135): (135) *The boat was sunk IMP, for the CIA, to collect the insurance. The argument to which purpose is ascribed must be distinct from the CIA here. This follows from Principle C as given in (122) above, given that IMP c-commands everything in the RatC. Note that the same effect of attentuation of the violation is achieved if the name is more deeply

The By -Phrase and Implicit

Arguments

113

embedded, and furthermore that a pronominal subject of the RatC can be coreferent with IMP: (136)

a. ?The boat was sunk IMPi [for [the people who worked for the CIAi] to collect the insurance], b. The boat was sunk IMPi for themi to collect the insurance.

So our account of RatC claims that RatC with overt subjects differ from RatC with PRO subjects in that those with PRO subjects must be understood so that their subject is also the argument of which purposiveness is predicated (except for in cases like (132b) where the books controls PRO), while RatC with overt subjects forbid this interpretation. However, there are cases where PRO appears not to need an antecedent. Cf. the following kinds of example: (137)

The boat was sunk by the torpedo [PRO to scare the enemy].

Lasnik (1985) proposes that what is going on here is "event control," i.e. the matrix Verb (or Infi) assigns PRO an index, giving the reading "the sinking of the boat scared the enemy." This is surely the correct reading for (137). In fact, Lasnik proposes that event control underlies all the cases of control by implicit arguments. There is, however, a range of evidence that indicates that event control (Ε-control) is a phenomenon distinct from the kind of control by implicit arguments (IMP-control) that we have been considering up to now in this section. Jaeggli (forthcoming^: fn. 20) gives arguments against Lasnik's proposal. He points out, first of all, that a crucial step in Lasnik's reasoning is that examples like (138) are bad: (138)

a. ?The ship was sunk by the navy by a torpedo, b. ??The navy sank the ship by a torpedo.

If the phrase by a torpedo here is the "double" of en, i.e. a passive fryphrase, then their ungrammaticality is explained and there is no structural Agent present to be predicated with the RatC in (137). The only possible controller in (137) is then the event of sinking the ship. However, as we have indicated, it is not clear that these examples are completely impossible; certainly, as Jaeggli points out, if by a torpedo is replaced with by torpedo they improve considerably. The only account for such acceptability is that these are not passive ό^-phrases, i.e. not the doubles of en, but rather instrumentals. The awkwardness of by in (138) could be due to a preference for with in such phrases. If this is so, then examples like (137) contain an IMP which is predicated with the RatC. Evidence for this, and against

114

The Representation of Implicit and Dethematized

Subjects

Ε-control, comes from the possibilities for anaphors to appear in the RatC: (139)

a.

The boat was sunk (by the torpedo) IMP; [PROi to aggrandize themselves;]. b. *The boat was sunk by the torpedo [PROi to avoid repeating itselfi].

This contrast shows that a controlling IMP can, via PRO, antecede an anaphor in the RatC while a "controlling event" cannot. A simple explanation is that there is no controlling event; such instances of E-control are really IMP control. Another argument Jaeggli gives against Lasnik involves non-agentive eventive matrix predicates: (140) *The house was struck by lightning [PRO to start a fire]. Although the main clause is eventive, this event cannot be understood as the controller of the PRO, or as predicated with the RatC. However, the event nominal corresponding to the main clause can be a subject, so there is no semantic reason for the ungrammaticality of (140): (141)

The lightning striking the house started a fire.

In a similar vein, consider non-agentive eventive predicates with animate Experiencer or Theme subjects; (142)

a. b.

John saw Mary leave [PRO to blackmail her later]. Fred noticed the painting [PRO to show off his knowledge].

Both of these examples have good readings where the main-clause predicate is understood agentively; on this reading, see is equivalent to watch and notice means "claim recognition." However, both also have anomalous readings, where see means "perceive visually" and notice means "become aware of." Both readings are clearly eventive (even if they concern mental events). The difference between them is a matter of agentivity: the former are agentive while the latter are not. An Ε-control theory of RatC cannot account for this contrast, while our IMP-control based theory can. These points conclude our account of RatC. Our account is based largely on the similarities between these and Class I Adverbs. The proposals made concerning those Adverbs in 3.2.1.2. carry over for the most part to RatC with the additional statement of (113) needed for PRO. Where an overt non-pronominal NP is subject of the RatC this NP must be disjoint from the NPs in the main clause, following Principle C. Finally, we argued

The By-Phrase and Implicit

Arguments

115

that Lasnik's proposal for Ε-control does not explain why non agentive eventive predicates cannot control into RatC, while a theory based on IMP-control does. (That Lasnik's proposal is incorrect for RatC does not, of course, mean that event control does not exist.) 3.2.4.1.3. Rationale Clauses and Purpose Clauses There is another class of purposive infinitival-adjunct clauses in English, which we have not considered up to now. This type of clause is illustrated in (143): (143)

I bought John a book to read.

(143) is in fact ambiguous between an infinitival-relative reading and a purposive reading. The respective structures for the VP in (143) are as in (144) (following Faraci (1974) and Jones (1985); although these authors give differing internal structures and categorial status to these adjuncts, their external positions differ exactly as shown): (144)

a. b.

Infinitival Relative: [vp buy John [Np a [ N booki [CP Oi [PRO to read ti]]]]]. PC: [vp buy John [NP a book]; [Cp Oi [PRO to read ti].

These two can easily be distinguished: infinitival relatives allow a wA-phrase to appear in the position occupied by the empty operator Oi (subject to a Case Visibility condition - cf. Levin (1983)) in (144a), while purposives do not allow this. Hence (145) can only be an infinitival relative: (145)

I bought a book from which to read.

Among the other differences between the two types of adjuncts noted by Jones are the following (the first two are taken from Faraci (1974)): (146)

a.

IRs do not allow a pronominal head *Hei to talk to t, is here. I brought himi to talk to ti. b. IRs cannot go inside tensed relatives *A pan [ that is stainless [ to fry omelettes ]]. I bought the pan [Reici John fries omelettes in ] [Pc to hammer nails with ti]. c. PCs do not allow aspectual have Bambi is a book to have read. *I bought Bambi to have read by tomorrow.

116

The Representation of Implicit and Dethematized

Subjects

We will not be concerned with IRs here, but only with PCs. We will show that PCs are predicates like RatCs, but of a different selectional type. They select Themes rather than Agents. The immediately apparent difference between PC and RatC is that the former but not the latter appear containing an obligatory gap (compare (143) with any of the preceding examples of infinitival adjuncts). In fact, Jones distinguishes two types of PC, according to the position of this gap. Subject PC (SPC) have an obligatory gap in subject position, and Object PC (OPC) one in object position. The two types of PC are illustrated in (147): (147)

SPC: They brought John along to show them the way. OPC: They brought John along to talk to.

Jones notes a number of differences between PC and RatC. Before looking at these, we briefly consider their similarities. Clearly, both involve a notional "purposiveness." However, RatC and PC predicate this purposiveness of different arguments. Compare the following simple examples: (148)

a. b. c.

Johni wrote the book [PROi to make money], (RatC) I gave John a booki [PRO to read tj. (OPC) I introduced John to an interpreter; [PROi to read the book]. (SPC)

(For us, SPC and OPC differ in that OPC involve wA-movement of the object, and SPC do not. Cf., however, Jones (op. cit.) for a detailed account of the differences between the two types of PC in rather different terms.) In (148a)/oArt is very clearly the argument of which purpose is predicated. We proposed in the last section how this reading comes about. In (148b) no one argument seems to be so unambiguously the argument with the purpose: not /, not John, not a book (even if this made sense) and not PRO. In light of our treatment of control in RatC, the natural suggestion to make is that PC are not predicated of the Agent. A further similarity is that both kinds of purposives allow parasitic gaps. We discussed the parasitic-gap properties of RatC in the previous section, and so we will not linger over them here. Consider then the possibility for a parasitic gap construction with a PC: (149)

What did you buy t; PRO to read ti

The construction in (149) parallels the properties of parasitic gap constructions seen earlier. (150) shows the Principle C violation noted by Contreras:

The By -Phrase and Implicit Arguments

117

(150) *I bought thenii PRO to read John's books;. The contrast in (151) indicates that John is not subject to Principle C when deeply embedded inside a reconstructed extraposed clause (cf. the contrast between (120) and (121)): (151)

a. *I sent a letterj to himi to make Johnj read tj. b. I sent a letterj to himi to make the guy staying with John; read tj.

In connection with (120) and (121), we saw how this contrast is used by Saito as evidence that parasitic gaps are extraposed. In (152) we see further evidence for this idea. Here the V-modifying Adverb cannot appear following the extraposed constituent: (152)

a. *Which booki did you steal tj for Mary to read ti quickly? b. Which book; did you steal ti quickly for Mary to read ti?

So PCs are extraposed at SS when they contain a variable that may violate Principle C if the clause remains in place. There are four main differences between the two kinds of purposives: (i) PC cannot contain in order. (153) *I bought a booki in order Oi PRO to read ti. ((153) has the irrelevant reading where read is intransitive, but this is not the one we are interested in here.) (ii) PC cannot link up with empty ö^-phrases in any way: (154) *John was given; a bookj IMPi PROi to read tj. (iii) the empty operator in a PC is unable to be coindexed with the subject: (155) *A booki was bought ti [Oi [PRO to read t j . Instead, the empty operator is always predicated of a VP-internal NP if there is one. (iv) PC are not transportable in the manner of RatC (cf. (102)): (156)

a. *To talk to, we brought John. b. *We have, to talk to, brought John.

Most of these facts fall out from one statement regarding PCs: (157)

PCs are predicated with the Theme.

118

The Representation of Implicit and Dethematized Subjects

So PCs appear in VP (and extrapose in the syntax), and are licensed by predication with a VP-internal NP, the Theme. This predication requirement is imposed by the empty operator of the PC. We will now show how (157) derives the differences between RatC and PC we listed above. Regarding (i), in order cannot appear in PC. This is because the properties of in order define RatC. We attributed the selection property of RatC to the Preposition which heads them, in order or 0 . If there is no overt Preposition, it is not possible, other things being equal, to distinguish RatC from PC, since either adjunct may contain a null head. However, the appearance of in order identifies an adjunct as a RatC since this Preposition's selection requirement underlies the properties of RatCs. The fact that PCs are predicated of an internal argument straightforwardly explains the lack of "implicit-argument" effects with these adjuncts as the IMP/6y-phrase can only be assigned the subject (usually Agent) θ-role. The next question is: why is (155) ungrammatical? Given our proposal that parasitic-gap adjuncts extrapose to avoid violating Principle C we can answer this question fairly straightforwardly. The SS (and, presumably, LF) for (155) is (158): (158)

A book; was buy+enj TI EJ EK [CPk Oi [IP PRO to read ti]

Here the variable /,· in the adjunct is bound by the main-clause subject. This violates Principle C. Note that if the adjunct does not move this will not save the structure as t, will still be bound by the subject. This explains the ungrammaticality of (155).13 Note that (155) involves an OPC, not an SPC (cf. (147)). The account of (155) predicts that SPCs should be predicable of subjects, since as SPCs do not involve wA-movement, the absence of variables precludes a Principle C violation. This seems to be true, as pointed out by Williams: (159)

It was bought to hold books.

We can explain the fourth property of PCs by maintaining the assumption from previous sections that "transportability" is really base-generation of the elements in question in their SS positions. The positions occupied by the PC in (156) are VP external, and the empty operator in purposives selects a VP-internal argument for predication. Hence the selection requirement can never be satisfied in (156). This is why these examples are ungrammatical (cf. also the discussion of (132b) and (133) in the previous section).14 How is PRO controlled in SPCs and OPCs? In SPCs PRO generally receives the index of the Theme. This is the case in all the examples of

The By-Phrase and Implicit Arguments

119

SPCs considered in this section. Here again, the parallel with RatCs is straightforward: PRO receives the index of the empty element which heads the adjunct clause. In SPCs this empty element is coindexed with the mainclause Theme, so PRO receives the index of the main-clause Theme. In OPCs, on the other hand, the situation is rather different. PRO in OPCs behaves rather as in wA-complements: (160)

a. b.

Wes brought them [PROi to read]. Wei asked [how PROi to fix the sink].

(161)

a. b.

John sent usi them [PROi to read]. John told us, [how PROi to fix the sink].

This is presumably related to the fact that OPCs involve wA-movement while SPCs do not. We have no account of control in wA-clauses however. This concludes our rather brief account of PCs and their relation to RatC. While RatC behave like Class I Adverbs, it seems that PCs resemble Class IV Adverbs: PCs cannot appear outside VP and may be predicated of Themes. We saw in 3.2.1.1. that these properties are shared by Class IV Adverbs. More generally, the parallels between infinitival adjuncts and Adverbs suggest that the generalization for AP modification made in (81a) holds also for infinitival adjuncts. We repeat (81a) in modified form here: (162)

Infinitival adjuncts are only predicated of structurally θ-marked NPs.

If predicated with an Agent, the adjunct can appear in peripheral positions in IP, but not in VP. If predicated with a Theme, the adjunct must appear in a peripheral position in VP. Together, (81a) and (162) suggest the following generalization about predication with adjuncts, both APs and clausal: (163)

Predication is only possible with structurally θ-marked arguments.

(163) is similar to the following generalization: (164)

"A predicative Adjective can be predicated of an NP if the NP is contained in a semantically unrestricted grammatical position." (Zubizarreta (1985:251), adapted from Bresnan (1982b)).

The basic semantically unrestricted positions are subject position and object position. These are the positions to which any θ-role may be assigned.

120

The Representation

of Implicit and Dethematized

Subjects

Other Α-positions are semantically restricted, in that they only admit particular θ-roles. However, a problem arises concerning how positions are restricted. Zubizarreta, following Bresnan, says that positions are restricted by a (non-dummy) Preposition. A further property which restricts a position may be inherent Case. Neither of these proposals explains why the following cases of predication are impossible, however: (165)

a. *We gave John; a book h a p p y / P R O i to read it. b. *It impressed me; astonished/PROi to investigate phenomenon further that... c. *John baked Maryi a cake drunk/PROi to eat it.

the

The impossibility of predication here follows straightforwardly from (163) on our view, as the arguments predicated with here are not assigned structural θ-roles, but rather Recipient, Experiencer and Beneficiary respectively. There is no need to invoke a distinction between restricted and unrestricted positions; the relevant distinction (which is clearly similar in intent) is that between structural and inherent θ-roles. These facts support the account of θ-role assignment given in 2.2.2. We conclude, then, that our account of modification extends from different classes of Adverbs to different classes of purpose clauses - the parallels between the two kinds of modifier being quite close. We have also seen further evidence for our general theory of modification and for our proposals for its relation to θ-role assignment. 3.2.4.2. Nonpurposive A djuncts Chomsky (1986) claims that implicit arguments cannot be coindexed with PRO inside nonpurposive adjuncts: (166)

a. ??The book was filed [without PRO reading it], b. ??The meal was eaten [without PRO cooking it].

However, the examples in (166) do not seem to be fully ungrammatical, although there is something odd about them. Note that this oddity disappears if a Class I Adverb or RatC is present: (167)

a. b.

The book was filed IMPi [without PROi reading it] [PROi to save time], The meal was reluctantly cooked IMPi [without PROi eating it]·

The By-Phrase and Implicit

Arguments

121

Also, it is possible to have a (3rd Plural) anaphor inside the RatC, which can only be bound by a PRO which gets its index from IMP: (168)

a. b.

The stories were written IMPi [without PRO; checking the facts] [in order PROj to incriminate themselvesi], The rumors were deliberately spread IMPi [in order PROs to embarass eachi other].

For the moment we conclude that coindexation between IMP and PRO is possible here and return to the account of the difference between (166) and ( 1 6 7 - 8 ) below. These clauses are base-generated in VP and optionally extrapose, we assume. The optional extraposition becomes obligatory in parasitic-gap constructions, as a function of Principle C. In fact, these are the standard structures which license parasitic gaps: (169)

Whati did you sell ti [without repairing ti]?

These structures are amenable to an extraposition analysis along the lines proposed by Saito and described in 3.2.4.1., as the following examples show: (170)

a. *John sold them* without framing Mary's paintingSj. b. ?John sold themj without removing the marks which disfigured Mary's paintingSj.

(171)

a. * Whati did John sell ti without framing ti quickly? b. What; did John sell ti quickly without framing ti?

(170) shows the effects with Principle C and names under reconstruction that we observed in the previous sections for both RatC and PC and (171) shows that VP-Adverbs cannot follow this parasitic adjunct. So we conclude that these parasitic adjuncts too are extraposed. Unlike RatC, these adjuncts appear to have no selection relation with an Agent NP in the main clause. Unlike PC, they have no relation with a VP-internal argument. However, they are usually predicated of the argument in subject position. PRO can be coindexed with the subject up to selectional incompatibility; if the whole adjunct is coindexed with the subject this follows from (113). The following examples appear to support this approach: (172)

a. They sold the books [without PRO telling us]. b. *They sold the books [without PRO being proofed].

122

The Representation of Implicit and Dethematized

Subjects

(173)

a. *The books were sold [without PRO telling us]. b. The books were sold [without PRO being proofed].

In the bad examples here, PRO cannot be construed with the matrix subject owing to selectional incompatibility. Actually, selection is not the whole story, as the following example shows: (174)*?The slaves, were sold [ without PROi telling us]. (174) is structurally parallel to (173a), except here there is no selection problem, and the example is still ungrammatical. Whatever rules out this example presumably also rules out (173a), so selectional incompatibility is not the only factor at work in that example. Moreover, there is no selectional incompatibility in any of the following ungrammatical examples: (175)

a. b. c. d.

*The chickens killed [without PRO telling us]. *They ate the chicken [without PRO killing]. *The chicken was eaten [without PRO killing]. *The chickens killed [without PRO strangling].

Once again the notions of agentivity and eventhood seem to be relevant, as in RatC, although these adjuncts are not the same as RatC, as the possibility of control by the matrix subject in (173b) shows. Also nonAgents can clearly control PRO in these adjuncts : (176)

a. b.

Bill struck John as smart [without PRO seeming pompous to him]. John seemed to be drunk [before PRO appearing to be incapable of speaking].

Compare these examples with similar cases with RatC. With RatC there is always a reading of secondary agentivity which is absent (although perhaps optionally present - the judgements are unclear) in (176): (177)

a. b.

The book was reviewed [in order PRO to seem erudite]. The car was driven carefully [in order PRO to strike the police as sober].

In contrast with (176), however, expletive subjects cannot license a nonpurposive adjunct: (178)

a. *It struck John that Fred was smart [before ... ]. b. *It seemed that Fred was smart [without... ].

The By-Phrase and Implicit Arguments

123

We can account for these facts by assuming (113) and (179): (179)

Non-purposive adjuncts are coindexed with Infi.

According to the proposals in 2.2.2., Infi has the index of the subject where the clause is agentive and eventive. So, by (113), PRO will receive the subject's index in such a clause. This is the case in examples like (172a). In clauses which are eventive but not agentive, we propose Infi has an index (which we call an Ε-index) distinct from that of the subject. (This is the temporal index on which V is dependent - see 4.2.). The adjunct clause, and therefore PRO, receives in this index examples like (176). This is also the index of Infi, and thus the adjunct and PRO, where the Agent is not the subject of the clause as in passives like (173b). In other words, these clauses involve Ε-control in the sense of 3.2.4.1.2.(139 - 142). In that section we saw that Ε-control and IMP-control are different; the differences between the non-purposive adjuncts under discussion here and RatC further support that claim. Note also that the examples of E-control which we saw in that section could not license a RatC license a nonpurposive: (180)

a. b. c.

The lightning struck the house [without PRO starting a fire]. John saw [Mary leave] [before PRO going out]. Fred noticed the painting [after PRO entering the room].

So we conclude that Ε-control exists, and is distinct from IMP control. Ε-control is characteristic of non-purposives, IMP control of RatCs. This last generalization follows from (113) and the proposed predication properties of these adjuncts. We clearly have to say more than this, however, in order to account for the possibility of PRO receiving the index of the implicit argument, as in (166) and (167 - 8). For these cases we can broaden the above generalization and say that non-purposives can be coindexed with any predicate in their R-modification domain. This formulation allows PRO to receive the index of a Class I Adverb or RatC in the same clause, and so accounts for (167 - 8). This proposal also allows PRO to receive the index of VP, but such a VP will always be eventive or agentive and so share its index with Infi, so the above proposal concerning E-control is unaffected. Following (113), when a non-purposive coindexes with a RatC or Class I Adverb, PRO receives this index. Since this is the index of the IMP modified by the Class I Adverb or RatC, PRO in the purposive is also coindexed with IMP. This explains the implicit-argument effects in (167) and (168) in contrast with (166). To illustrate this proposal, we repeat (167a) with the proposed coindexing:

124 (181)

The Representation of Implicit and Dethematized

Subjects

The book was filed IMPi [s without PRO* reading it] [ Pi PROs to save time].

If the RatC were contained in the non-purposive this coindexing would be impossible. Compare (181) with the following, where the context strongly favors the reading where the RatC modifies the subject of the non-purposive, and is therefore contained in the non-purposive, and so is not in the nonpurposive's R-modification domain: (182) ??The book was filed [without reading it to learn French]. The awkwardness of this example is due to the fact that the selection requirement of the non-purposive cannot be met for structural reasons. We have to add a little more to the above account of the predication properties of non-purposives in order to account for why these adjuncts can also be predicated with a 6^-phrase. This is shown by the possibility of an anaphor in the adjunct phrase: (183)

The book was filed by John* [without PROj reading it to himself,].

What underlies the contrast between (183) and (166)? We have no explanation for this at this point, but will return to this in 3.3.4., where we observe similar contrasts in the ability of IMPs to bind anaphors. One question that has been left open here concerns the fact that extraction from these adjuncts is impossible: (184) *?Whoi did they sell the books without telling ti? It is clear that the phrase without telling is not subcategorized by V, or by anything else. Hence, it is not lexically governed. Since without telling is a CP, this CP will be a barrier and so will the IP immediately dominated by CP (pace Chomsky (1985)). Hence the adjunct acts as an island whether it is in VP or not. This means that (184) is ruled out by subjacency. This concludes our account of infinitival adjuncts. We have seen that RatC must appear in a configuration of Agent θ-role assignment, PC select a Theme, and non-purposive adjuncts select a predicate (which may be Infi, a RatC, a Class I Adverb or V). These differences account for many properties of these adjuncts, and for implicit arguments fairly straightforwardly. Also, the general theory of predication and θ-role-assignment that we have put forward here makes interesting predictions about modification, as we saw at the end of the previous section. This theory of predication and modification has a good deal of independent motivation and accounts for the facts concerning implicit arguments. We conclude,

The By-Phrase and Implicit

Arguments

125

then, that implicit arguments are structurally present elements in passive constructions. 3.3. REFERENTIAL PROPETIES OF IMPLICIT ARGUMENTS

So far in this Chapter we have concentrated on the predication properties of IMPs and 6y-phrases. In the preceding sections we have accumulated a good deal of evidence that this element enters into predication relations. This evidence, combined with our theory of predication, has led us to conclude that the IMP is structurally present. This conclusion supports our theory of passive. If the IMP/6j-phrase can be the "subject" of predication, it must be a referential expression of some kind. In this section we address the question of the referential properties of IMPs and, in less detail, ^-phrases. Our discussion of the referential properties of IMPs falls into two parts, each dealing with a different aspect of grammatical theory relevant to reference: binding theory and quantification. We will cover these areas and build up a fairly clear picture of the semantic properties of IMPs: we will see that they are variables carrying the subject θ-role and are bound under existential closure in LF. Our proposals for IMPs predict, given general principles of binding and quantification, certain rather surprising constraints in the interpretation of passives. The fact that these constraints can be explained in general terms is a further argument for the general theory of passives proposed here. Furthermore, our investigation of the binding and quantificational properties of IMPs will lead us to reformulate some of the relevant principles in ways which receive independent motivation from constructions unrelated to passives. 3.3.1. Binding Theory and Θ-theory In 1.4.(31) we presented the Binding Principles of Chomsky (1981). Since these principles will be of some importance in what follows we repeat them here: (185)

A: An anaphor must be Α-bound in its binding domain. B: A pronominal must be Α-free in its binding domain. C: An R-expression must be A-free.

Recall that X binds Y iff X c-commands Y and X and Y are coindexed. In the presentation in 1.4., we focussed attention on the ways in which the binding principles determine the distribution of empty categories. These principles also place conditions on the distribution of overt NPs. The class

126

The Representation of Implicit and Dethematized Subjects

of anaphors includes lexical reflexives (himself etc.) and reciprocals (each other), as well as NP-traces. Both overt and empty pronouns are pronominals, and names and variables are R-expressions. Recall also that we imposed the following condition on binding relations in Chapter Two (118): (186)

If X binds Y then X agrees in grammatical features with Y.

As we saw in Chapter Two (pp.59-60), (186) rules out examples like (187): (187) *Johni likes herselfj. The binding domain for X is the smallest NP or CP containing X, the governor of X, and a subject accessible to X (but cf. fn. 5 of Chapter One for the more recent formulation in Chomsky (1986)). The accessible subject for X as the closest c-commanding subject to X which can be coindexed with X. Taking the binding conditions as axioms, we can derive a number of theorems about the distribution and reference of elements in a binding domain. Those of interest to us here are the following: (188)

a. b.

In a normal transitive clause, an argument in VP is coreferential with the subject only if it is an anaphor. Reflexives and reciprocals have identical distribution.

The truth of (188a) can be seen from examples like the following: (189)

a. *Johni saw Johni. b. *Johni saw him;. c. *Hei saw John;. d. John; saw himselfi. e. Theyi saw each other;.

The truth of (188b) can be seen from (189d,e) and also from: (190)

a. *Himselfi/each otheri saw thenii. b. *Theyi want Fred to wash each other/themselvesi.

However, there are ^-phrases where (189a) fails to hold (cf. Lees & Klima (1963:21), Postal (1971:7f.,15f.)): (191)

a. ??The man was seen by himself in the mirror, b. ??She was inadvertently electrocuted by herself.

The By-Phrase and Implicit Arguments

127

c. ??You are tortured by yourself/selves. d. ??We are shaved by ourselves. The same is also true of IMPs: (192)

a. b.

They They were were killed. admired.

(192a) cannot mean "they committed suicide" and (192b) cannot mean "they admired themselves." This does not seem to be a pragmatic constraint, as the following contrast between passives and middles shows: (193)

a. b.

John was shaved, John shaves easily.

In (193) the understood Agent cannot be John, while (193b) it can. This minimal contrast indicates that there is no pragmatic constraint ruling out these readings in (192), or the overt reflexives in the Λ^-phrases of (191). The explanation for the ungrammaticality of (191) and the impossibility of a reflexive reading for the IMP in (192) is the chain-formation algorithm (CFA) of Chapter One (16), combined with the θ-criterion and the Projection Principle. The CFA requires that each position in a chain locally bind the next, θ-chains are those chains for which the θ-criterion holds. The Projection Principle effectively requires that subcategorized arguments appear in well-formed θ-chains at all syntactic levels (cf. 2.2.3.(103)). We assign to (192a) the representation in (194): (194)

Theys were kill+eni ts

IMPi.

In (194), they and en are both in non-0-positions. They must be in a wellformed θ-chain in accordance with the Projection Principle as it is the subcategorized argument of kill. The passive argument of kill, the chain (en, IMP), must also appear in a well-formed θ-chain following the condition that root clauses must be CFCs (Chapter Two (107)). So both arguments must form chains with some element in a θ-position. In fact, this requirement is already met by en thanks to the existence of the clitic (DS) chain formed between en and IMP. The problem is then: how can the SS subject, they, form a well-formed θ-chain with tP. Since they is coindexed with en there is no way to form chain which meets the θ-criterion, in accordance with the CFA. So either the CFA is violated, by forming a chain which "skips" the intervening coindexed en, or the θ-criterion is violated by the formation of a chain with two θ-roles, or the Projection Principle is violated by the presence of the argument they without a θ-role.

128

The Representation of Implicit and Dethematized

Subjects

Rizzi points out that this kind of account will rule out all cases which are schematized as follows: (195)

XitzY.ti] (where X is in a non-9-position and Y is an argument).

In these terms, an example like they were killed is prevented from having the reading where the individuals in question committed suicide, as in that case Xi = they and Y, = en. So we can account for the impossibility of this reading. Moreover, examples containing reflexive 6y-phrases, as in (191), have the S-Structure (196): (196)

Theyi were see+eni t, by themselves;.

This is another instance of the schema in (195), so (191) is ruled out in the same way as (192). There are two problems with this account of (191) and (192). First, some Verbs seem to allow a reflexive passive Z^-phrase: (197)

a. ?I am amused by myself. b. ?We are impressed by ourselves. c. ?You are worried by yourself.

These examples are certainly better than those in (191). We explain the relative acceptability of (191) by exploiting the fact that the Verbs in question are all "psych-Verbs." That is, they are Verbs whose argument structure involves a Theme and an Experiencer, and no Agent. Belletti & Rizzi (forthcoming) argue that such Verbs are underlyingly ditransitive unaccusatives, having two DS objects and a nonthematic subject. If this is so, then it follows as a subcase of the 1-Advancement Exclusiveness Law, which for us follows from the θ-criterion (see 2.2.1.2.3.), that these Verbs cannot be passivized. If they cannot be passivized then the only possible derivation for examples like (197) involves an Adjectival passive. Assuming a standard view of Adjectival passives, such as that in Wasow (1977), these passives are derived in the lexicon. The passive morphology simply forms a derived Adjective, and there is no reason to assume the presence of a logical-subject argument in any way. If the morphology is not an argument, then the examples in (197) do not instantiate the schema in (195), hence there should be no problem with their grammaticality. Support for the view that these are Adjectival passives (or at least passives where en is not an argument) comes from the fact that implicit arguments are not found here:

The By -Phrase and Implicit (198)

Arguments

129

a. *I am amused by myself [to pass the time]. b. *We are voluntarily impressed by ourselves. c. *You are intentionally worried by yourself [to get attention].

It is possible that the above account can be related to the fact that the examples in (191) are not as bad as the intended readings of (192). There is, to a certain extent, the possibility of understanding the examples in (191) as Adjectival passives. In (192) there is no such possibility, if the IMP is be interpreted as coreferential with the SS subject. Compare (191) with examples which could not possibly be Adjectival passives: (199)

a. b. c. d.

*John is considered brilliant/a fool by himself. *Mary was baked a cake by herself. *Fred was promised to leave by himself. *John was sent a letter by himself.

Furthermore, it is a general property of Adjectival passives that they are stative (see 4.1.). Stative predicates cannot appear in the progressive, in pseudoclefts, in imperatives or in complements to control Verbs. If the examples in (191) are interpretable as Adjectival passives, then they should be ungrammatical in these environments. Cf.: (200)

a. b. c. d.

*The man was being seen by himself in the mirror. *They forced the man to be seen by himself. *What the man did was be seen by himself. *Be seen by yourself!

These examples cannot be Adjectival passives as the aspectual conditions on this construction are not met, and they cannot be syntactic passives for the reasons spelled out above (cf. 4.1. for more detail on Adjectival passives and a comparison betwieen them and middles). So, once we extrapolate away from the possibility of an Adjectivalpassive interpretation, the examples with reflexives in ^ - p h r a s e s become much worse. These sentences fall unambiguously under the schema in (195), and so are ruled out by conditions on θ-chains. The second problem for the above account of examples like (91) and (199) is that reciprocals are possible in ^-phrases: (201)

a.

They *They b. They *They

are considered smart by each other. are considered smart by themselves, were seen by each other. were seen by themselves.

130

The Representation of Implicit and Dethematized Subjects

Consider also the following examples: (202)

a. b. c. d.

The men were seen by each other in the mirror. They were inadvertently electrocuted by each other. We are amused by each other. You are tortured by each other.

To account for these contrasts we have to consider the inherent differences between reflexives and reciprocals. One proposal for the difference between reflexives and reciprocals appears in Lebeaux (1983). Lebeaux notes the following contrast: (203)

a. John and Mary brought some friends for each other to meet, b. * John would like some books for himself to read.

(204)

a. ??John and Mary wonder whether each other will win. b. *John wonders whether himself will win.

Lebeaux concludes that reflexives can only appear in properly governed positions, while reciprocals can appear in non-properly governed positions. So Lebeaux postulates certain differences between the two kinds of anaphor that derive this contrast while leaving Principle A intact. For reciprocals, Lebeaux proposes a rule adjoining each to the nearest VP. Here we propose a slight extension of Lebeaux's proposal. We will exploit the proposals he makes about each and other, respectively, to the full: (205)

a. b.

each is a quantifier. other is a disjunctive pronoun.

That each is a quantifier is uncontroversial. As a quantifier, we expect it to undergo a form of QR, along the lines described in Chapter One (25 - 7). We concur with Lebeaux's suggestion that each undergoes QR to adjoin to VP. This proposal means that each appears as a bare quantifier at LF. Each can also appear as a bare quantifier at SS: (206)

a. b.

The men each saw the other(s). John and Mary each like chocolate.

Jaeggli (1982, Chapter Two), following a suggestion made by Belletti (1982a), proposes that bare quantifiers are in general anaphors. He cites the following English paradigm to demonstrate this:

The By-Phrase and Implicit Arguments (207)

a. The kids b. The kids c. *The kids d. *The kids

131

will all go to the movies. tried [PRO to all go to the movies]. forced John [PRO to all go to the movies]. said that John would all go to the movies. (Jaeggli's (2.55), p. 77).

The same paradigm holds for bare each (as pointed out in Chomsky (1973)): (208)

a. The kids b. The kids c. *The kids d. *The kids

will each go to the movies. tried [PRO to each go to the movies]. forced John [PRO to each go to the movies]. said that John would each go to the movies.

If bare quantifiers in general are anaphors, and each of each other undergoes QR out of this phrase at LF, then we expect it to be an anaphor at LF. When each and each other undergoes QR, it adjoins to the nearest VP, we have said. There is evidence that the bare each of (208) can adjoin to VP but cannot appear inside VP. The evidence is that bare each can appear anywhere in a sequence of auxiliaries (although with increasing awkwarsness), but cannot follow the main Verb: (209)

a. They each must have been being beaten. b. They must each have been being beaten. c. They must have each been being beaten. d. ?They must have been each being beaten. e. ??They must have been being each beaten. f. *They must have been being beaten each.

So we propose that each in reciprocals appears at LF in a position which can be occupied by "bare" each at SS - namely, adjoined to VP. To complete the parallel, we allow for the possibility that reciprocal each can appear before the first auxiliary, and therefore outside VP, at LF, just as bare each can at SS, as (209a) shows.15 So we conclude that reciprocal each is a bare quantifier at LF adjoined to VP (or perhaps IP) at LF, and therefore an anaphor. We assume that other is a disjunctive pronoun. We interpret this requirement to mean that it must be contraindexed with each. Other is the head of each other, and so its index percolates to the NP. An example like they were seen by each other, then, has a representation like (210): (210)

Theyi each; were see+enj ti by [tj other^

This representation is not an instance of the ill-formed θ-chain configuration

132

The Representation of Implicit and Dethematized

Subjects

of (195) since for (195) to result in a violation Xi and Yj must both be arguments, i.e. subject to the Projection Principle. Neither each nor U are arguments, and so (210) is allowed. Owing to the contraindexing of other, en, which is an argument, does not cause a violation of conditions on chains, as happens where there is a reflexive in the ό^-phrase. Note that, if the indexing in (210) means that each and other belong to different θ-chains then the two parts of the reciprocal receive different θ-roles, each being in a θ-chain with they and other with en. This is intuitively the right account of the thematic properties of reciprocals: reciprocal clauses like they hit each other are interpreted such that the members of the set denoted by the antecedent of each hit other members ofthat set exhaustively. The assumptions we have made about each other, following Lebeaux, seem compatible with this interpretation. 16 As for reflexives, Lebeaux proposes that (203b) and (204b) are ruled out by the ECP as the reflexive is not in a properly governed position at LF. Lebeaux makes two suggestions about how to implement this idea. One is that reflexives involve X-self movement in the derivation to LF. The LF representation for (203b) and (204b) would then be as in (211): (211)

NPj

self + V

[CP [ C O M P

for/wh ]

[IP

ti

...

Here the clitic self is unable to antecedent-govern the lower subject because there is something in Comp. Hence ti in (211) violates the ECP. However, this account runs into a problem, illustrated in (212): (212)

They were shown to themselves in the mirror.

(212), with X-se#"-movement, has an LF structure which corresponds to that in (195), and is therefore ruled out by the CFA and the θ-criterion. However, the sentence is clearly good (this is pointed out in Epstein (1986), who proposes retaining an account based on (195) while changing the definition of local binding). Because of this, we abandon this approach in favor of Lebeaux's other alternative: the view that reflexives become null at LF. If the reflexive becomes null at LF, what kind of empty category is it? Since reflexives have features of number, person, gender and Case, the empty category they "become" at LF is presumably pronominal. Moreover, as Lebeaux points out, they could not be any empty category except pro: "that it [i.e. the null reflexive - IGR] is .. not a wh-trace is predictable from its being Α-bound; that it is not an NP-trace is predictable from the fact that its antecedent is in a theta position; that it is not PRO is predictable from its being in a governed position" (727). Hence the null reflexive must be pro.

The By-Phrase and Implicit

Arguments

133

If the null reflexive is pro, it is not subject to the ECP. However, it is subject to the identification requirement on pro given in (127) of Chapter Two. This identification requirement on null reflexive pro is as follows: (213)

a. b.

Null reflexive pro must be governed by a lexical head X. Null reflexive pro must appear in Pred(X).

Pred(X) is defined here as in Chapter Two (108). This identification requirement is very similar to the lexical government requirement of the ECP, hence Lebeaux's ECP-based account of the ungrammaticality of (203b) and (204b) will apply mutatis mutandis: the null reflexives will be ruled out at LF as they will not meet the identification condition (213). (213) contains no requirement that the null reflexive pro must be governed by the element which θ-marks it; this is to allow for cases where reflexives appear as ECM subjects (but they cannot appear as subjects of tensed clauses as this position is not governed by the matrix Verb). This provision also allows us to extend the identification conditions on null reflexive pro to each, and thus to anaphors in general. 17 Now Lebeaux's account of the contrast in (203 - 4) applies in full, and without the problems which a ie/f-movement hypothesis poses with respect to chain-formation. 18 In this section, we have seen how binding theory and θ-theory interact to rule out the ungrammatical passives of (191) and (192). We proposed (i) that the ungrammaticality of these examples can be explained in terms of conditions on θ-chain formation and (ii) that reflexive/reciprocal differences in 6^-phrases can be explained by a Lebeaux-type account of each other. (191) and (192) are cases of strong crossover in passives (cf. Postal (1971), Jackendoff (1972)), so our account of these examples is a contribution to the theory of crossover, which is generally held to be derived from binding theory. In the next section, we will extend the Θchain account of crossover. 3.3.2. Strong Crossover, Chains and Principle C In this section we extend the account of strong cossover given in the previous section to a number of other cases discussed by Postal (1971), Lasnik (1985) and Rizzi (1983). Second, we consider the relationship between a chain-based account of crossover and the account of "standard" cases (i.e. those which involve w/z-movement, cf. (32) of Chapter One), which involves Principle C. We propose first that the chain-based account of the previous section (which is inspired by Rizzi (1983)), combined with the Bijection Principle expressed as a condition on chains (see Chapter One (28) for a standard formulation of this principle), can explain standard strong crossover cases (the same point is argued in Hoekstra (1984)). This

134

The Representation of Implicit and Dethematized

Subjects

leads to a redundancy with respect to Principle C. We will show that a chain-based approach to crossover is both empirically and conceptually superior to one based on Principle C. The conclusion is that the part of Principle C which deals with variables might be derivable from conditions on chains. Third, we consider the consequences of this approach for movealpha: we are led to regard move-alpha as simply an operation which moves categories; coindexation of the moved element and its trace is not a part of this operation but rather supplied by conditions on chainformation. The chain-based account of crossover in passives proposed in the previous section is based on a proposal of Rizzi (1983)). Rizzi's example is the following: (214) *Giannij si§ e stato affidato t'i t". G to-himself was entrusted. This example, like those discussed in 3.3.1., falls under the general schema of (195): Gianni moves from Λ to the subject position, and si is an intervening coindexed element. Hence the structure is impossible. Another class of cases discussed by Rizzi involves raising. We predict the grammaticality of (215a) and (216a), in contrast with (215b) and (216b): (215)

a. Theyi seem to [eachs other]j ti to like John, b. ??Theyi seem to themselvesi ti to like John.

(216)

a. Theyi strike [eachi other]] as ti intelligent, b. ?*Theyi strike themselvesi as ti intelligent.

(215b) and (216b) are ruled out by the θ-criterion in the same way as the passive examples we saw in the previous section: they cannot form a well-formed θ-chain with its trace because of the intervening reflexive. (215a) and (216a) involve no such violation, given our proposals about reciprocals. Rizzi (1983) discusses (215a) and (216a) and suggests that the reason these cases do not violate the θ-criterion is that each other does not c-command the subject position of the lower clause, and hence does not count as an intervening binder interrupting the formation of the chain between they and /,·. With seem the lack of c-command between each other and the lower subject is explained as a function of the presence of the Preposition to. For (216), Rizzi suggests the following structure (p. 19): (217)

NP [Vp [vp strike NP] as .... ]

Our proposals, on the other hand, require that the position occupied by

The By-Phrase and Implicit

Arguments

135

the anaphors in (215) and (216) does c-command the lower subject. Then we will find exactly the contrast between reflexives and reciprocals found in passive 6y-phrases, and for exactly the same reason. Rizzi gives three arguments to support the proposal that there is no c-command between the position occupied by the anaphors in (215) and (216) and the embedded subject position. First, non-clitic anaphors are possible as the indirect object of the Italian counterpart of seem, sembrare (although for some, probably independent, reason they must appear clauseinitially): (218)

a.

b.

Ormai, perfino a se stessoi Gianni; sembra [es non fare il suo dovere]. "Now, even to himself, G seems not to do his duty." L'uno all'altroi, i due candidate risultavano [ei poter vincere]. "To each other, the two candidates appeared to be able to win."

For these examples, it clearly seems that the Preposition a prevents the anaphor from c-commanding out of its PP and so prevents it from interfering in the formation of the chain between the matrix subject position and the embedded subject position. However, Italian and English differ independently with respect to the availability of a rule which "restructures" a Verb with a subcategorized Preposition. The evidence for this restructuring process in English comes from the existence of the pseudopassive construction, illustrated in (219): (219)

a. b.

The bed; was slept in tj. The subjects was spoken of ti.

In some sense, sleep and in here become a complex Verb. We suggest that this happens as a result of the free adjunction of Ρ to V (a kind of "incorporation" in the sense of Baker (1985b)). The adjunction of Ρ to V in English allows en to receive ACC from in {sleep, as an intransitive, has no ACC: cf. *It was slept). This restructuring can only take place with adjacent, subcategorized complements, as (220) shows: (220)

a. *The bar was danced near, b. *John was given a book to.

After Ρ adjoins to V, the object of the complex Verb c-commands everything in VP. This is the situation in English. In Romance languages, on the other hand, this kind of adjunction is impossible, as the impossibility of pseudopassives in these languages shows. Hence the difference in the

136

The Representation of Implicit and Dethematized

Subjects

possibility of Preposition-incorporation in English as compared to Italian could underlie the difference between (218) and (215b, 216b). Rizzi's second argument that each other c-commands the lower subject, at least in (216a), is based on the fact that this position can license a parasitic gap. This is shown in (221): (221) ?Whoi did the pamphlet strike ti as insulting to ei. If, as argued in Engdahl (1983) and Chomsky (1982), the position which licenses a parasitic gap cannot c-command that gap, then clearly the object of strike in (221) cannot c-command the object position of the lower clause, and so does not c-command any position in that clause. However, we saw in 3.2.4.1.2. that the generalization that the position which licenses a parasitic gap cannot c-command that gap has been questioned, and alternative accounts have been proposed which allow for c-command between licensing gaps and parasitic gaps, at least at certain levels. In terms of the account we adopted in 3.2.4.1.2., the ability of the object of strike to license a parasitic gap in the as-phrase becomes a question of the ability of üi-phrase to extrapose. In that case, (221) is not evidence against the proposal that the object of strike c-commands the subject position of the a^-clause. Third, Rizzi observes that (222) has the flavor of a weak crossover violation: (222)

a. ?John strikes everyonei as more intelligent than himi. b. ?John seems to everyone; to be more intelligent than him;.

If the SS position of everyone here does not c-command the embedded subject position, (222) has an LF-structure where everyone binds two variable-chains, in violation of the Bijection Principle of Chapter One (28). Hence (222) is an instance of weak crossover exactly like the paradigm case illustrated in Chapter One (34). The above account of the ungrammaticality of (222) only goes through if we assume that everyone does not c-command him. This account, however, also rules out examples like (223), which do not seem so bad: (223)

a.

John seems to everyone, to like hinij.

b.

John strikes everycnej as more intelligent than he; is.

Also, compare these examples with the much worse (224): (224)

a. *Mary seems to him; to like John,. b. *It strikes hinij that Mary likes John;.

The By-Phrase and Implicit

Arguments

137

The latter examples are ungrammatical. This can be explained in terms of Principle C, if we allow that him c-commands John here. This approach is supported by the grammaticality of Mary seems to hisi mother to like Johnu The effect of Principle C here seems much stronger than the violation in (222), so this is a further argument that there is c-command between him and John here, i.e. the indirect object of seem does c-command the subject of the clausal complement. In general, then, it is possible to maintain the view that in English there is c-command between the indirect object position and the embedded subject position in examples like (215) and (216) (this conclusion also means that we assume a more conventional complement structure for strike rtian that in (217)). The examples of (215) and (216) are pointed out by Postal (1971). Another case which was pointed out by Postal is that of tough-movement. Here again we find the same reflexive-reciprocal contrast: (225)

a. *They are tough for themselves to understand, b. They are tough for each other to understand.

The proposal made to account for the reflexive-reciprocal contrast in the previous section carries over here whether we assume an NP-movement or a wA-movement derivation of tough sentences. The contrast is illustrated in (226): (226)

a. b.

Theyi are tough for themselves; (Oi PRO,) to understand ti Theyi are each; tough for [ti other^ (Oi PROi) to understand tj.

If these examples are derived by NP-movement, the θ-criterion is violated in the same way as in the passive and raising cases discussed earlier. If they are derived by wA-movement, we must attribute their ungrammaticality to the binding relation between the reflexive and one of the categories in the lower clause which has this index. Note that the above distinction carries over to the impossibility of interpreting an empty for-phrase as coreferent with the subject. Again, an NP-movement analysis subsumes this under the account of the similar impossibility of interpreting empty ^ - p h r a s e s in passives as coreferent with the subject. A further example that may fall under the θ-criterion viewed in this way, given certain additional assumptions, is that of "symmetric predicates," illustrated in (227): (227)

a. b.

John and Mary met. John shaved.

138

The Representation of Implicit and Dethematized

Subjects

These Verbs are interpreted with obligatory coreference between their two arguments. In fact, (227a) has a reciprocal interpretation for its second argument, while (227b) has a reflexive interpretation. These interpretations are impossible in the passives which correspond to (227): (228)

a. b.

John was met. John was shaved.

In (228) coreference between IMP and the subject is impossible. We can explain this in terms of chain-formation. On the reading where the ivyphrase is coreferent with the subject, (228) has the representation (229): (229)

Johni was meet+ens ti I M P i .

Clearly, this is another case where no well-formed chain can be formed which includes John, so the θ-criterion rules this reading out. The fact there is no analog to the earlier reflexive/reciprocal distinction here indicates that the earlier distinction was structural rather than semantic, as here there is the same semantic distinction as before but there is no structural distinction. Finally, Lasnik (1985) discusses a number of cases of "illicit NP movement." These are instances of NP-movement which are generally ruled out by principle A, but accidental coindexation of the trace with a pronoun provides a "loophole" in binding theory: (230)

a. * Johni is believed he; likes ti. b. *Bill tried Johni to be believed that he; likes ti.

In both of these examples, the NP-trace is bound in its binding domain by the subject pronoun, which is "accidentally" coindexed with it. As a result, Principle A is satisfied. However, the chains in these examples are ill-formed in exactly the same way as the other cases we have considered: there is no well-formed θ-chain which includes exactly John and its trace, as the pronoun which saves the structure from Principle A prevents chain formation between these two categories. Our adaptation of Rizzi's proposals about chain-formation account for the restrictions on the interpretation of passives we observed in the previous section, and similar restrictions on raising, /owgA-movement, symmetric predicates and examples discussed by Rizzi and Lasnik. As we pointed out, the passive, raising and tough cases are discussed in Postal (1971) under the general heading of "strong crossover." The connection between these cases and standard cases of strong crossover is clear: both involve representations of the general form of (195), repeated here:

The By -Phrase and Implicit (231)

Arguments

139

XiYiZi

where X c-commands Y and Y c-commands Z. The only difference between the cases discussed above and standard strong crossover is that X, in (231) is in an Α-position in the former cases but in an A'-position in the latter. The paradigm case of strong crossover with wA-movement is (232) (repeated from (32) of Chapter One): (232) *Whoi does hei love ti? The standard account, sketched in 1.4., is that this structure violates Principle C, as the variable is Α-bound by hei. However, the similarity to the examples discussed up to now suggests that a deeper generalization could be obtained if we extended a chain-based account of strong crossover to cover A'-chains. To do this, we propose that A'-chains are those chains which obey the Bijection Principle, and the Bijection Principle requires operators and variables to form chains where they are in a bijective relationship (cf. Chapter One (28, 29)). Although the Bijection Principle is logically independent of the θ-criterion, an A'-chain may also be θ-chain, and therefore subject to the θ-criterion in addition to the Bijection Principle. This is in fact the case with wA-movement from argument position, since the wA-phrase is both an operator and an argument. This move allows us to rule out examples like (232) in combination with the θ-criterion. In (232) there are four possible chains headed by wh: (233)

a. b. c. d.

wh, he, t. wh. wh, he. wh, t.

The chain in (233a) satisfies the Bijection Principle, but is ruled out by the θ-criterion, as the chain formed by wh and the two variables it binds has two θ-roles. If wh forms a single member chain, as in (233b), both the θ-criterion and the Bijection Principle are violated because wh has no θ-role and the variable is free. The chain (wh, he) in (233c) satisfies the θ-criterion, as it has only one θ-role, but the Bijection Principle is violated as he and the wh-trace must be counted as distinct variable chains, leaving the wh-trace as a free variable. Finally the chain (wh, t) of (233d), although it satisfies the θ-criterion and the Bijection Principle, cannot be formed because he is an intervening binder between wh and the trace. In standard cases of strong crossover, then, it is possible to derive the effect of Principle C from an extension of chain formation. This allows a unified account of all cases of strong crossover.

140

The Representation of Implicit and Dethematized

Subjects

At this point we should address the theoretical issues implicated by our theory of strong crossover. There are two issues: the relationship between chains and Principle C and the relationship between chains and movement. In 3.2.4.1.2.(122) we formulated Principle C, following Chomsky (1986), as follows: (234)

a. b.

A name must be Α-free in the domain of its operator, A name must be A-free.

The two parts of this condition apply disjunctively, i.e. (234a) applies if it can, otherwise (234b) applies. Effectively, (234a) applies to variables, i.e. wA-traces, and (234b) to non-pronominal, non-anaphoric NPs. Imposing the Bijection Principle as a condition on A'-chains derives the effects of (234a) for strong crossover. Can we now eliminate (234a)? We also invoked (234a) to force extraposition in parasitic gap constructions. These constructions resemble strong crossover in containing two variables apparently bound by the same operator. However, Chomsky (1985) has argued, on the basis of evidence from subjacency, that the parasitic gap is bound by its own null wA-element. The evidence is that all the standard island effects are found inside parasitic gaps: (235)

This is the man who I interviewed before ... a. ??...I knew which job to give to. b. ??...hearing the rumor that noone likes. c. ??...meeting the woman who loves.

Chomsky concludes that parasitic gaps have the structure in (236a) rather than that in (236b): (236)

a. b.

whi ts [Of U]. whi ti [ ef|.

In (236a), according to Chomsky, the null operator and the overt wh form a "composed chain." One of the requirements on the composed chain is a lack of c-command between the "real" variable and the null operator. If we eliminate (234a) as a separate principle, we can still force extraposition of parasitic adjuncts in the relevant cases by applying (234b) to the null operator. This has exactly the result of deriving the anti-ccommand requirement in all the relevant cases. So (234a) may be derivable from conditions on chains, leaving (234b) as the statement of Principle C. We noted in 3.2.4.1.2. that the condition on names behaves differently to the condition on variables under reconstruction. Now we can see why

The By-Phrase and Implicit

Arguments

141

this is: the two conditions are quite different, one being a condition on chains, the other being a condition on names. If the condition on the null operator is the same as the condition on names, however, this result will not be achieved. So we need to impose a stronger condition on null operators. The relevant condition is a reformulation of the Bijection Principle: (237)

Each variable (chain) can only be bound by one and only one operator, and each operator can only bind one and only one variable (chain).

(237) prevents the null operator from being c-commanded by the overt wh in parasitic-gap constructions, and so forces extraposition of the parasitic adjunct. (237) effectively replaces (234b) for the null operator, and so we can maintain the difference in the behavior of names and parasitic gaps under reconstruction. ((237) holds at LF only; we still assume that A'binding is the SS requirement on variables, as in 1.4.(29a)). We have now derived much of Principle C from the Bijection Principle and the θ-criterion viewed as conditions on chains. This condition is still independently needed in order to rule out simple cases of Α-bound names such as (238): (238) *Johni saw Johni. However, it may be possible to extend the agreement condition on binding relations of Chapter Two (118) to rule out agreement between "overspecified" elements. The intuition behind this proposal is that only elements which are in some way underspecified for features must agree. Full NPs may not agree with each other as they are not underspecified. If agreement is construed as copying of features from one element to another this idea can be implemented; copying of features is a consequence of binding, and full NPs cannot have features copied onto them (cf. the discussion of clitic chains and clitic-doubling in 2.2.4.). In any case we conclude that much of Principle C can be derived from the Bijection Principle and the θ-criterion as conditions on chains. Turning to the the second issue, we now consider the relation between chain-formation and move-alpha. One issue which arises once we have θ-chains in the theory concerns whether move-alpha is properly to be considered a representational or a derivational operation. In other words, can DS be factored out from SS by means of a chain-formation algorithm which duplicates move-alpha, or are DS and SS truly different levels of representation, associated by the movement rule move-alpha? There is a redundancy between the two mechanisms: chain-formation guarantees that

142

The Representation of Implicit and Dethematized

Subjects

DS can be "factored out" from SS and move-alpha actually derives SS from DS. It thus appears that we do not need both. In fact, it is only true that there is a redundancy between chain-formation and move-alpha if we assume (i) the Extended Projection Principle, and (ii) that the operation of move-alpha involves leaving a coindexed trace. Given a version of the Projection Principle like the one we have proposed (cf. 2.2.3.(103)), and the consequent possibility of movement from subject position into VP, we must assume that move-alpha does not obligatorily leave a trace. If it did, our reformulation of the Projection Principle would be pointless as Principle A would always rule out the trace of downward movement, so our theory of passive would be inoperable no matter how we construed the Projection Principle. If move-alpha does not have to leave a trace, there is not an absolute redundancy between chain formation and movealpha. The presence of traces can be regarded in general as a function of the Projection Principle, and the coindexation of traces as a function of chain-formation - subject to the θ-criterion (ultimately again because of the Projection Principle). In that case, we can say simply that movealpha moves categories and nothing more, while chain formation is responsible for the assignment of an index to the trace left behind by movement. The independence of these operations can be seen in passives: movement of the DS subject into VP is an instance of move-alpha where no trace is left, as the Projection Principle does not apply to this position, so chain-formation is irrelevant. The further movement of a coindexed DS object to subject position leaves a trace whose coindexation with en will violate the θ-criterion, and whose failure to coindex with en (and therefore the subject) will violate the local-binding condition on chain formation. Move-alpha operates oblivious to all this, however. Chain-formation itself is a formal relation between positions that is logically independent of conditions on the representation of arguments like the Projection Principle. Hence not all chains are subject to the Projection Principle, since chains are not just representations of arguments, and conversely certain chains which represent arguments may be subject to stronger conditions than the Projection Principle. The class of chains which are independent of the representation of argument structure includes composed chains formed between parasiticgap chains and the chain headed by the overt wh (cf. (236b) and Chomsky (1985)), "control chains" (cf. Chapter One fns. 3 and 5, and 3.2.) and the chains formed between auxiliaries and other elements (cf. 2.2.1.3.1.). There is also at least one kind of chain whose existence is not a function of the Projection Principle (alone): clitic chains. Coindexation between a clitic and the Α-position it governs remains throughout the derivation. The Projection Principle depends on a definition of the notion of "argument." Although we will not attempt such a definition here, it is clear

The By-Phrase and Implicit Arguments

143

that one property of arguments is that they are integral in the sense that they cannot be broken up during a derivation. Since clitic chains represent "discontinuous arguments" they cannot be broken up; the coindexation and government between the clitic and the Α-position must be constant throughout the derivation. This stronger requirement is independent of the Projection Principle, in fact it is a precondition for the Projection Principle. This concludes our account of strong crossover and its relationship to chain-formation and Principle C. We have proposed a unified account of both NP-movement strong crossover and wA-movement strong crossover in terms of chain-formation. This account allows us to eliminate at least part of Principle C, and we suggested that perhaps the whole of Principle C can be derived from the interaction of other principles. In the next section we turn to weak crossover and the semantic properties of implicit arguments. 3.3.3. The Interpretation of Implicit Arguments In this section we will consider what the properties of the chain (en, IMP/ &y-phrase) are at LF. We continue to assume that this is a clitic chain (see 2.2.4.) where IMP is a variable, locally bound in syntax by en, and the ^ - p h r a s e doubles en. At LF, however, IMP must be bound by an operator, following our interpretation of the Bijection Principle (which we now assume to hold at LF in the form given in (237)). Much of this section will be devoted to showing the various ways in which IMP meets this requirement. In this connection, we will consider the scope and crossover properties of IMP and, to a lesser extent, the fry-phrase. 3.3.3.1. Weak Crossover in Passives In addition to the cases of "NP-movement strong crossover," discussed in the previous section, there are examples which resemble weak crossover involving NP-movement: (239) *[Hisi mother^ was killed tj

IMPi.

(239) is ungrammatical with the coindexing given, i.e. it cannot be interpreted as a statement about matricide. The purpose of this section is to provide an account of this restriction on the interpretation of passives. The account that we gave of strong crossover in the previous section in terms of chain-formation will not carry over to weak crossover because in weak crossover there is no c-command between the pronoun and the variable and therefore no question of forming a chain. However, according

144

The Representation of Implicit and Dethematized

Subjects

to Koopmann & Sportiche (1982) (cf. 1.4.), it is precisely because of this lack of c-command that the Bijection Principle rules out standard whmovement cases, as there are two distinct variables which do not c-command each other bound by a single operator (for Koopmann & Sportiche this follows from their definition of variable). However, it is not obvious how to rule out the NP-movement case. For (239) to be ruled out by the Bijection Principle we would need an operator-variable structure, a kind of structure which is not obviously present in passives. We have been assuming, however, that agentless passives at least contain a variable. We proposed in 2.2.4. that this variable satisfies the Bijection Principle at SS by A'-binding by en and at LF by existential quantification over the implicit argument. The quantification arises via a mechanism of existential closure, so in that case passives do in fact involve an operator-variable structure, as follows (cf. 2.2.4. (128)): (240)

Hxi [hiss mother]j was kill eni tj ti.

This LF configuration can now be ruled out by the Bijection Principle exactly like (239). If the existential in (240) is present as a function of properties of the implicit argument, we predict that " N P movement" weak crossover only appears with passives. This is true, as the grammaticality of otherwise comparable raising examples shows: (241)

[Hisj mother]j seems to Johns [ tj .... ]

Our accounts of strong and weak crossover interact to produce this result, as we rule out cases of strong crossover by the θ-criterion and the Bijection Principle together, and cases of weak crossover by the Bijection Principle alone. Hence for weak crossover to arise there must be an operator-variable structure. Such a structure is present in passives but absent in raising constructions. 19 In order for our account of weak crossover in passives to go through, certain assumptions need to be made about the representation of implicit arguments: (i) that they are structurally represented, and (ii) that they are represented at LF by an existential quantifier which undergoes QR to form an operator-variable structure of the kind seen in (240). The first assumption is motivated independently in terms of our account of modification and predication in 3.2. The motivation for the second comes from semantic properties of implicit arguments. The next sections deal with the latter topic in an attempt to independently motivate the above account of weak crossover.

The By-Phrase and Implicit Arguments

145

3.3.3.2. Implicit Arguments as Quantifiers In this section we will motivate the claim that IMP is interpreted as a quantifier, in fact an existential. We refer to this quantifier henceforth as the "implicit existential." There are three main ways to show that something is a quantifier: one can show that it is interpreted quantificationally; one can show scope ambiguities between it and negation or other quantifiers; and one can show that the assumption that it undergoes QR in the mapping to LF has desirable consequences, e.g. regarding weak crossover. In all three respects, IMP acts like an existential quantifier. We have already suggested that weak crossover may occur in passives; we will return to this point at the end of the next section after reviewing the other two quantifier-like properties of IMP. So we now consider the issues involved in positing the existence of an implicit existential. First, it is intuitively clear that the agent of an agentless passive is interpreted as an existential quantifier. In other words, (242) has the interpretation in (243): (242)

Mary was kissed.

(243)

a. b.

Mary was kissed by someone, Someone kissed Mary.

Chomsky (1966:53) formulates the passive transformation as deriving (243a) from (243b) with subsequent optional deletion of by someone. Also, Bresnan (1982a) says: "The predicate argument to which 0 [i.e. inability to be realized as a grammatical relation - IGR] is assigned is existentially interpreted" (38). She points out that this existential always has narrow scope, a point we return to below. And Keenan (1980) gives the following schema for "the relation between passive VPs and the TVPs [transitive VPs - IGR] they are derived from:" (244)

(Pass, tvp) (x) = (3y) (tvp(x)) (y).

(244) is "that VP function which assigns to each individual χ the truth value represented by 'for some y, y kissed x" (12). The idea common to these accounts is that the position in the argument structure of the predicate which is not specified by an overt Z^-phrase is given a referential value by means of existential quantification. Where this happens by means of conversion of a free variable into a variable bound by the existential at the relevant level (for us, LF), the process is known as "existential closure."

146

The Representation of Implicit and Dethematized

Subjects

We adopt the "existential closure" approach here and examine its consequences in the context of the general approach that we have taken to passives. So we begin by assuming an LF for (242) as in (245): (245)

3xj Maryi was kissedj ti Xj

We will now see to what extent the existential quantifier in (245) behaves like other existentials: does it show scope ambiguities and does it undergo QR in the mapping to LF? Taking the question about scope ambiguity first, consider the following example: (246)

Every man was killed.

Given our assumptions about LF (cf. 1.4.), if an existential is present we expect either order of quantifiers to be possible. This means that we expect (246) to be true whether there is one person who killed every man (wide-scope existential) or whether a different person killed every man (wide-scope universal). In fact, both readings do seem to be available for (246). However, this does not constitute evidence for the implicit existential, because the wide-scope existential reading is entailed by the narrow-scope reading; the situations in which a sentence with a wide-scope existential is true are a proper subset of those in which a sentence with a widescope universal is true. In other words, the availability of both scope readings for (246) does not show that this sentence need be formally ambiguous in the sense of having two LF representations. Both of the readings available for (246) arise if every man has wide scope. So the ambiguity of (246) is not evidence for the presence of an implicit existential which undergoes QR in the usual way. Another test for the scope properties of a putative existential like IMP is to see how it interacts with negation. Wide-scope existentials allow existential generalization, i.e. the truth of a sentence containing an existential outside the scope of negation entails the existence of someone or something as the value assigned to the variable it binds. No such entailment holds for existentials inside the scope of negation. Consider then (247): (247)

John wasn't killed.

This sentence cannot mean that there exists someone who did not kill John. It can only mean that there is no-one who killed John. The lack of the first reading means that the existential cannot be outside the scope of the negation here. Compare this with (248):

The By -Phrase and Implicit (248)

Arguments

147

John didn't kill someone.

(248) clearly can have the meaning that there exists someone who John did not kill. This contrast casts doubt on the hypothesis that the implicit quantifier is an existential quantifier in anything like the usual sense. However, a closer look at negation suggests that this conclusion might be too hasty. A standard assumption is that clausal negation appears in Infi. This assumption has some independent syntactic motivation (from the fact that not always appears following the first auxiliary and is inverted with the first auxiliary) and is consistent with the assumption that scope domains are determined by c-command: if not appears in Infi it c-commands the rest of the clause and so can negate whichever part of the clause does not QR. However, not can also appear in NP and in VP: (249)

a. b. c.

[NP Not a single person] arrived. He has twice [VP not opened the door]. The men have all [VP not left].

In these examples not has scope only over the maximal projection inside which it appears; it does not have clausal scope. So (250a) but not (250b) is a possible reading for (249a), (251a) but not (251b) is the reading for (249b), and (252a) but not (252b) is the reading for (249c): (250)

a. b.

No person arrived. A person didn't arrive.

(251)

a. b.

There were two events of his not opening the door, There weren't two events of his opening the door.

(252)

a. b.

None of the men have left. Some of the men have left (and some haven't).

So in these examples not has scope only over the VP which contains it (which in (249b,c) is not in fact VP/S, as have adds a further VP cf. the discussion of auxiliaries in 2.2.1.3.). We have seen that the existential in (247) cannot have wider scope than clausal negation. Can it have wider scope than VP negation? The answer seems to be that it can. To show this, we need examples where the negative is embedded in a sequence of auxiliaries, and hence unambiguously in VP. The problem with such examples is that the best cases involve modals, as negation is harder to embed in sequences of aspectuals, and, since modals create opaque contexts, existential generalization cannot be used as a test.

148

The Representation of Implicit and Dethematized

Subjects

Another possibility is to find an operator that is restricted to VP in scope, show that it can have wider scope than negation and then show that the existential can have wider scope than the negation. This would show that the lack of existential generalization in examples like (247) is not the consequence of the absence of IMP but of some restriction on IMP's scope domain. Two operators whose scope may be restricted to VP are twice and floated all. Compare the following examples with (249b,c): (253)

a. b.

John hasn't twice opened the door, The men haven't all left.

In these examples the order of negation and the operator is the reverse of (249b,c), and the scope order is reversed as well. (253a) means (251b) but not (251a), and (253b) means (252b) but not (252a). The scope differences between (249b,c) and (253) can be accounted for in terms of c-command if we assume that twice and floated all are in VP in both cases, and not is in VP in (249b,c) but in Infi in (253). Further evidence that these elements appear in VP comes from examples with modals like (254): (254)

a. b.

John must have twice opened the door, The men must have all left.

Here the modal occupies Infi, have heads a VP and so twice and all must be inside at least the VP headed by have (or perhaps adjoined to this VP, since all cannot appear to the right of the main Verb - cf. the discussion of bare each in 3.3.1.). Another piece of evidence for the view that floated quantifiers are restricted to VP scope comes from the behavior of bare each. In general, each requires wider scope than other operators. If floated each has VP scope, and therefore narrower scope than clausal negation, this requirement cannot be met. In fact, examples with floated each inside clausal negation are not very good. Compare: (255)

a. The men each didn't leave, b. *The men didn't each leave.

So twice and floated all are restricted in scope to the VP they appear in, and, as (249b,c) show, they can have scope over negation. So we can now test for the presence of an existential in passives by seeing whether the implicit existential has wider scope than a negation in the scope of one of these elements:

The By-Phrase and Implicit (256)

a. b.

Arguments

149

The door has twice not been opened, The papers have all not been graded.

In each of these examples, existential generalization is possible. So (256a) can mean "there's someone who did not open the door twice," and (256b) can mean "there's someone who didn't grade all the papers." Moreover, the existential in these examples seems to be able to have wider scope than the VP operator, i.e. (256a) can be true where the same person failed to open the door on two occasions, or where it was a different person the two occasions, and (256b) can be true where the same person failed to grade all the papers or where different people failed to grade different papers (although the former reading is preferred). These facts indicate that there is in fact an implicit existential in VP, but its scope domain is limited to VP. Another piece of evidence that there is an implicit existential with a scope domain limited to VP comes from the interaction of negation and certain types of adjunct. Examples like (257a) are three-ways ambiguous, the readings are given in (257b - d): (257)

a. b. c. d.

John NOT NOT John

didn't kill his wife for this reason. [John killed his wife for this reason]. [John killed his wife] for this reason. killed his wife NOT [for this reason].

These readings have distinct entailments: (257c), but not the others, entails the existence of a reason for John's (refraining from) action, and (257d), but not the others, entails that John's wife is dead. So negation can have scope over both the main predicate and the adjunct. These facts indicate that our simple theory of scope of negation as determined essentially by its SS c-domain must be elaborated. We propose two things to account for the readings in (257): (i) that certain adjuncts are optionally basegenerated under CP but outside IP, and therefore outside the c-domain of a negation in Infi; (ii) that there exists an LF rule of "neg-attraction," whereby a negation adjoins to a secondary predicate, and is thus interpreted as negating only that predicate and not the main one. Assumption (i) is relevant for the wide-scope adjunct reading (257c) and assumption (ii) for (257d). The implicit existential interacts with negation here in that existential generalization is possible only when the adjunct alone is negated. Cf.: (258)

John's wife wasn't killed for this reason.

This example entails the existence of a killer of John's wife only on the

150

The Representation of Implicit and Dethematized

Subjects

reading where the adjunct is negated, i.e. where it means "Somebody killed John's wife but not for this reason." So existential generalization is only compatible with the entailment that John's wife is dead. The existential interacts with another fact about sentences like (256). Extraction of the adjunct forces the reading where only the main predicate is negated (this was pointed out to me by Luigi Rizzi, p.c.). In other words, (259a) can only be interpreted as a question about the fact that John did not kill his wife, and the quantifier in (259b) must have wide scope, i.e. (259b) corresponds only (257c): (259)

a. b.

Why didn't John kill his wife? John didn't kill his wife for many reasons.

These examples imply the existence of reasons for John's actions, and do not imply that John's wife is dead. Since the adjunct cannot be negated in negative questions or when quantified, we predict that existential generalization is not possible in the passives of (259): (260)

a. b.

Why wasn't John's wife killed? John's wife wasn't killed for many reasons.

This seems to be true; (260a) cannot be interpreted parallel to (258), i.e. it means (261a), not (261b), and (260b) means (261c), not (26Id): (261)

a. b. c. d.

For which reason did no-one kill John's wife? For which reason did someone not kill John's wife? There are many reasons for which no-one killed John's wife. There are many reasons for which someone didn't kill John's wife.

We will return to these facts in the next section. For now, the interaction of the implicit existential with other operators of various kinds indicates that it is correct to assume that the implicit existential is structurally present. It is not obvious how one would explain the facts above, particularly those in (239) and (255), otherwise. The implicit existential is semantically like a quantifier, in allowing existential generalization, interacts with other operators in VP, as (255) shows, and interacts like other quantifiers as predicted by the principles of the theory in producing weak crossover effects. So we conclude that there is an implicit existential in the LF representation of passives, i.e. that IMP is bound under existential closure. We now have to explain the scope restrictions on this quantifier. Formulating a proposal for this is the main topic of the next section.

The By-Phrase and Implicit

Arguments

151

3.3.3.3. Implicit Arguments and Extraction We have so far seen that implicit arguments have the following properties at LF: (262)

a. b.

Implicit arguments are interpreted as existential quantifiers, The implicit existential that binds the variable in the implicitargument position has a scope domain restricted to VP.

Assuming the correctness of (262a), we will now investigate what underlies (262b). The first point to consider is whether overt quantifiers inside ^-phrases can have wider scope than subjects or sentential negation, etc. In general, this seems to be possible, as the ambiguity of examples like (263) shows: (263)

a. b. c.

Someone was seen by everyone. Two languages are spoken by everyone in this room. Two women are loved by many men.

Although there is the usual preference for the relative scope of the quantifiers to be the same as their relative order, these examples allow a wide-scope reading for the quantifier in the £j>-phrase. In other words, different people can be seen in (263a), different languages spoken in (263b) and different women loved in (263c), as a function of the value of the quantifier in the Z^-phiase. These examples seem to be as clearly ambiguous as their active counterparts. On the other hand, sentence pairs such as the following, originally noticed by Lakoff (1965), and discussed by Jackendoff (1972) among many others, show that a quantifier inside the 6_y-phrase must remain within the scope of sentential negation: (264)

a. b.

Many arrows didn't hit the target. The target wasn't hit by many arrows.

This pair is said to differ in that the first sentence has either scope order for not and many, while the second allows only wide scope not. Other relevant examples are given in (265): (265)

a. ?John wasn't killed by someone, b. John wasn't killed by everyone.

(265a) is of somewhat questionable grammaticality, but it clearly can only mean "there is no-one such that John was killed by him." (265b), similarly,

152

The Representation of Implicit and Dethematized

Subjects

only allows the narrow-scope universal reading and hence cannot mean "there is no-one who killed John," but rather means "not everyone killed John." The same is true in other cases. Consider for example the negative versions of (263b,c): (266)

a. b.

Two languages are not spoken by everyone in this room, Two women are not loved by many men.

These examples do not allow the quantifier in the 6^-phrase to have wider scope than the subject. So (266a) is true only where the same two languages are unknown to everyone in the room, and (266b) can only be true where the same two women are not the object of men's affections. So the conclusion is that quantifiers in 67-phrases are limited in scope in negative sentences. This recalls the result of the previous section, where we saw that the implicit existential is limited in scope to VP in negative sentences. The absence of evidence that the implicit existential is limited to VP in positive sentences leads us to the suggestion that there is a complete parallelism between overt and implicit quantifiers in Z>>>-phrases. In fact, in (266) the relative scope of negation itself and the quantifiers is freer than in (264) and (265). Here the natural reading is to take the ^ - p h r a s e quantifier as wider in scope than negation. So (266a), for example, is naturally interpreted as "There are two languages that no-one in this room speaks," although it can also mean "There are two languages that some people in this room speak." 20 To summarize so far, then, we have observed the following facts about quantification in ^-phrases: (267)

a. b.

c.

ity-phrase quantifiers can have wider scope than the subject in positive sentences. 5^-phrase quantifiers must have narrower scope than a quantified subject in negative sentences (but may have wider scope than negation). Zty-phrase quantifiers must have narrower scope than negation in negative sentences with no quantified subject.

To a certain extent, implicit existentials behave in the same way. We cannot test for scope properties of the implicit existential in positive sentences for reasons discussed in the previous section. We saw in the previous section that the unavailability of existential generalization in negative agentless passives shows that the existential must have narrower scope than sentential negation. The implicit existential also has narrower scope than negation where there is a quantified subject, as the failure of existential generalization in (268) shows:

The By-Phrase and Implicit (268)

Arguments

153

Two men weren't killed.

We will return to this difference below. For the moment what interests us is the parallel between overt and implicit quantification in i^-phrases. To pursue this parallel further, the next thing to test is the relative scope of a floated quantifier in VP and a quantifier in the ^ - p h r a s e . In positive sentences there is a straightforward ambiguity: (269)

The papers have all been graded by two men.

This sentence can be true if each paper was graded by a different pair of men, or the same two men graded all the papers. Compare now the following two sentences: (270)

a. b.

The papers have all not been graded by two men. The papers have not all been graded by two men.

These two examples differ, like (253) and (256) of the previous section, in both the surface order and the relative scope of all and not. Aside from this difference, however, it appears that two men can vary freely in scope with respect to all. The possible relative readings of these quantifiers are as in (269). Again overt quantifiers in the by-phrase parallel the implicit existential. Consider next cases where negation is ambiguous between applying to an adjunct, to a main predicate, or to both. Does scope of negation interact with the potential scope of a quantified NP in the ό^-phrase? Is wide scope for the 6^-phrase quantifier possible only when the adjunct alone is negated? Cf.: (271)

Two women were kissed by every man for this reason.

The positive sentence is again perfectly ambiguous; there could either be a different pair of women for each man or the same pair. Compare (271) with its negated counterpart: (272)

Two women weren't kissed by every man for this reason.

As in (256), the negation is three-ways ambiguous. The readings are illustrated in (273): (273)

a. b. c.

NOT[two women were kissed by every man for this reason]. Two women were kissed by every man NOT[for this reason]. NOT[two women were kissed by every man] for this reason.

154

The Representation of Implicit and Dethematized

Subjects

The reading in (273c) entails that a reason exists for the lack of action on the part of the men; (273b) entails that the women were kissed, and (273a) has neither of these entailments. The judgements concerning the scope possibilities of the quantifiers in these examples are rather difficult. It seems fairly clear that in (273a), every man cannot have wide scope. This is exactly analogous to the judgement of (265b) and (266) above. On the other hand, (273b) fairly clearly allows both scope readings for the quantifiers, like a positive sentence. (273c) is the difficult case; my intuition is that it does not allow wide-scope for every man, but this is far from crystal clear. We saw in the previous section that the scope possibilities of the implicit existential interact with extraction of the adjunct, since this extraction requires negation to have scope over only the main predicate. We thus predict, parallel with (259), that wide scope for a 6y-phrase quantifier is not possible where the adjunct has been extracted in a negative. Cf.: (274)

Why weren't two women kissed by every man?

Here the prediction seems clearly to fail; every man can have scope over two women. In fact, this seems to be the preferred reading (although the other reading is also possible). Notice also that every man must be inside the scope of negation here; the sentence does not mean that nobody kissed two women. The parallel example to (259b), with a quantified adjunct, differs, however. It seems that a sentence like Two women weren't kissed by every man for many reasons has only the wide-scope adjunct interpretation, with every man narrower in scope than both negation and the subject. So, aside from (267b) and (274), similar constraints on quantification hold in passive VPs where there is a 67-phrase containing an overt quantifier and where there is only the implicit argument. This parallelism supports the idea that the implicit argument is quantificational, but only in a limited way. It is clear that ^ - p h r a s e quantifiers have quite peculiar scope properties. We need to account for these properties for two reasons: first, to shed further light on the implicit existential, and, second, to reconcile the above data with the account of "NP-movement weak crossover" that we proposed in the previous section, since that account relies on a representation like (240) where the implicit existential has scope over IP. There is one well-known set of cases where negation interacts with extraction. This is the phenomenon of "inner islands," discussed by Ross (1967,1984). Ross noticed contrasts like the following: (275)

a. b.

How did you solve the problem? *How didn't you solve the problem?

The By-Phrase and Implicit

Arguments

(276)

a. b.

What did no book discuss? *How much did no book cost?

(277)

a. b.

How did you say [we used to get there]? *How did you say [we never got there t]?

155

The generalization underlying these examples seems to be the following (pointed out by Rizzi (class, Spring 1985)): (278)

For a certain class of adjuncts A, no member of A may bind a variable in the scope of a negation.

The generalization in (278) accounts for the contrasts in (275 - 277); in each of the (b) examples the trace of an adjunct is in the scope of the negation. Note that this restriction does not apply to argument traces, as (276a) shows. (278) also accounts for the interaction of scope of negation and extraction discussed in (260) and (261) of the previous section. If we adopt the generalization in (278) and assume ^ - p h r a s e s are (structural) adjuncts we can account for the scope restrictions on both overt and implicit fry-phrase quantifiers. 21 We interpret (278) as allowing QR to take place within the scope of negation, what is ruled out is QR from a position inside the scope of negation to a position outside it. Construed this way, (278) explains the basic constraint on the scope of overt ^ - p h r a s e quantifiers in negative sentences of (267c) and the similar behavior of the implicit existential which underlies the absence of existential generalization in negated agentless passives. Instead of being an argument against the implicit quantifier, the absence of existential generalization in negated agentless passives now follows from (278). Since (278) is a general constraint on extraction, including QR, the fact that the implicit existential obeys it is evidence for the existence of this quantifier. Some of the other data discussed above appear to pose problems for (278). In (270), for example, the öy-phrase quantifier is in the scope of the negation, and yet is able to be extracted. For (270a), which has the order not all, we propose that the scope of the negation is limited to the quantifier itself, hence the fry-phrase is not in the scope of negation and so extraction is allowed. This account can be extended to (270b), only here the negation has narrower scope than the quantifier. The crucial point about both cases is that the scope domain of the negation is just the floated quantifier phrase. Second, while (278) explains a number of aspects of the interaction of extraction, negation and quantifier scope, for example (256, 272, 273), it does not explain the interaction of overt Z>j-phrase quantifiers and quantified subjects summarized in (267b), or the extraction example (274).

156

The Representation of Implicit and Dethematized

Subjects

In (274), the negative appears in the inverted auxiliary. We assume the auxiliary appears in a position attached to IP. Hence QR of the &>"-phrase quantifier to a position inside the scope of negation but wider in scope than the subject is possible. This captures the crucial fact that everyone is inside the scope of negation in (274) and the contrastive judgement with respect to the QR example parallel to (259b). We have no explanation for (267b), however. This example involves a kind of "connectedness effect" (see Kayne (1984)), in that a licit extraction licenses an otherwise illicit extraction. The details in this case are far from obvious, however. Since this is one respect (the only really unambiguous one) where overt and implicit quantifiers differ, whatever connectedness effect there is clearly does not extend to the implicit quantifier. So our account of scope restrictions on 6^-phrase quantifiers rests on the generalization about inner islands in (278). We will not attempt to investigate here what underlies this generalization (although we will return to the topic of negative polarity in Chapter Four). Adopting (278) allows us to make a further prediction concerning the scope possibilities of fryphrase quantifiers: negatively quantified subjects should always have wider scope than a 67-phrase quantifier. This is shown for adjunct extraction in (276b). Presumably the LF-representation for such a sentence has the following form: (279)

[CP how muchi (did) [ip no bookj [IP tj cost tj ]]].

Here the adjunct variable f, is c-commanded by the negative quantifier, and therefore in its scope, this illicit extraction thus falls under (278). The same seems to be true for quantified fry-phrases. Compare the following examples with (263) and (266): (280)

a. b. c.

Only two languages are spoken by everyone in this room. Two languages alone are spoken by everyone in this room. Few languages are spoken by everyone in this room.

In all these examples, the group of known languages is constant, it does not vary according to the value we choose for everyone. So once again, the 6y-phrase quantifier obeys the inner-island constraint in (278). The same observation was made by Keenan (1979:15), who noticed that the following pairs of examples are non-equivalent: (281)

a. b.

No politician is admired by each student, Each student admires no politician.

(282)

a. b.

Few pupils were kissed by everyone, Everyone kissed few pupils.

The By-Phrase and Implicit

Arguments

157

(281a) can be true where the students disagree about the class of admirable politicians, although individual students admire certain politicians, while (281b) can only be true where not one of the students finds anything admirable in any politician. In both examples, then, the scope order follows the surface order; in (281a) this is because of (278), and in (281b) it is because each always has wide scope when it can. In (282a), the same few pupils were kissed by everyone, while in (282b) it can either be a different group or the same group. Once again, (278) accounts for the impossibility of wide-scope everyone in (282a). The implicit existential also has narrower scope than a negatively quantified subject, as the failure of existential generalization in (283) shows: (283)

a. b. c.

Nobody was killed. Few people were seen. Only two of us were arrested.

As with the other failures of existential generalization, this is accounted for by (278): the implicit existential cannot raise to a position outside the scope of the negative quantifier, leaving an adjunct variable in a position c-commanded by that quantifier. It appears, then, that (278) has some explanatory power with regard to the behavior of quantifiers in 6y-phrases. A consideration of the interaction of Class I Adverbs and RatCs with the implicit existential gives further evidence for (278), and interacts with our proposals for the syntax of modification in 3.2. The first relevant example is the negation of the active with a modifier of one of these types: (284)

John didn't kill Bill deliberately/to collect the insurance.

(284) is ambiguous: on one reading both the main predicate and the modifier are negated, so it does not entail that Bill is dead; on the other only the Adverb is negated, so it entails that Bill is dead (and John killed him but not in the manner specified in the adjunct). Unlike examples like (256) and (272), these kinds of adverbs do not allow main-clause negation only, i.e. (284) cannot mean "John deliberately refrained from killing Bill." Given (278), the fact that the modifier is always in the scope of negation means that we should not be able to question it. This is true, as (285) shows: (285) *How didn't John kill Bill? (Cf. fn. 7 on the other extraction properties of these Class I Adverbs and RatCs). Our proposals for the modification properties of Class I Adverbs and RatC allow a straightforward account of this restriction. Only CP-

158

The Representation of Implicit and Dethematized

Subjects

modifiers can be free of the scope of clausal negation, and these modifiers cannot be construed as CP-modifiers, as they select an Agent (usually the subject) and an event (see 3.2.). Because of their selectional requirements, then, they must R-modify predicates and arguments, which means that the highest position in which they can appear is one adjoined to IP, and therefore in the c-domain of Infi. Hence Class I Adverbs and RatCs cannot be free of clausal negation. On the other hand, a Class I Adverb or RatC can trigger neg-attraction, making the main predicate free of negation. This structure gives rise to the reading of (284) which entails that Bill is dead. Following (278), then, we predict that the existential can have wide scope. Consider: (286)

a. b.

John wasn't deliberately killed. John wasn't killed in order to collect the insurance.

Both of these examples allow the reading where there exists a killer for John. Hence the existential has wider scope than the negation. The same is true with an overt quantifier. So compare (286) with: (287)

a. b.

Two men weren't voluntarily killed by everyone. John wasn't killed by everyone to collect the insurance.

The same seems to be true of the examples in (287): this is more clearly seen with (287b), which can mean both "not everyone killed John to collect the insurance" (i.e. some people did it because they don't like him) and "no-one killed John to collect the insurance" (they all did it out of some other motive). The former reading corresponds to wide-scope clausal negation, the latter to adjunct negation and wide-scope everyone. On the other hand, if the Adverb appears in pre-aux position it cannot be negated, as the readings for (288) show (we leave RatC out of consideration here, as RatC are less readily transportable than Adverbs, probably because they are heavier): (288)

John voluntarily didn't kill Bill.

In (288) either the whole clause is negated, or just the VP. So this example, unlike (284), can mean that John voluntarily refrained from killing Bill, and cannot mean that John did not voluntarily kill Bill. In this example, then, we must assume that neg-attraction optionally adjoins negation to VP, and cannot adjoin it to AP, while the opposite possibilities obtain for neg-attraction in (284). Although it is unclear what constraints on neg-attraction bring this situation about, the prediction about quantifier scope, given (278), is clear: Z^-phrase quantifiers are unable to have wide scope. The prediction is confirmed in (289):

The By-Phrase and Implicit (289)

a. b.

Arguments

159

John voluntarily wasn't killed. John voluntarily wasn't killed by everyone.

It seems clear that (289a) cannot mean that John has a particular killer who voluntarily didn't kill him, neither does it entail that John was killed. Similarly, (289b) only has the reading where everyone has narrower scope than negation, and there is no entailment as to whether John was killed. These examples both have the reading "It was voluntary of John that it was not the case that someone/everyone killed him" (the Adverb must be understand as modifying the subject here - cf. 3.2.1.1.). (290) is an intermediate case between (286) and (289): (290)

John hasn't intentionally been killed.

In this case, the Adverb can either be a daughter of VP or of IP. When a daughter of VP, the Adverb modifies the implicit argument and the implicit existential can have clausal scope, as in (286). If the Adverb is a daughter of IP, and so modifies John, the existential is limited to narrow scope, as in (289). Our assumptions about neg-attraction and the independently motivated generalization in (278) account for these facts. Notice further that other kinds of modifiers do not permit any kind of neg-attraction, i.e. negation only has scope over both the main predicate and the modifier. Cf.: (291)

John didn't file the book before reading it.

We cannot understand (291) to mean either "John filed the book, but not before reading it," or "Before reading the book, John didn't file the book." So (278) predicts that this adjunct cannot be questioned, and that the agentless passive of (291) will not give rise to existential generalization. Cf.: (292)

a. b.

*When didn't John file the book? John wasn't killed before checking his insurance status.

Again the prediction seems to be correct. We conclude from all of this that there is an existential quantifier in agentless passives, and that this existential is subject to essentially the same constraints as an overt quantifier, i.e. (278). (278), while it remains mysterious, has considerable explanatory power in various domains, the behavior of the implicit argument being just one of them. We must now return to the question which led us to initially posit the implicit existential, namely the phenomenon of " N P movement weak crossover," illustrated again in (293):

160

The Representation of Implicit and Dethematized

Subjects

(293) *[Hisi motherjj was killed tj ti. We can retain the account for this case that we gave earlier, since we have found no evidence to refute the claim that agentless passives involve QR of an existential. So (293) has the LF-representation of (294), a representation which violates the Bijection Principle: (294)

3xj [hiss motherjj was kill+eni tj Xi.

However, this account will not extend to the negation of (293), since (278) prevents extraction of the existential from outside the scope of the negation. For these examples we must find an alternative account. So this part of the NP-movement weak crossover phenomenon remains to be explained. Leaving this problem, then, we conclude this section. Here we have seen that apparent counterevidence to the proposal that agentless passives involve an implicit existential can be accounted for in terms of an independently motivated, if rather mysterious, principle relating extraction and quantifier scope to scope of negation - this principle is Rizzi's explanation for the inner-island phenomenon discovered by Ross. This principle also affords an account of at least a significant number of properties of overt 6y-phrase quantifiers, and interacts in a way we would predict with Class I Adverbs and RatC. Finally, since we can maintain the idea of an implicit existential, we can subsume "NP-movement weak crossover" under the standard, bijection-based account of weak crossover. The one outstanding problem is then to account for NP-movement weak crossover in negatives. 3.3.4. Binding, Coreference and Implicit

Arguments

In this section, we will review the binding and coreference properties of IMPs. Two issues are involved here: first, we still need to account for NP-movement weak crossover in negative clauses, and, second, we need to explain how it is that IMPs are somewhat "syntactically inert" with regard to binding and coreference. The latter fact has led to the conclusion (e.g. in Chomsky (1986)) that IMPs are not structurally present. Since we are now committed to the view that IMPs are structurally represented, we have to explain their "inertness." In this section we will see that their lack of binding properties can be straightforwardly explained: it is their marginal ability to bind which then requires explanation. This explanation intimately involves the notion of arbitrary reference, a notion which we will attempt to elucidate a little here. The discussion of binding centers on anaphors, i.e. Principle A, although we will briefly consider Principles Β and the residue of Principle C.

The By -Phrase and Implicit

Arguments

161

Implicit arguments are claimed to be unable to bind anaphors, e.g. in Chomsky (1986). Relevant examples are given in (295): (295)

a. *Testimony was given about himself. b. ?*Letters were sent to themselves. c. ??The blankets were put on top of each other. d. ??Obscene photographs were shown to each other.

These examples show that the empty ^ - p h r a s e is at best somewhat awkward when binding anaphors, and in many cases quite impossible. We account for the basic impossibility of binding by IMPs in terms of the agreement condition on binding relations proposed in 2.2.4.(118), repeated here: (296)

If X binds Y, then X and Y must not conflict in features.

In fact, the version relevant for anaphor binding (but not necessarily chainformation - see 2.2.4.) is stronger, being of the form (297): (297)

If X binds Y, then X and Y agree in features.

The stronger condition would follow automatically from the suggestion made earlier that anaphors have to be identified in featural content by their antecedents at LF (see 3.3.1.), as in that case anaphors would by definition have the features of their antecedents. Given (297), we can explain the inability of the IMP to bind by saying that it lacks grammatical features. Taking the set of grammatical features relevant to (297) to be person, number and gender features, this seems to be true, as IMP seems to lack any specification for person, number or gender. This account is also consistent with the fact that an NP in an overt ^ - p h r a s e can bind (as shown by a simple contrast like that between (295a) and Testimony was given about himself by the suspect). These NPs will have person, number and gender features. This account correctly rules out examples like (295a), but it also rules out all the other examples in (295). This is a little too strong, as plural anaphors are more readily bound by IMP than singular ones, and each other more readily than themselves. How can we account for these contrasts? One way to put the distinction between each other and themselves, on the one hand, and other anaphors, on the other hand, would be to say that each other is and themselves may be plural indefinites, and as such involve quantification. Singular anaphors, with the exception of oneself, which we discuss directly, are neither plural nor indefinite, and plural anaphors are generally definite (although cf. the remarks on ourselves below;

162

The Representation of Implicit and Dethematized

Subjects

also, to the extent that yourself/selves functions like oneself it is indefinite). So we will retain (297) as the basic constraint barring binding by IMPs, and try to explain the improvement with indefinite plurals in terms of other principles. The idea which immediately suggests itself is to introduce a feature [ ± definite], and make this feature subject to (297). IMP, oneself, each other and, marginally, themselves are [-definite], while other anaphors are [+definite]. This is problematic as it predicts that definite anaphors cannot have indefinite NPs as antecedents, which they clearly can (e.g. a man saw himself in the mirror etc.). However, the idea that the notional indefiniteness of both antecedent and anaphor is relevant is important. Indefinite plural pronouns are closely related to arbitrary pronouns, so an account of these indefinite anaphors ought to provide the basis for a general account of arbitrary reference, something we will sketch without going into in detail. Our approach will be based on the idea that indefinite plural pronouns are quantificational in that they undergo QR in the mapping to LF. If indefinite anaphors can be bound by IMPs, then we expect oneself to be allowed in examples like (295). This is true, as the following examples from Rizzi (class, Fall 1984)) show: (298)

a. b.

Such privileges should be kept to oneself. Damaging testimony is always given about oneself in secret trials.

Oneself also appears bound by the arbitrary pronoun one and by arbitrary PRO, as the following examples illustrate: (299)

a. b. c.

One; always likes [PROi to admire oneself]. One, often sends letters to oneself. PRO, to write letters to oneself is the sign of a well-adjusted personality.

So, to understand the nature of oneself, we must first understand the conditions which license one (we leave aside arbitrary PRO here - cf. Chapter One, fn. 5, 3.2.4.1.2. and fn. 24 below). Three aspects of one are of interest: the fact that it is quantificational, the distributional constraints it is subject to and its behavior with respect to Principle B. The evidence that one is quantificational is of a standard kind: it cannot corefer with another pronoun and it gives rise to weak crossover effects. These properties are illustrated in (300) and (301): (300)

a. *Onei often goes to parties and hei has fun. b. Onej often thinks he, is being followed.

The By -Phrase and Implicit Arguments (301)

163

a. *Hiss mother should always say ones is good, b. Onej should always say hisi mother is good.

We take these contrasts as evidence that one is a quantifier. The relevance of this conclusion will become clear below. Turning to the distributional constraints on one, the first thing to note is that in general, one can only appear in subject position (although there is some dialectal variation): (302)

a. *They see one often. b. *People send books to one. c. *It is inconvenient for everyone for people like that to live near one. d. One is often sent books. e. One likes to visit one's friends. f. It seems that one is always pressed for time. g. The book is easy for one to read.

Moreover, even in subject position, one must appear with a modal or stative simple present tense: (303)

a. ??One went to the bank last Friday, b. ??One is going to the bank tomorrow.

This is an SS requirement, as (304) shows: (304)

a. b.

One is always asked for money at the airport, One seems to have so little time.

These facts suggest the following constraint: (305)

One must be governed by a modal or by a stative Tense.

Third, one, as a pronoun, is subject to Principle B. Cf. the following contrast: (306)

a. One is always persuaded that one is the best, b. *One always believes one to be the best.

The striking thing about (306b) is that there is no way to interpret the second occurrence of one as disjoint from the first. This is not usually the case with pronouns; an example like (307) always allows a disjoint interpretation for the second pronoun, in which case Principle Β is satisfied:

164

The Representation of Implicit and Dethematized

(307)

He believes him to be the best.

Subjects

In fact, this behavior with regard to Principle Β is explained by the constraint in (305): one can only appear in positions other than those licensed by (305) - e.g. subject of NP as in (302e) - if coindexed with another occurrence of one in a position licensed by (305), and not in violation of Principle B. This possibility is illustrated in the contrast that follows: (308)

a. One often reads books about one's problems. b. *One often reads books about one. c. *John often reads books about one.

(308a) is grammatical, as the one in subject position is licensed following (305) and the one in the PP can be coindexed with it without violating Principle B. In (308b), if the second one is coindexed with the first it violates Principle B, and if it is not it violates (305). In (308c), one violates (305). In a similar way, the anaphor oneself can appear as long as its antecedent meets condition (305) and subject to Principle A:22 (309)

a. One often reads books about oneself. b. *One often believes people to like oneself. c. *One is going to shave oneself.

The latter observation extends to the case where the IMP is the antecedent, as the contrast between (298) and (310) shows: (310)

a. *It is expected that John will talk to oneself, b. *Letters are being sent to oneself.

It is possible to give a unified account of these phenomena which sheds light on arbitrary reference in general and accounts for the ability of IMP to bind indefinite anaphors. First, we make the natural supposition that one undergoes QR in the mapping to LF. This assumption is necessary to account for the weak crossover effect shown in (301). Second, we understand the condition in (305) to restrict the positions which one can QR from. In a sense, then, we consider one to be a kind of semantically empty quantifier that must receive its content, i.e. have its domain of quantification identified, under government by a modal or stative Tense. After QR, one forms a chain with the variable in the position it vacated. If one is coindexed with some other element X, which the variable ccommands, then we have a structure like the following: (311)

onej ... ti ... Xi

The By-Phrase and Implicit

Arguments

165

(300) shows that Xi is necessarily bound by one. This is true whether Xi is another occurrence of one or of oneself (on definite pronouns bound by one see below). Xi must function as a variable bound by one,. Our third proposal is that the coindexed elements in (311) form a chain. This cannot be a θ-chain, as it would violate the θ-criterion, but it is a wellformed variable-chain in the sense that it meets the Bijection Principle (237). The chain proposal accounts for the facts about one and its derivatives in (300 - 310). The relevance of this account of one for the issue of the binding properties of IMP can be seen in connection with the two conditions on featureagreement we have proposed. One condition - (297) - requires absolute feature-agreement; this is the one necessary for binding. The other, (296), simply rules out conflicting features; this is the one relevant for chains. We propose that indefinite anaphors and pronouns are those anaphors and pronouns which form chains with quantifiers and variables which formally bind them, while definite anaphors and pronouns do not. This is a purely formal distinction in the way "bound variable" interpretations arise with anaphors and pronouns - semantically, it has no consequences that I am aware of. However, it allows us to say that condition (296) holds for oneself, each other and themselves, while the stronger condition (297) holds for definite anaphors. The anaphors subject only to condition (296) can be bound by a quantifier with no features, i.e. IMP, while the others cannot. This is our account of the contrasts in (295).23 This account allows us to clear up an issue left open in earlier section. In 3.2.4.2., we proposed that non-purposive adjuncts could not be predicated of IMPs but they could be predicated of overt 6^-phrases because of a feature-agreement requirement on predication with just this kind of predicate. This explained the contrast between (166) and (183), since IMP lacks the features to agree with anything, while overt NPs clearly have them. It was left unclear why this feature-agreement requirement should hold for non-purposives but not for purposives. Now we can resolve this issue by saying that PRO always requires feature agreement in order to be controlled, and that PRO in non-purposives is definite, while in purposives it is indefinite. In that case, PRO in non-purposives is subject to the strict agreement requirement in (297), and so cannot ever agree with the featureless IMP, while PRO in purposives is indefinite and therefore subject only to the weaker no-conflicting-features requirement of (296). In other words, PRO in non-purposives is a phonologically empty coreferential pronoun, unable to corefer with a quantificational element like IMP, while PRO in purposives is bound by IMP in an extended chain. Clearly, much more could be said about this distinction, but the basic idea captures the relevant facts and seems to be intuitively correct. We now predict that examples like (295c,d) are perfect, which they clearly

166

The Representation of Implicit and Dethematized

Subjects

are not. At issue here are the restrictions on indefinite pronominal quantifiers like one and IMP. One is subject to condition (305). Is IMP subject to a similar condition? There is no condition which rules out IMP alone, but there is one additional condition relevant for binding: IMP must be modified in order to bind. Compare the following examples with (295). They show that if we add a Class I Adverb or RatC IMP is more acceptable as a binder than when it is unmodified: (312)

a. ?Damaging testimony was inadvertently given about themselves. b. Letters were cleverly sent to each other. c. The blankets were put on top of each other to keep warm. d. Obscene photographs were shown to each other to break the ice.

Our proposals for Adverbs in 3.2. predict that (312a,b) should be less good where the Adverb is between the subject and Infi. This prediction seems to be correct: (313)

a. ?*Testimony inadvertently might have been themselves. b. ??Letters cleverly must be sent to each other.

given

about

Whatever the exact formulation of this condition on the binding ability of IMP, these examples motivate the existence of such a condition, a condition which is violated in (295c,d). Furthermore, IMP and one seem to be alike in being subject a particular licensing condition. In both cases, we could regard this as a kind of identification requirement; IMP needs to be identified - at least as an antecedent - by an agentive, eventive modifier, while one seems to require government by a modal element. So our account of the binding properties of implicit arguments is that they cannot bind definite anaphors, due to their lack of features, but they can form chains with indefinite anaphors, principally each, but marginally arbitrary themselves. Since the formal distinction between the two classes of anaphors rests on definiteness, this analysis effectively claims that IMPs can only bind indefinite anaphors. IMP and one, the two quantificational indefinites discussed up to now, share the following properties, then: (i) they have a particular identification requirement; (ii) they are quantificational, in the sense that they do not corefer and they produce weak-crossover effects; (iii) they require coindexed elements to be bound (in variable chains); (iv) they are subject to the weaker feature-agreement condition of (296), rather than (297). Another arbitrary element which patterns in a similar way is that which instantiates another kind of implicit argument: the benefactive implicit

The By-Phrase and Implicit Arguments

167

argument discussed in Epstein (1984). Epstein proposes that examples like (314) contain an implicit benefactive argument which is interpreted as a universal quantifier, as shown in (315): (314)

It is fun [PRO to play baseball].

(315)

(Vx) it is fun for χ [for χ to play baseball].

Such implicit benefactives can only appear with a limited class of predicates (iunclear, possible, difficult, easy, fun, nice, amusing, etc.). This contrasts with the passive implicit argument, which can appear with any Verb that can be passivized. Also, the implicit benefactive is always precisely a benefactive, never an argument of any other kind, while the implicit argument of a passive can receive any θ-role a 6>>-phrase can receive (cf. 2.1.4.). These facts suggest that the identification requirement particular to this element is that it be assigned a benefactive θ-role. This is property (i) of arbitrary elements given above. The second property of arbitrary elements is also shared by the implicit benefactive: it is clearly quantifictional (Epstein gives evidence of scope ambiguities involving this element). How does the benefactive implicit arguments pattern with respect to binding? The basic paradigm for binding seems to be the same as for the passive IMP: (316)

a. *It is fun ei [PROi to b. ?It is fun e, [PROi to c. ?It is fun e; [PROi to d. It is fun e; [PROi to

shave shave shave shave

himself]. ourselvesi]. each other,]. oneselfi].

(316a) is ruled out in the same way as (295a); the implicit argument lacks the features to bind himself, and himself, being a definite anaphor cannot form a chain with the quantifier. (316d) is good because oneself, arbitrary PRO (which is in this respect like one) and the variable in the benefactive position all form a chain with the implicit universal. This is possible because oneself is an indefinite anaphor. Each other is allowed in (316c) for the same reason; similarly, themselves should be marginally OK here. So the benefactive implicit argument displays properties (iii) and (iv) above: it forms variable chains with coindexed elements, and, assuming it is inherently featureless, it is subject to the weaker agreement requirement of (296), and so can only bind indefinite anaphors. So the parallel between passives and benefactives is almost complete, except for one difference: ourselves is possible in benefactives, as shown by (316b). We propose that this difference is a consequence of the fact that the benefactive implicit argument is "inclusive" in the sense that the

168

The Representation of Implicit and Dethematized

Subjects

domain of quantification of the implicit universal can include the speaker, while the passive implicit argument is exclusive, and so the domain of, quantification of the implicit existential excludes the speaker. Note that this is a feature of universal as opposed to existential quantifiers generally: everyone can include the speaker but someone generally does not (cf. also Jaeggli (1986) on this distinction among types of arbitrary construction in Spanish). 24 If implicit arguments can bind, we should next ask how Principles Β and C are satisfied. It appears in fact that these principles operate as expected. Names cannot be construed as coreferential with the implicit argument: (317)

a. *Stories were told IMPi about Johni. b. *Letters were sent IMPi to Maryi.

In each of these examples, the reading indicated is impossible. We can attribute this to the residue of Principle C which prevents non-pronominal, non-anaphoric NPs from being bound - (234b) of 3.3.2. On the other hand, we suggested in 3.3.2. that some form of agreement condition on binding may replace this part of Principle C. If so, then no coindexing relationship between a full NP and an IMP could be ruled out by this Principle or its equivalent. Instead, a different principle seems to be at work. Noting examples like (317) and (318), Zubizarreta (1985) proposes the condition in (319): (318)

a. * Mary went to hiss farm with John;. b. *Mary sang his; song to Sally with John;.

(319)

If X is an argument of Ζ and Y is an adjunct of Z, then X cannot be referentially dependent on Y. (Zubizarreta's (27)).

Assuming that the PPs in (317) count as "adjuncts o f ' the main Verb in the way that the w/VA-phrases of (318) do, (319) accounts for both of these cases together independently of Principle C. In fact, (319) accounts for a coreference restriction on overt NPs in passive ö^-phrases. Coindexing with overt NPs in fry-phrases is impossible: (320) *Hisi mother is loved by Johni. We propose to adopt a version of (319) to account for examples like (320). The crucial notion in (319) is the distinction between the relation "argumentof" and the relation "adjunct-of." If we take "argument-of to include

The By -Phrase and Implicit Arguments

169

the GFs subject-of, object-of etc., then "argument-of' may be related to the notion "occupies an Α-position." So we reformulate (319) as (321): (321)

If X is immediately dominated by Ζ, Ζ in an Α-position, and Y is immediately dominated by W, W in a non-A position, then referential dependency of X on Y is impossible if X and Y are in the same clause.

(321) is different from (319) in detail, but close to it in spirit. (319) contains two references to "immediate dominance;" it is the phrase immediately dominating each potentially coreferring NP whose status as an A or A'-position is crucial. We illustrate how (318a) is ruled out, as follows: (322)

Mary went to

PP=W is not out by (321). X and Y are following type (323)

[np=z [ x

hisi] farm] [pp=w with

[γ

Johni ]].

in an Α-position while N P = Z is, so this example is ruled (318b) is ruled out in a similar way. (321) only holds if in the same clause. That this is correct is shown by the of example:

It was believed by Johni that hisi mother was sick. (Zubizarreta's (26)).

It might seem that (321) must refer to the order of the pronoun and the name, given contrasts like (324): (324)

a. ??Fred was introduced to hiss mother by Johni. b. ?Fred was introduced by Johni to his; mother.

However, we can assume that (324b) is the result of extraposition of to hiSi mother. After extraposition, this PP is in an A' position, and so (321) is inoperative, and coreference is hence (marginally) possible. Also, an NP more deeply embedded in an Α-position such that it is not immediately dominated by the category in the Α-position allows coreference: (325)

a. b. c.

A picture of hisi mother was given to Mary by Johni. Rumors about his; mother were spread by Johni. Hisi mother was killed in front of Johni.

So it seems that (321) - or something close to it - is descriptively correct, and that no residue of Principle C is relevant for accounting for the coreference properties of IMPs. 25

170

The Representation of Implicit and Dethematized

Subjects

I MPs appear also to obey Principle B. This is shown by the following examples, assuming that the pronouns in the PPs in the examples below are in the same binding domain as their antecedents: (326)

a. *Stories were told IMPj about himi. b. * Letters were sent IMPi to heri.

However the same question as above arises concerning the agreement condition on binding in (326). Presumably this condition holds also for the antecedent-pronoun relation covered by Principle B. We said earlier that the implicit argument cannot be an antecedent because it lacks agreement features. If this is so, then (326) cannot be ruled out by Principle Β since the pronouns are not in fact bound by the implicit argument as they cannot agree in features with it. So Principle Β is not at work here. Instead, we propose that the pronouns in (326) are variables bound by the implicit existential (without forming a chain with it). Now, if there is c-command between the implicit-argument position and the pronoun, we have a Strong Crossover violation, i.e. ultimately a θ-criterion violation. If there is no c-command between the implicit-argument position and the pronoun, weak crossover is violated, i.e. the Bijection Principle again. So again it is possible to account for these examples with independent principles, and so maintain the idea that IMPs are basically unable to bind. An assumption behind our account of the ungrammaticality of (326) is that the implicit existential must bind a pronoun that it c-commands (and therefore, given the above proposals, form a chain with it). More generally, we maintain the following: (327)

If a pronoun is coindexed with an inherent quantifier (in the sense introduced in 1.4.) then that pronoun must be bound by the quantifier.

We retain the rough characterization of the class of inherent quantifiers given in 1.4. (note that (327) is needed to rule out (300), in addition to the assumption that one is an inherent quantifier). The quantifiers in (328) are inherent quantifiers, then: (328)

a. *Hisi wife wonders which man* left. b. *Heri mother saw someone* leave. c. *Whose, mother does he, love?

In (328), the pronoun is coindexed with the quantifier but not c-commanded by it. Hence the pronoun cannot be bound by the quantifier, and must

The By-Phrase and Implicit

Arguments

171

instead be coreferential with it. But if these are inherent quantifiers, then by (328) coreference is impossible. So no pronoun - quantifier relation is allowed and the indexing shown is ruled out. (327) also allows us to account for all cases of NP-movement weak crossover, negative and positive. We repeat the relevant negative example once more: (329) *[Hisi mother]; wasn't killed tj IMPi. (329) is impossible on the indicated reading, i.e. it cannot be a statement about matricide. In 3.3.3. we accounted for the impossibility of the positive counterparts of (329). In these cases the implicit existential can have clausal scope and, since this operator has the index of his, it binds both hist and ei. This results in a violation of the Bijection Principle. Such an account fails to rule out the negative (329), however. However, if the implicit existential is an inherent quantifier, then NPmovement weak crossover is ruled out in exactly the same way as the examples in (328); the pronoun cannot be coreferent with the quantifier because it is the wrong kind of quantifier, and cannot be bound by it because it is not c-commanded by it. Hence the indexing is disallowed. In this section, we have argued that the fact that implicit arguments do not enter into binding relations is a function of the agreement condition on antecedence in (297). This condition on binding prevents all definite anaphors from being bound, as IMP lacks features. The only anaphors that IMPs can bind are indefinites - oneself, each other, and marginally themselves. Here, however, we suggested that chain-formation, not just binding, was the issue, and that the agreement condition holds in the weaker form (296) with chain-formation. We also observed certain constraints on the implicit existential, related to modification. Both of these properties of IMPs seem to be shared by other elements with arbitrary reference: one and the benefactive implicit argument. For coreference, we saw that two conditions, one on coreference between arguments and adjuncts, and one on pronoun-quantifier relations, were enough to account for coreference restrictions on both overt NPs in Zvy-phrases and NP-movement weak crossover. So it appears that with implicit arguments, both benefactive and passive, we are dealing with a variable bound by an implicit quantifier. This seems to be the only statement that we need to make about the LF properties of passive implicit arguments. If the presence of the quantifier is the consequence of the fact that the implicit argument has an arbitrary interpretation, then all we would need to say about implicit arguments is that they have an arbitrary interpretation, a statement which any theory of passives must make about the agent of agentless passives. However,

172

The Representation of Implicit and Dethematized

Subjects

it is clear that the suppressed subject of a middle also has an arbitrary interpretation, so if we are to reduce our theory of passive to the issue of arbitrary subjects, we need to specify how it is that middles are different. This will be the topic of the next Chapter. 3.4. CONCLUSION

This chapter's goal has been to argue by demonstration that the implicitargument effects found with passives are the reflex of the presence of a structural position in VP created by the downward movement of the original external argument in this construction. We have carried out this demonstration in two ways: first by showing that the modification properties of implicit arguments fit well with a fairly minimal structurally based theory of modification and predication, and second by pointing out and accounting for the referential properties of implicit arguments. The referential properties we observed provide the strongest evidence for our theory. We saw that implicit arguments can produce both weak and strong crossover effects: the latter indicate that the implicit argument is in fact an argument in the sense that it is relevant for the formation of θ-chains, while the former indicate that implicit arguments are interpreted quantificationally. We saw in some detail that the scope restrictions on the "implicit quantifier" are largely the same as those which hold for 6y-phrase quantifiers generally, and suggested that these restrictions may be related to the phenomenon of "inner islands." Finally, we considered the binding capacities of implicit arguments. These are severely constrained by the fact that implicit arguments lack agreement features. However, we suggested that implicit arguments can bind indefinite anaphors because anaphors in this class can form chains with inherent quantifiers that bind them. We saw two other elements which act like passive implicit arguments in this regard: one and the benefactive implicit argument discussed by Epstein. Our proposal that all these elements are quantificationally represented amounts to an outline theory of arbitrary reference, as all these elements clearly are arbitrary in reference. The main topic left open for future research here is the status of arbitrary PRO; can this element be assimilated to one and the various kinds of implicit argument (see Clark (1985) and fn. 24)? In any case, if this "quantificational theory of of arbitrary reference" can be maintained, then all the facts concerning binding and weak crossover with implicit arguments follow from the fact that this argument is arbitrary in reference, and so all that we would need to say (and all that the child would need to know) concerning the non-overt subject of an agentless passive is that this element corresponds to a subject argument with arbitrary interpretation. From this statement, which any theory of passives must make about the agent of agentless

The By-Phrase and Implicit

Arguments

173

passives, all facts discussed in this Chapter follow, given the assumptions we have made and attempted to supply with independent motivation. This, then, is our theory of passive. In the chapters to follow, we explore how this basic approach extends or fails to extend to other constructions involving the "suppression" of a subject argument: middles, nominals and impersonals. All these cases might be argued to involve arbitrary subjects, also. We will see how the arbitrary subjects in these cases are similar to and different from the IMP of passives. NOTES TO CHAPTER T H R E E 1. The data and judgements given here are from Chomsky. Rizzi (class, Fall 1984) points out that examples like (i) are considerably better than (6a): (i)

Such privileges should not be kept to oneself.

Also, reciprocals seem slightly better than reflexives: (ii)

?Letters were sent to each other.

We will discuss these facts in 3.3.4. 2. Roeper (1983) gives examples like (i), which appear to contradict (9): (i)

The game was played barefoot.

Here it is clearly the players who are barefoot. However, barefoot is arguably a non-/y Adverb, and does not directly modify the implicit argument, but rather the VP. 3. Actually, as is typical with selection violations, the book can be pragmatically construed as meeting the selection requirement of the Adverb, given a sufficiently unlikely real world situation. Note that whenever we construe the Adverb as able to modify an argument, the clause has an event reading. Middles in general lack event readings (see 4.2.), and so they are incompatible with Adverbs of this type. The difference with (4a) is that (4a) requires no such bizarre construal. That this is due to the presence of passive morphology in (4a) and its absence in (4b) seems obvious, since this is the only difference between the two examples. So, without denying that selectional restrictions may be bound up with pragmatics, we regard the contrast in (4) as essentially syntactic. 4. We have neglected speech-act Adverbs in our exposition. These Adverbs can appear initially and are often very similar in meaning to Class III: (i)

Ultimately, is this going to be the best analysis?

(ii)

Frankly, as for myself, I don't believe in Austin.

These are possibly CP-modifiers, and as such may be predicated with C. Note that they cannot appear anywhere inside CP without comma intonation: (iii)

*John frankly is a fool.

(iv)

*John is frankly a fool.

174 (v)

The Representation of Implicit and Dethematized

Subjects

John is a fool, frankly.

5. Note that (i) is also bad: (i)

*Is definitely he coming?

Here the Adverb is presumably adjoined to IP. Perhaps it is unable to predicate with I because it is excluded by a segment of the adjoined maximal projection of I. Cf. Chomsky (1985) on exclusion and Jackendoff (1972:84), Jaeggli (1983), Kayne (1984) on inversion. In 3.3.3.3., we will have reasons to propose that subject-aux inversion adjoins aux to IP (contra Chomsky (1985)), so it is not clear that the account for (23c) is correct. 6. Andrews gives V' where we have given VP. In all other respects, the representations are identical to those given by Andrews. 7. Movement of Adverbs to sentence-initial position is possible, as in (i): (i)

Deliberately, the book was sold (by the publisher).

Since the iy-phrase/IMP can be understood as predicated of the Adverb here, we regard this as an instance of topicalization of the Adverb from'its base-generated position. This topicalization is subject to a number of constraints. First, it is clause bound: (ii)

'Deliberately, they told me that the book had been sold t.

We discuss a similar constraint on predicative Adjectives in fn. 10. The account given there carries over to this case, and also the parallel data with Rationale Clauses: (iii) (iv)

To collect the insurance, the boat was sunk t. *To collect the insurance, they told me the boat had been sunk t.

On the general parallel between Class I Adverbs and Rationale Clauses, see 3.2.4.1. Chomsky (1985) notes a further fact about Class I adverbs and extraction. This is that they cannot be extracted. Cf.: (v)

*How deliberately did John read the book/was the book read?

Again, the same is true of subject-predicated Adjectives. The suggestion in fn. 10 that IP is a barrier for APs, in combination with Chomsky's stipulation that " [ + w h ] " elements cannot adjoin to IP, will account for this in terms of antecedent-government. For more details on these notions, see 3.2.2. and fn. 10. Cf. also 3.3.3.3. on extraction of adjuncts in negative environments. 8. It is hard to determine whether this distinction is a function of selectional properties of individual Adjectives or whether, given the right context, any Adjective is interpretable in either reading. Certainly it is not hard to construct examples where the same Adjective is interpretable in either reading: (i)

a. b.

John beat the cat angry. John walked the dog exhausted.

If angry and exhausted are predicated of John here, the reading is that John was in the relevant state before the time of the action described in the VP, i.e. the Adjectives are interpreted

The By-Phrase and Implicit

Arguments

175

as temporally independent. On the other hand, if the Adjectives are predicated of the cat, the dog, respectively, then the reading is the temporally dependent one; the animal got angry or exhausted as a result of the action described. So it appears that Adjectives can vary between temporally dependent and temporally independent readings. However, some Adjectives are hard to get in on one or the other reading, e.g. tall and other stative Adjectives are hard to construe as temporally dependent (cf. 4.2. for an account of stativity that captures this). Moreover, in these cases temporally dependent readings can be obtained if a sufficiently bizarre state-of-affairs is imagined. This kind of pragmatic construal for otherwise anomalous readings is typical of selectional violations (cf. fn. 3 above). In any case, this matter has little bearing on our treatment of Adjectives; the important point is that Adjectives must be predicated of some NP. 9. These examples can be used to support the claim made earlier that lexical governors of empty categories must be in situ. Extraction of the object of an AP small clause is possible: (i)

a. b.

Who do you consider Bill fond of t? Who do you wonder whether to consider Bill fond of t?

Example (84) in the text shows that extraction of the predicate AP is possible. However, extraction of both is impossible: (ii)

a. b.

*How fond, did you wonder whoj to consider Bill ti of tj? *WhOj did you wonder how fondi to consider Bill ti of tj?

(NB the examples show that the ungrammaticality here is not a crossing effect). Our suggestion that lexical governors must be in situ will account for the contrast between (i) and (ii). 10. A further contrast between (86a) and the others regards topicalization, as shown in (i): (i)

a. b. c.

Sadi, Johni left the party. *Sadi, John met Maryi. *Sadi, John made Maryi.

How can we account for this? There are two possibilities: either the Adjective moves or it does not. If we say it moves, then we have to explain how antecedent-government is possible, given the earlier account of the non-extractability of these subject-predicated Adjectives. If it can be base-generated in topic position, then we have to weaken the generalizations in (80). The latter course seems to be the worse of the two, since whatever approach we adopt here will contradict the data in (70) and (71), as it will allow all cases of subject-predicated Adjectives to appear outside VP. Furthermore, the facts of (71) seem to hold for APs in topic position. Cf.: (ii)

a. b.

Angry at himself/*Johni, he; left. Which pictures of himself,/John,, did he* like?

(iib) shows the parallel between these cases and standard instances of reconstruction. So we must now explain the following contrast in terms of antecedent-government: (iii)

a. b.

*How sad did John leave the room t? Sad, John left the room.

In the text, we claimed that the barrierhood of VP blocked extraction of non-dependent

176

The Representation

of Implicit and Dethematized

Subjects

APs. Suppose instead (i) that topicalization is adjunction to IP, and (ii) that, for APs, IP is a barrier. Although VP is a barrier for APs, as it is for everything else, AP escapes its effects by adjunction, as is the case for all other categories, (iiia) is ruled out because IP is a barrier for APs and adjunction to IP is impossible for [ + w h ] elements (the latter assumption is made in Chomsky (1985)). Further movement, as in (iv), is impossible: (iv)

*Sadi, John said Mary, had left the room.

Sad cannot be predicated of Mary here. The account for this is the same as that for (ii) of fn. 7. Movement into Comp is possible for topics, as they are [-wh] elements and CP cannot be adjoined to as it is an argument. However, CP is a complement of the matrix Verb, so it is not a barrier to movement of AP. At this point, the third way a category can be a barrier to antecedent-government (but not to movement) becomes relevant: the projection of a head Ζ intervening between X and Y is a barrier to government of Y by X in the configuration .. X .. [ z n . . Y .. ]. Chomsky (1985) calls this the Minimality Condition. In the examples in question, the presence of C' means that the minimality condition prevents antecedent-government of the trace adjoined to IP by the next trace of sad (adjoined to the matrix VP). I have no suggestion to make as to why IP should not be "defective," i.e. be a barrier, for APs but for nothing else. 11. This is too simple an account as it stands. First, there are cases of what appear to be Α-control where IMP can control: (i)

a. b.

John was asked PRO to leave, It was decided PRO to leave.

For Verbs which allow this possibility, we can say that the condition that extended-chain formation is restricted to DS positions does not hold (although cf. fn. 24 for another proposal for these examples). Note, however, that Jaeggli's observation in (110a) still has some force here. Compare (ii) with (i): (ii)

a. ??John was asked by Bill, PRO; to leave, b. ??It was decided by Bill· PRO, to leave.

Further facts suggest that the D S restriction on extended chain formation does not always hold even for Verbs like promise. The facts are known as Visser's generalization: (iii)

a. b. c. d.

Fred, promised [PROi to leave]. *Fredi was promised [PROi to leave]. Fredi was promised [PROi to be left alone]. *Fred was promised IMPi [PROi to be left alone].

(iiia) is a straightforward case of Α-control, (iiib) falls under the condition in the text if we modify this condition so that only D S arguments in their D S positions are allowed to form extended chains. This amounts to saying that for some Verbs extended-chain formation takes place at D S (this would have to be a requirement on Verbs and not on PRO if PRO's control property follows from the Visibility Condition). So this example is ruled out because promise requires its D S subject to form the extended chain, (iiic) is the example which indicates that the chain formation condition can be violated by promise. Promise may fail to impose a D S requirement on extended-chain formation. In that case, PRO forms an extended chain at SS with an available argument, one possiblity being then the SS subject, as in (iiic).

The By-Phrase and Implicit

Arguments

177

The ungrammaticality of (iiid) suggests that there is a further Thematic Matching Requirement on the extended chain, independently of the level at which chain-formation takes place (cf. Rüzicka (1983) on Thematic Matching). We now have two factors which determine Α-control: (i) Thematic Matching and (ii) D S or SS chain-formation. Some Verbs seem to impose no requirements, i.e. to allow SS chain-formation freely. Examples of these are decide and ask, as in (i) and (ii). Note, however, that even here Thematic Matching is relevant; (iv)

John was asked (by Bill·) P R O i to be allowed to leave.

So further clarification of these issues is needed. Since the conditions determining control properties of Verbs are so obscure, we have not attempted in this book to provide an account of the control properties of implicit arguments. See also fn. 12. 12. The locality conditions on Α-control are very unclear in cases like (111a). Chomsky (1986) notes that non-locally-bound PRO acts like a pronoun in allowing split antecedents and in not being "subject-oriented:" (i)

Bill wanted Tom to agree PRO to swim across the pond together.

(ii)

They thought I said the decision PRO to feed each other was foolish.

PRO in (111a) acts like this: the antecedent of PRO can be split between they and / (as shown by the possibility of substituting ourselves for each other)·, and the example is clearly not obligatorily subject-oriented, cf.: (iii)

They thought I had suggested to Mary, that feeding herself, would help.

This is in fact preferable to taking the subject of suggest as antecedent; (iv)

?They thought I suggested that feeding myself would help.

The only suggestion we could make for these "pronominal" PROs is that these are instances of coreference and not of the binding-like relation of standard cases of control (as in fn. 11). Possibly, then, these examples involve coreference with a null pronoun rather than chainformation (see also 3.3.4. where this distinction is exploited in a different context). In any case, the data as it stands does not counterexemplify our proposals for control. 13.1 am indebted to Mamoru Saito for this suggestion. 14. As noted by Jones, VP-movement and deletion rules obligatorily analyze PCs as part of VP: (i)

?John bought Moby Dick [to read] and Fred did, too, [to keep on his shelf]. (Jones' (65), p. 29).

(ii)

*John said he'd buy it to read, and buy it he did to read. (Jones' (68.2), p. 30).

(iii)

*Buy it though John may to read, he'll never get past the first chapter. (Jones' (71.3), p. 30).

(iv)

*What Mom did to pick up the laundry was send John out. (Jones' (75.2), p. 31).

178

The Representation of Implicit and Dethematized

Subjects

Given Saito's analysis of parasitic gaps, these data lead to a contradiction. How is it that PCs are obligatorily analyzed as part of VP by these rules and yet can extrapose when they contain a parasitic gap, as, following Saito, we assume is the case since these constructions allow parasitic gaps? In fact, the situation is even more acute. Assuming that wA-movement of an empty operator is involved in OPCs, extraposition must always take place in these structures or Principle C will be violated. This means that we must either alter our definition of c-command, so that something can be in VP for the VP-movement rules and yet not be c-commanded by complements of V - a possibility not allowed by the Aoun/Sportiche definition of c-command assumed here - or conclude that the "VP-movement" rules that we have been using as tests for VP-consituency are not reliable, in that they can and, in cases like those under discussion, must involve other constituents in addition to VP. Although we incline towards the latter solution since less is at stake theoretically if we take this tack, we will leave this important question open here. 15. Arguably, bare each also QRs at LF. The evidence for this is the fact that bare each seems to be outside the scope of Tense, unlike reciprocal each (Lebeaux, p.c.): (i)

They each saw the other.

(ii)

They saw each other.

In (ii), we understand that the individuals saw each other simultaneously, while with (i) there is at least a preference for the interpretation where the individuals saw each other at different times. We could capture this by saying that bare each is in an extra-sentential position, and so outside the scope of tense, at L F in (i). 16. The statement of the locality requirement on each and other is quite tricky. Consider: (i)

a. b.

*John & Maryi said that John had brought some friends for eachi other to meet. *They, claimed that Fred was wondering whether each, other would win.

These examples can be ruled out by Principle A if we assume that each can only move as far as the nearest VP. Why should each movement be so restricted? There are two reasons why we cannot appeal to the ECP to provide the answer to this question. First, the assumption that each must antecedent-govern its trace will rule out precisely the examples we are concerned with, namely those where each is in subject position of a wA-complement, since the ECP will be violated whenever each moves out of the IP in which it appears at SS. Second, as Lebeaux points out, examples like the following are ambiguous: (ii)

John and Mary knew that they liked each other.

This sentence can mean either of the following: (iii)

a.

John and Mary knew that f John liked Mary t

b.

I Mary liked John J. r John knew that John liked Mary 1 I Mary knew that Mary liked John J.

Lebeaux proposes that (iiia) corresponds to narrow-scope each and (iiib) to wide-scope each. Hence he allows each to move to higher VPs. So, if we concur with Lebeaux's analysis, no locality constraint on each can be proposed. However, we could perhaps answer these arguments by pointing out (i) that the contrast

The By-Phrase and Implicit

Arguments

179

between /Acf-complements and wA-complements which holds for wA-extraction is paralleled by each movement, as shown in (iv) and (v): (iv)

a. b.

*They thought that each other would win. *Who did they think that t would win?

(v)

a. ??They wondered whether each other would win. b. ??Who did they wonder whether would win.

These contrasts could both be explained by the ECP. Second, Debby Schlindwein (p.c.) points out that the ambiguity in (ii) is not limited to clauses containing each. It is also found in examples like John and Mary knew that they liked fish. Since it is hard to decide whether the ECP is operative in constraining the distribution of each, we could try to capture the distribution of reciprocals by imposing a constraint on other. A possible proposal is that other must be contraindexed in its CFC. We leave this question open. 17. Given our definitions of C F C (Chapter Two (106)) and Pred(X) (Chapter Two (109)), the result seems to be that reflexives must be governed in their Pred and bound in their CFC. The one crucial element missing from this characterization is a specification of the role played by Tense in determining the distribution of reflexives. The obvious move is to incorporate Tense into the definition of Pred. This is a task we will not undertake here (although see Chapter Four on the relation between predicates and Tense generally). Note also that the extension of this approach to each may solve the problem discussed in fn. 16. 18. Further support for the reflexive/reciprocal distinction we have made might be gleaned from the following contrast: (i)

a. b.

They gave themselves to each other, T h e y gave each other to themselves.

The VPs in these examples have the following LF representations, given what we have said: (ii)

a. b.

[vp eachi [Vp gave ei [pp to fc other ]]]. [VP eachi [vp gave [t; other] [ PP to $]]].

The configuration in (iib) is ruled out by the Bijection Principle, as the quantifier each locally binds two variables (assuming the null reflexive = pro must count as a variable here - cf. 1.4. and 3.3.4. on pronouns as variables). So this is a weak crossover violation (cf. 1.4.). Note that the existence of a similar contrast in (iii) suggests, by parity of reasoning, that the first object must be embedded in some further constituent: iii)

a. b.

They gave each other themselves, *They gave themselves each other.

These facts could support Kayne's (1984) proposal for a [PP NP] structure for double-object constructions, with PP headed by a phonologically null Preposition (contra the conclusion of Barss & Lasnik (1986)). The force of these observations is weakened by the fact that Prepositions like to do not in general interfere with c-command relations. Cf. (iv), which, as discussed in 3.3.2., can

180

The Representation of Implicit and Dethematized

Subjects

be ruled out by Principle C on the assumption that to does not prevent him from c-commanding John: (iv)

*It seems to himi that John, is a genius.

However, it is possible that Prepositions only fail to count for c-command when they are adjacent to V. This can be accounted for if we assume that Prepositions in English can adjoin to V under adjacency (cf. the discussion of seem to vs. sembrare a in 3.3.2.). Note also the contrast in (v): (v)

a. b.

Theyi expect eachi other to hate themselvesi. *Theyi expect themselvesi to hate eachi other.

This contrast is also amenable to a "weak-crossover" account, given the assumptions of the text. On the subject of binding relations in double-object constructions, the following example poses a potential problem for the account of NP-movement strong crossover in terms of chain-formation given in 3.3.2.: (vi)

They, were given themselvesi.

If (vi) can be interpreted analogously to (vii) in dialects that allow this kind of passive, Rizzi's chain-formation account must be wrong, since themselves intervenes between they and its trace, preventing the formation of a well-formed chain: (vii)

A booki was given John ti.

However, since it is impossible to judge intuitively whether the Theme or the Recipient has been passivized in (vi), the status of this sentence as a counterexample is unclear. 19. Is there weak crossover with /ougA-movement? Cf.: (i)

Hisi mother is easy to love.

Apparently (i) allows the reading where the benefactive implicit argument here is understood as coreferential with his;. This suggests that there is no implicit argument here, or that if there is one it is rather different to the passive one in its LF properties. For more on benefactive implicit arguments, see Epstein (1984) and 3.3.4. (where Epstein's proposals are adopted). 20. Note that if the subject quantifier is stressed it can have narrow scope: (i)

TWO languages are not known by every linguist.

This example can mean "every linguist does not know two languages," where the languages differ according to the which linguist is chosen. It seems to be a general property of focussed quantifiers that they have narrow scope. 21. There are two big problems with this idea. The first is that there is no confirmation that fty-phrases count as adjuncts for the ECP. Compare the following, for example: (i)

?By whom did you wonder whether John was killed t?

(ii)

*How did you wonder whether [John was killed t]?

The By -Phrase and Implicit

Arguments

181

Second, overt Zy-phrases can be extracted at SS out of the scope of negation: (iii)

Who wasn't John killed by t?

We can account for these facts and take a step towards subsuming (278) under the ECP if we assume that en antecedent-governs a f>_y-phrase trace at SS but not at LF. In terms of Lasnik & Saito's (1984) theory, en gamma-marks a iy-phrase trace at SS, but not at LF. This licenses SS movement, but not LF-movement of the kind under discussion. However, it should be pointed out that it is in any case far from clear that the inner-islands facts should be subsumed under the ECP. To do this, it is necessary to show how negation prevents antecedent-government between extracted adjuncts or quantifiers and their traces, and also construct a structurally based theory of negative polarity. No such proposals exist at present, as far as I am aware. 22. Interestingly, (309b) is nowhere near as bad as its counterpart with a definite anaphor: (i)

*Johni believes people to like himself.

The same is true of French soi, as pointed out by Pica (1985). Cf.: (ii)

a.

b.

On ne souhaite jamais que les gens ne regarde que soi. One neg wishes never that the people neg look-at only oneself. "One never wants people to only look at oneself." * lis ne souhaitent jamais que les gens disent du mal les uns des autres. They neg wish never that the people say of bad the ones of-the others. "They never wish that people speak ill of each other."

There's also evidence that variable chains formed by arbitrary pronouns are exempt from the θ-criterion. An example like (iii) is good even though it should violate the CFA: (iii)

It's fun [PROi to shave oneself].

Clearly examples like this require further investigation. 23. The definite/indefinite distinction proposed here is not a complete stipulation. Arguably, one other semantic property of definites is not syntactically manifested at LF, while the same property of indefinites is. It is quite likely that both definite and indefinite NPs should be treated as quantified expressions from the point of view of semantics, however, only non-specific indefinites display the properties which motivate treating them as quantifiers in the syntax of LF. Non-specific indefinites, unlike specific indefinites and definites, cannot corefer and induce weak crossover effects: (i)

I want to meet the/a Norwegian and Bill wants to meet himi too.

(ii)

I want hisi mother to love the/a Norwegian!.

In these examples, the definite is fine and the indefinite must be interpreted specifically, despite appearing in a context which in general allows a non-specific reading. This can be explained if non-specific indefinites are treated as quantifiers at LF and specific indefinites as names; the specific reading is forced in (i) and (ii) because the non-specific is impossible, being unable to corefer in (i) and responsible for a weak crossover violation in (ii). In conjunction with the proposals of the text, these facts suggest that the distinction between definiteness and indefiniteness is a matter of LF accessibility for certain processes. This

182

The Representation of Implicit and Dethematized

Subjects

difference may underlie many syntactic and semantic differences between the t w o types of NP. A l s o , the idea that a weaker condition on agreement holds for binding than for coreference is supported by the well-known lack of number-agreement in cases of binding but not of coreference (cf. Evans (1977), H a i k (1984)). 24. W e will not discuss arbitrary P R O here, except to note one apparent parallel between this element and passive implicit arguments. L e b e a u x (1984) identifies an interesting property of arbitrary P R O . He points out that the t w o occurrences of arbitrary P R O in the f o l l o w i n g examples are restricted in that they must pick out the same individuals, even though they are both arbitrary: (i)

a.

P R O b e c o m i n g a movie star involves/entails P R O being recognized by everyone,

b.

P R O winning this game requires P R O mastering the endgame.

Consider next (ii): (ii)

* P R O i b e c o m i n g a movie star involves PRO] recognizing youi.

If the first occurrence of P R O in (ii) has the index o f you, then the second one must and the example is then ruled out by Principle B. T h e passive implicit argument allows such linked readings, as the natural construal of (iii) shows (although this is not a requirement): (iii)

F o r this g a m e t o be w o n the endgame must be mastered.

Arbitrary P R O also has the same binding properties we observed f o r implicit benefactives: (iv)

PROi to shave oneselfi/each 0theri/?0urselvesi/themselvesi/*himselfi is boring.

So arbitrary P R O has properties (iii) and (iv) of arbitrary elements discussed in the text. M o r e o v e r , it is possible to maintain that arbitrary P R O is quantificational (cf. C l a r k (1985) and C h a p t e r One fn. 5), thus this element c o n f o r m s with property (ii) o f arbitrary elements. Whether arbitrary P R O has a particular licensing condition is unclear, however. A n o t h e r prediction which emerges f r o m the account of binding by I M P s is that I M P s should be able to bind arbitrary P R O . This seems to be true, as examples like (i) o f fn. 11 show: (v)

It was decided IMPi PROi to leave.

Cf. also; (vi)

J o h n was asked IMPi h o w P R O i to fix the sink.

25. A further issue regarding coreference is raised by the f o l l o w i n g examples (from Jaeggli (forthcoming, a: 14)): (i)

a. b.

(ii)

a. b.

Johni lost hiSi way. ??HiSi w a y was lost by Johni. Johni raised hisi hand. *His handi was raised by Johni.

The By-Phrase and Implicit

Arguments

183

In the (a) examples, John and his must refer to the same person; while in the (b) examples this is impossible. Jaeggli attributes this to the fact that the ^ - p h r a s e is not assigned a compositional θ-role by VP. However, we saw in Chapter T w o , examples (35 - 38) that compositional θ-roles are assigned to the iy-phrase. Moreover, we proposed in 2.1.4. that only Agents can be compositional arguments, and so we d o not want to consider the Experiencer/Agent ambiguity in (iii) to be a question of compositional θ-role assignment (as proposed in Chomsky (1981), for example): (iii)

John; broke hisi leg.

So we propose an alternative to Jaeggli's account for (i) and (ii). These examples - and (iii) - all involve inalienably possessed NPs. One proposal (from Gueron (1983)) for the possessive pronouns in inalienably possessed N P s is to treat them as anaphors. Support for this idea comes from the following: (iv)

a.

They, all lost theiri way.

b.

Theyi all raised thein hand(s).

The sentences in (iv) have only the interpretations in (v): (v)

a.

(Vx) (x lost x's way)

b.

(Vx) (x raised x's hand(s))

So we see that their is obligatorily bound by they in (v). Consider also: (vi)

*Johni forced Bill to lose hisi way.

If inalienable his is an anaphor, we can account for (vi) by Principle A. We choose this example because of the quasi-idiomatic meaning of way here ( = roughly "knowledge of the correct way"). Other examples, like (iiia), are always construable on a non-inalienable reading. Perhaps this is because English lacks a morphological distinction between alienable and inalienable possessive pronouns. In Romance languages, the definite article appears in place of possessive pronouns to give an inalienable reading. So in French we find: (vii)

(viii)

(ix)

Jeani a leve la, main. J has raised the ( = h i s ) hand. *Jean, a persuade Pierre de lever lai main. J has persuaded Ρ to raise the ( = h i s ) hand. lis, ont tousi leve lai main. They have all raised the (=their) hand(s).

The idea that inalienable possessive pronouns are anaphors would explain why John and his must refer to the same person in (ia) and (iia) and cannot refer to the same person in (ib) and (iib). We view the latter examples as analogous to (x): (x)

*Himselfi was seen by Johni.

Chapter Four

Middles and Nominals

In this chapter we consider how to analyze middles and nominals in the light of the analysis of passives presented in the preceding chapters. We also briefly consider adjectival passives and -able Adjectives, in relation to these constructions. We will consider two proposals for middles. The first takes middles to be derived by the lexical rule of "Externalize Theme," as in Williams (1981). We adopt a slightly different conception of the nature and functioning of Williams' rule. The lexical approach accounts for many restrictions on middles: their limitation to transitive Verbs of a particular type, and more general similarities to Adjectival passives and ableAdjectives. It is also possible to formulate, drawing partly on ideas proposed by Rizzi (1985), an account of the interpretation of the subject θ-role, and of the apparent operation of "Burzio's generalization" with middles. Our alternative proposal is that middles are derived by move-NP in syntax, just like passives (so our second proposal, but not the first, follows Keyser & Roeper (1984) and Hale & Keyser (1985)). A crucial role is played in this account by the observation that middles are Stative Verbs. Stative formation has two parts: non-assignment of the Agent θ-role and "temporal obviativity" of the Verb with respect to Tense. These two properties are necessarily linked by our proposals for structural θ-role assignment and its relationship to temporal dependencies in 2.2.2. and 3.2.2. We also tie these notions to Case-assignment, in order to account for NP-movement in middles. This leads to a formulation of Burzio's generalization as a condition on Case-assignment. Finally, we discuss two further puzzling aspects of middles: the constraints on which Verbs form middles, and the apparent requirement that middle VPs be modified. We can characterize the former in terms of the Dowty (1979)/Vendler (1967) aspectual categorization of Verbs, which in fact allows us to define a notion of Theme. The latter issue is much less clear: we will revise the identification requirement on predicates of 2.2.3. (102) to account for this. Middles are exemplified in (1) (the examples are all from Keyser & Roeper (1984)):

186 (1)

The Representation of Implicit and Dethematized a. b. c. d. e. f. g. h.

Subjects

Bureaucrats bribe easily. The floor paints nicely. This book translates quickly. Those chickens kill easily. The baggage transfers efficiently. Messages transmit rapidly by satellite. The letters transpose easily. Boxes like these will not transport easily.

Keyser & Roeper discuss the differences between middles and ergatives; for the moment we restrict our attention to the class they isolate as middles. We return to their proposed distinction between middles and ergatives in 4.4. In 4.1., we present the main properties of middles which distinguish them from passives and from transitives, and introduce the lexical account based on the rule E(Th). In 4.2., we present the evidence that middles are stative, define stativity in terms of our formalism for temporal indexing, and sketch how this can account for many properties of middles, even if we take middle-formation to be a syntactic process. In 4.3., we briefly outline the Vendler (1967)/Dowty (1979) aspectual classification of Verbs in an attempt to clarify the connection between θ-roles and aspect, and more precisely the nature of Theme θ-roles. We then integrate this discussion with the temporal indexing system in way intended to capture the constraints on which Verbs can form middles. This also leads us to a potentially illuminating reformulation of Burzio's generalization. In 4.4., we discuss Keyser & Roeper's analysis of the distinction between ergatives and middles, and finally, we propose a principle of predicate identification that determines the interpretation of the non-assigned Agent and imposes the modification requirement on middles. In section 4.5., we develop a rather more sketchy and speculative account of nominals. We will propose that the central property of nominals is that NPs lack an Infi node. The absence of Infi has consequences for aspect, θ-role assignment and Case-assignment in NP. We also discuss the phenomenon of "passive in NP," which we argue to be another instance of E(Th), and provide an account of implicit-argument phenomena in NPs. 4.1 PROPERTIES OF MIDDLES

Keyser & Roeper are not the first (cf. Belletti (1982), Ruwet (1972), Williams (1981), Zubizarreta (1982)) to point out that middles appear to have exactly the properties ascribed to passives in Chomsky (1981) (cf. Chapter Two (10)):

Middles and Nominate (2)

a. b.

187

No θ-role assigned to NP/S. No Case is assigned to NP/VP for some NP in VP.

Such a proposal is intended to capture the intuition that the examples in (1) are somehow "agentless"; since the θ-role assigned to the subject position is usually the Agent θ-role, if θ-role assignment to the subject is somehow prevented, the clause will indeed be agentless. Because of (2b), the Case Filter will cause the object NP to move to subject position, where it can be assigned ΝΟΜ. So this account captures the essential properties that middles have in common with passives: that the surface subject corresponds to the logical/selected object, and that the logical/selected subject is somehow "suppressed". It also predicts that (3) violates the Case Filter: (3)

*It/there killed the chickens easily.

(2a) and (2b) are related by "Burzio's generalization" (cf. Chapter Two, (15)). These similarities lead Keyser & Roeper to conclude that middle formation ".. is fully analogous to passive formation, where both lexical rules of morphology and a syntactic movement rule are needed to complete the operation. Although middles have no overt morphological characteristics, they still require lexical operations on their thematic and case structures" (p. 402).

We pointed out a number of problems for the theory of passive based on (2) in 2.1. However, those criticisms do not rule out the possibility that such statements might be relevant for some other construction. We can ask the same question about the properties in (2) seen as properties of middles, as about these same properties as putative properties of passives: what are these properties a consequence of? Our initial proposal is that the properties in (2) are the result of the operation of the lexical rule which externalizes Themes. Externalization rules operate on θ-grids as follows: (4)

"E(X): erase the underline on the external argument, if there is one, and underline X. If X = O, then underline nothing." (Williams (1981: 92).

We gave structural definitions of external and internal argument, in Chapter Two (109). Here we adopt Williams' notational convention of underlining the θ-role to be assigned to the external argument - for our purposes here this is a more convenient notation than the one adopted in 2.2.2.

188

The Representation of Implicit and Dethematized

Subjects

(86) where the external argument is marked as the non-subcategorized argument by default. This is intended to show that the argument which receives this θ-role must be the external one (given the assumptions about θ-role assignment put forward in 2.2.2., this notation is strictly speaking redundant for Agents - we assume it for the time being, however, as a convenient way of talking about externalization rules, and return to the point in the next section). Middle-formation, we propose, is the operation of the E(Th) variant of this rule. So middles are formed by the following operation on θ-grids: (5)

kill:

[Gagent, Otheme]

— [ 0

a g

e n t ,

Otheme]·

A very important aspect of this formalization is that it makes the claim that the Agent θ-role is not eliminated, but simply prevented from being assigned to the external argument. In fact, Williams explicitly states that E(X) keeps "the number of arguments in the input and the output the same" (99). The Agent θ-role cannot be assigned to the external argument by kill in (5) after the operation of E(Th) because this would violate the θ-criterion: two θ-roles, Agent and Theme, would be assigned to the same position, the external argument position. Note that, according to the definition of internal argument in Chapter Two (109), the Agent θ-role is not "internalized", i.e. required to be assigned to some internal-argument position. Internal arguments are those arguments θ-marked by a head and not governed by the category containing all that head's subcategorized arguments; they are not governed by this category because they must be inside it in order to be θ-marked by its head under government. Since no category receives an Agent θ-role in this way (but cf. 4.5.), the Agent cannot be assigned to an internal position. Since the Projection Principle applies only to subcategorized arguments, and the Agent θ-role does not (in fact cannot, cf. 2.2.2.) correspond to any subcategorized position, the Projection Principle is not violated by (5). If we assume further that underlined positions in θ-grids do not correspond to subcategorized positions (which in no more than the assumption that our notation and Williams' notation for internal and external arguments are equivalent), then the Theme θ-role will not correspond to a subcategorized position either, avoiding a second potential violation of the Projection Principle. What, then, is the status of the Agent θ-role after the operation of E(Th)? To answer this question, we introduce the notion, following an aspect of Relational Grammar, of chomeur θ-role. A θ-role becomes a chomeur when some lexical rule changes its realization without deleting it. Chomeur θ-roles cannot be realized structurally either as external or internal arguments, although they can be realized in structurally internal positions

Middles and Nominals

189

which are not forced by the Projection Principle (i.e. non-subcategorized positions). We will indicate chdmeur θ-roles with curly brackets, so "{θχ}" means that θχ is a chdmeur. Now we restate (5) as (5'): (5')

kill'. [ 6 a g e n t ,

Otheme]

=

=

=

! >

[{Oagent},

Otheme]·

(Once again, the relevant translation can trivially be made into the notation shown in 2.2.2. (86)). Two further questions remain: (i) why can the Agent θ-role not be assigned to some "extra" internal position (e.g. a ftj'-phrase of some kind)? and (ii) why is there no possibility of the Theme argument not being realized, even though, as an externalized argument, it is no longer subject to the Projection Principle? The simple externalization account, along with our theory of the mapping from the lexicon to syntax presented in 2.2.3., does not prevent either of these results. We will propose in 4.2. that these results are the consequence of aspectual constraints on middles - the fact that middles are statives. The constraints in question do not hold for nominals, and so nominals do not have to realize their Theme argument, and the Agent can appear in an internal position, i.e. a byphrase (see 4.5.). All of this means that kill, after the operation of E(Th), lacks the ability to realize the Agent θ-role of the corresponding transitive. This gives the effect of (2a). What gives the effect of (2b)? We suggest (6) as a principle governing the association of Case frames with Verbs: (6)

If a Verb assigns ACC, it has an internal argument.

(6) leaves open the possibility that a Verb may not assign ACC and still have an internal argument, which is needed for ergative Verbs and Verbs with PP internal arguments {speak to etc.). However, no Verb with an ACC feature can lack an internal argument. So kill will lack an ACC feature after the operation of E(Th) in (5). So the effects of (2) are derived from E(Th) as the middle-formation rule, if we assume (6). In fact, the proposal that middles are derived from transitives by E(Th) has two very interesting consequences for the general middle/passive contrast, which recommend it highly. First, the account just given makes no mention whatever of morphological elements, either affixes or clitics. As middles have no morphological properties distinct from those of actives, this seems to be correct. Second, as the external θ-role is simply not assigned - either to an affix, clitic or argument - we do not find expect to find implicit arguments. And, as pointed out earlier, we do not. We repeat the relevant examples (cf. Chapter Three (4,5)): (7)

a. The book was sold voluntarily, b. *The book sold voluntarily.

190

The Representation of Implicit and Dethematized

(8)

a. The book was sold to make money, b. *The book sold to make money.

Subjects

Keyser & Roeper have a different account of what underlies the properties in (2). They suggest that English has a phonologically n u l l ( s e e Chapter Five on the properties of the various s/Ae-morphemes in Romance). They say: "English has an abstract si clitic that absorbs case and the agent theme, but it is inexpressible. All of the facts follow from this assumption" (406). This proposal is actually incompatible with the facts given the theory of argument clitics adopted in the analysis of passives in the preceding Chapters: if we proposed a null clitic si which was assigned the external θ-role, we would, other things being equal, predict that (7b) and (8b) are grammatical (but cf. Chapter Five, fn. 4). This is a false prediction. In fact, Keyser & Roeper's theory of middles does not really distinguish middles from (syntactic) passives. Hence, in addition to predicting implicitargument effects, their account predicts a range of possible middles that are not found. The following constructions, all of which allow passive in English, do not allow middles: (9)

Idioms a. * Advantage takes of John easily. b. Advantage was taken of John.

(10)

Perception verbs with small-clause a. *John sees singing easily. b. John was seen singing.

(11)

Exceptional Case-Marking Verbs a. *John believes to be a fool easily. John was believed to be a fool b. *This theorem proves to be true easily. This theorem was proved to be true.

(12)

Small-clause Complements T h e s e problems consider easy at MIT. These problems are considered easy.

complements1

Clearly, accounting for middles in terms of a lexical rule which makes reference to thematic properties of predicates like E(Th) will prevent middles like these from being formed. More generally, the account entails that middles will exhibit the property that Williams calls "thematic constancy," meaning that the rule can only affect Verbs with Themes. Hence, where the Verb has no thematic relation with a complement, as in (9-12), middles

Middles and Nominals

191

cannot be formed. Also, middles cannot be formed by externalizing nonTheme complements: (13)

Double-object Verbs a. *Orphans give presents easily at Christmas. b. *Libraries send books best in boxes. c. *Mary bakes a cake easily. d. *John begrudges his wealth a lot.

{Mary is intended as the benefactive argument in (13c).) The parallel with Adjectival passives should be fairly clear. Middles are ruled out wherever Adjectival passives are ruled out. So compare (ΙΟΙ 3) with the following: (14)

a. * Advantage went untaken of John. b. *John was unseen singing. c. *John was unbelieved to be a fool *The theorem was unproven to be true. *John was unconsidered a fool. d. * Orphans are ungiven presents at Christmas. *The books remained unsent to the library. *Mary went unbaked a cake. *John is unbegrudged his wealth.

Note further that Adjectival passives do not give rise to implicit-argument effects: (15)

a. *The book was unsold to make money, b. *The book was voluntarily unsold.

Assuming for the moment (but cf. 4.4.2.), following Wasow (1977) and Williams (1981), that Adjectival passives are lexically derived, the parallel with middles should be enough to suggest that middles are lexically derived. Since Williams' theory offers a simple and general account for both rules, it seems natural to account for both middles and Adjectival passives in terms of E(Th). Two properties cluster with these lexical rules: thematic constancy and lack of implicit arguments. Clearly, the syntactic passive has neither of these properties. A further fact that the E(Th) account of middles captures is that middles cannot be formed from intransitives: (16)

*It runs easily.

192

The Representation of Implicit and Dethematized

Subjects

Note that there is nothing semantically anomalous about (16): it means "running is easy". However, intransitives like run have no Theme argument, so E(Th) cannot apply to them. Moreover, Jaeggli (p.c.) points out that the Romance middle si (or se, as his examples are from Spanish) differs from English middles in ways that indicate precisely that se is a "syntactic middle" (cf. 5.2.). First, the external argument of Romance middles can be modified by a RatC or Class I Adverb: (17)

Las manzanas se comen para adelgazar/voluntariamente. the apples SE eat for to-slim/voluntarily. "Apples eat (are eaten) to get slim/voluntarily."

These implicit-argument effects with se indicate that this morpheme is an argument, or is linked to an argument, rather in the manner of passive en. Second, Romance middles can be formed with NP-movement from the exceptionally Case-marked subject of a small clause, either in the complement to a perception Verb or an epistemic Verb: (18)

a.

b.

Esos problemas se consideran faciles en MIT. these problems SE consider easy at MIT "These problems consider (are considered) easy at MIT." Los jugadores se ven cansados despues del partido. the players SE see tired after of-the match "The players see (look) tired after the match."

From these contrasts (among others - see below) we conclude that it is hard to see how these differences should follow from the existence of a null clitic on the one hand (in English), and the existence of a phonologically spelled-out clitic (in Spanish) on the other. Instead, we develop an account for se which is similar in most respects to that given in Belletti (1982b), see 5.1. and 5.2. An important feature of these accounts of se is that this morpheme is treated either as an argument or as linked to an argument. This treatment of se clearly parallels our treatment of passive en. To sum up so far, we have seen that middles parallel Adjectival passives with respect to both thematic constancy and the lack of implicit arguments, while syntactic passives and the Romance middle se show neither of these properties. These differing clusters of properties indicate that Keyser & Roeper's "null si" proposal cannot be correct, as it stands. Assuming, then, that middles are derived by E(Th), an empirical question arises concerning the definition of the term Theme. In this connection, further contrasts between the English middle and the Spanish .se-construc-

Middles and Nominals

193

tion noticed by Jaeggli are instructive (in each of these examples, the Spanish sentence is an exact translation of the English one): (19)

"Psychological-activity" Verbs a. *Romance languages acquire easily. b. Las lenguas romances se adquieren facilmente.

(20)

Perception Verbs a. *The mountains see easily. b. Las montanas se ven facilmente.

These facts are subsumed under the Affectedness Constraint by Jaeggli (forthcoming, a): (21)

"If a complement of X is unaffected, it is impossible to eliminate the external theta role of X." (p. 24).

Since middle formation appears to involve the elimination of the original external argument, the Affectedness Constraint entails transitive Verbs which take unaffected objects should be unable to form middles. The notion "affected object" corresponds closely to the notion Theme (cf. Roeper (1983) and the notion of Theme as "moved object" in Gruber (1965) and Jackendoff (1972)). We will make this correspondence precise in 4.3. To the extent that the affected argument is the Theme, the proposal that middles are formed by E(Th) captures the Affectedness Constraint. However, at this point the parallel between middles and Adjectival passives breaks down. Adjectival passives are able to appear in examples parallel to (19) and (20): (22)

a. b.

The language was unacquired, The mountains were unseen.

In fact, it has been pointed out by both Fabb (1984) and Levin & Rappaport (1984) that Adjectival passives allow non-Themes to be externalized. Cf.: (23)

a. b. c.

The charity went unaided. The student was unhelped. I remain unthanked for my efforts. (cf. Fabb (1984, (4.205)).

(24)

a. b.

The children were left untaught, The customers went unserved. (cf. Levin & Rappaport (1984, (14,18)).

194

The Representation of Implicit and Dethematized

Subjects

Levin and Rappaport formulate the Sole Complement Generalization (SCG) to account for these facts: (25)

An argument that may stand as the sole complement to a verb can be externalized by APF [=Adjectival Passive Formation - IGR].

Levin & Rappaport conclude that externalization is free with Adjectival passives, subject to the SCG. Note that SCG explains the ungrammaticality of (14d), independently of thematic constancy. They also propose that externalization need not be independently stipulated, but rather follows from general properties of Adjectives. So the only statement Levin & Rappaport need to capture Adjectival passives is that Adjectival Passive Formation derives an Adjective from a Verb. We will briefly reconsider Levin & Rappport's proposals at the end of 4.4.2. The important point for the moment is that the contrast between (19,20) with middles, and (22), with Adjectival passives, does not pose a problem for the E(Th) account of middles. We can maintain that middles are formed by E(Th), and that the Affectedness Constraint is a function of the definition of Theme, and that Adjectival passives are derived by E(X), subject to the SCG. Hence Adjectival passives are not subject to the Affectedness Constraint, but are subject to a form of thematic constancy, in that the externalized argument must be an argument of the deverbal Adjective as shown by (14a-c). Also, a derivation via E(X) precludes implicit-argument effects. In addition to the Affectedness Constraint, middles have a number of properties that the simple statement that they are formed by E(Th) will not account for. First, they are subject to aspectual restrictions, as (26) and (27) show: (26)

Verbs in the progressive2 a. * Bureaucrats are bribing easily. b. *Chickens are killing nicely.

(27)

Verbs in the preterit with punctual interpretations a. ?*Yesterday, the mayor bribed easily. b. ?*Last week, the chickens killed nicely.

The above properties of middles are all discussed by Keyser & Roeper, among others. We will account for them by developing a partial theory of the interaction of aspectual interpretation and θ-roles in the next two sections. Finally, our account of middles does not say anything about modification, but some form of VP-modification seems to be obligatory with middle VPs. This is most commonly an Adverb:

Middles and Nominals (28)

a. b.

195

Bureaucrats bribe *(easily). Chickens kill *(easily).

However, a modal or even contrastive stress on the Verb will suffice: (29)

a. b. c.

Bureaucrats BRIBE. The floor might wax. The bread DOES cut.

Also, negation is enough (this was pointed to me by Klaus Nieter): (30)

a. *This bread cuts. b. This bread doesn't cut.

It is not obvious that the negation facts can be subsumed under some generalized notion of VP-modification, as even a negatively quantified subject seems to suffice for middle-formation. Cf. the following contrast: (31)

a. * Many bureaucrats bribe. b. Not many bureaucrats bribe. c. ?Few bureaucrats bribe.

Furthermore, middles are negative polarity contexts, in the sense that the subject can be quantified with any. Cf.: (32)

a. Any bureaucrat bribes. b. *Any bureaucrat was bribed.

So we need to find an account for the aspectual restrictions in (26,27) and for the licensing conditions for middles more generally, i.e. the requirement for a modifier, and the negative polarity phenomenon. The first question we will consider is that of the aspectual properties of middles. In this section we have seen how E(Th) can derive certain properties of middles - those in (2), and the lexical restrictions on which Verbs can form middles - while leaving others unaccounted for. The rest of this Chapter is devoted to accounting for the remaining properties. We will see that it is possible to return to a syntactic theory of middles, once the mechanisms which constrain middles are properly understood. 4.2. STATIVITY A N D MIDDLES

We now propose an account for the aspectual properties of middles, some of which were observed at the end of the previous section. The account

196

The Representation of Implicit and Dethematized

Subjects

also represents an initial attempt to relate the argument structure of Verbs to aspectual constraints on interpretation and to temporal binding (cf. Gruber (1965), Lakoff (1965), Anderson (1971), Dowty (1979)). The main ideas are that middles are stative Verbs, that middle-formation is a process of stativization and that stativity is a kind of temporal obviativity. The distinction between stative and nonstative Verbs was originally made by Lakoff (1965). Paradigm stative Verbs are know, exist, contain, resemble etc., while run is a good instance of an intransitive nonstative, and build or write are typical transitive nonstatives. Statives can be distinguished from nonstatives in the following ways (the presentation and examples are from Dowty (1979:55-6)): (33)

Only nonstatives occur in the progressive a. *John is knowing the answer. b. John is running. c. John is building a house. b. Run! c. Build a house!

(36)

Only nonstatives cooccur with Class I Adverbs a. *John deliberately knew the answer. b. John ran carefully. c. John carefully built a house.

(37)

Only nonstatives appear in pseudoclefts a. *What the box did was contain books. b. What John did was run. c. What John did was build a house.

(38)

Only nonstatives have iterative simple present a. *John exists every day at 6. b. John runs every day at 6. c. John writes a poem every day at 6.

In addition to these properties, Εης (1985) points out that statives have temporal interpretations which differ from those of non-statives in two ways. First, statives allow a reading for a past complement clause embedded under a past matrix that is impossible with non-statives: (39)

a. b.

John said Mary left. John said Mary knew the answer.

In (39a), the time of leaving must precede the time of saying (this is what

Middles and Nominals

197

Εης calls a "shifted reading" of the past tense in the lower clause), while in (39b) the time of Mary knowing the answer can either precede or be simultaneous with the time of saying. Such "simultaneous readings" (Eng's terms again) are only possible with statives. Second, Stative Verbs do not have to be included in the interval denoted by temporal Adverbs: (40)

a. b.

John visited China last year. John knew the answer last year.

(40a) cannot mean that John's visit to China lasted any longer than last year, although it can occupy any part of last year. (40b), on the other hand, does not have this restriction. In (40b), John can still know the answer at the time the sentence is used; for this sentence to be true it just has to be the case that John's knowledge of the answer lasted for a time at least some of which is included in last year. We can sum up these two properties, again following Εης, by saying that statives are temporally independent. Keyser & Roeper noticed that middles pattern like statives with respect to (38) (cf. *Bribe, bureaucrat!). Moreover, we have already seen that middles are incompatible with Class I Adverbs. Given the similarities of Class I Adverbs and RatC, we expect stative Verbs will be ungrammatical with RatC: 3 (41)

*John knew the answer to impress everyone.

Middles pattern just like statives with respect to the other tests, too: (42)

Complement to object-control Verb *John forced the bureaucrat to bribe easily.

(43)

Pseudoclefts *What the chicken did was kill easily.

(44)

Iterative simple present4 * Bureaucrats bribe every year at Christmas.

To complete the parallelism, we should show that statives like know resist a punctual interpretation in the preterit: (45) ??Yesterday evening John knew the answer. (45) is odd on the reading where where "knowing the answer" is taken as a property of John for the same interval as hitting the ball in (46):

198

The Representation of Implicit and Dethematized

(46)

Yesterday evening John hit the ball.

Subjects

Moreover, middles are temporally independent in exactly the way statives are, according to En